Intelligent teaching evaluation method based on multi-dimensional feature fusion

Abstract

This invention discloses an intelligent teaching evaluation method based on multi-dimensional feature fusion, belonging to the field of intelligent teaching evaluation technology. The method includes: collecting real-time multi-source data during the teaching process; generating standardized multimodal teaching data streams; performing semantic parsing to extract multi-dimensional teaching behavior feature vectors; calculating raw scores for three dimensions: teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation; correcting the raw scores to generate a comprehensive evaluation score; identifying strengths and weaknesses to generate personalized evaluation reports; pushing customized teaching skills enhancement micro-courses; and automatically updating evaluation thresholds based on changes in the distribution of group evaluation data. This invention, through multimodal data fusion and deep learning technology, achieves an objective and comprehensive evaluation of teachers' teaching quality, solving the problems of traditional evaluation methods relying on subjective judgment and having a single evaluation dimension, thus improving the objectivity and intelligence of the evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching evaluation technology, and in particular to an intelligent teaching evaluation method based on multi-dimensional feature fusion. Background Technology

[0002] Intelligent evaluation of teaching refers to a teaching quality evaluation method that uses technologies such as artificial intelligence, computer vision, speech processing, and big data analysis to automatically collect, analyze, and evaluate multi-dimensional information such as teachers' teaching behavior, language expression, classroom organization, students' participation status, and learning feedback during the classroom teaching process. Compared with traditional manual evaluation methods, intelligent evaluation of teaching can break through the limitations of subjective experience judgment and achieve objective quantification and dynamic monitoring of the teaching process.

[0003] Currently, classroom teaching quality assessment relies on traditional questionnaires, supervisory observations, or single-dimensional quantitative indicators. This fails to systematically integrate and deeply analyze multi-source heterogeneous data generated during the teaching process, such as teacher behavior, vocal emotions, student facial expressions, and classroom interactions. Consequently, the assessment is often one-dimensional and highly subjective, making it difficult to comprehensively reflect the teacher's true teaching level and classroom effectiveness. Furthermore, existing methods struggle to eliminate assessment biases caused by different subjects and class backgrounds, and cannot accurately pinpoint specific weaknesses in the teacher's teaching process to form a personalized improvement loop. This results in insufficient guidance for the assessment results, hindering the effective promotion of continuous improvement in teaching quality.

[0004] Therefore, we propose an intelligent teaching evaluation method based on multi-dimensional feature fusion to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent teaching evaluation method based on multi-dimensional feature fusion to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent teaching evaluation method based on multi-dimensional feature fusion, comprising the following steps: S1. Collect real-time multi-source data during the teaching process through a multimodal sensor array deployed in the classroom, including teacher video stream, teacher voice stream, student facial expression sequence, and classroom interaction event log.

[0007] S2. Perform timeline alignment and preprocessing on real-time multi-source data, including audio noise reduction, video keyframe extraction, and heterogeneous data normalization, to generate a standardized multimodal teaching data stream.

[0008] S3. Based on the knowledge graph of teaching behavior, perform semantic parsing on the multimodal teaching data stream to extract multidimensional teaching behavior feature vectors, including teacher gesture trajectory, voice emotion features, student head-up rate and interaction frequency.

[0009] S4. Utilize the constructed multi-task deep learning model to jointly evaluate the multi-dimensional teaching behavior feature vectors, and calculate the original scores for three dimensions: teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation.

[0010] S5. Combining course content characteristics, student post-class assessment feedback, and historical teaching evaluation records, the original scores are corrected through a dynamic weighted calibration algorithm to eliminate assessment biases caused by subject differences and class foundation, and a comprehensive teaching evaluation score is generated.

[0011] S6. Based on the comprehensive evaluation scores and feature vectors, the attribution analysis model is used to locate the teacher's strengths and weaknesses in the teaching process, and generate a personalized evaluation report that includes a visualized teaching trajectory.

[0012] S7. Input the analysis results of weak links into the adaptive learning engine to push customized micro-courses to improve teachers' teaching skills, and automatically update the evaluation threshold based on the distribution changes of group evaluation data.

[0013] Preferably, the real-time multi-source data acquisition in S1 includes the following steps: S11. By combining a high-definition camera array with skeletal point recognition technology, data on the teacher's podium positioning trajectory, gestures, and blackboard writing area location are collected.

[0014] S12. Collect teacher's voice signals and ambient sounds through a microphone array, and collect students' facial orientation, expression changes, hand raising, and classroom interaction events through student area cameras and terminal devices.

[0015] S13. The collected teacher skeleton data, voice waveforms, student facial features and interactive events are uniformly added with hardware timestamps and aggregated through the data bus to generate a time-synchronized original multimodal data matrix.

[0016] Preferably, generating a standardized multimodal teaching data stream in step S2 includes the following steps: S21. Background noise suppression of the teacher's speech stream is achieved by using spectral subtraction combined with Wiener filtering, and the denoised Mel frequency cepstral coefficients are extracted as the basis for speech features.

[0017] S22. Keyframes are extracted from the video streams of teachers and students using the inter-frame difference method and object detection algorithm, retaining effective images containing teaching actions and facial expression changes.

[0018] S23. Using minimum-maximum normalization, data such as voice decibels, movement amplitude, and facial expression intensity of different dimensions are mapped to a unified numerical range, and then aligned to the same time scale through linear interpolation to generate a standardized multimodal teaching data stream.

[0019] Preferably, the extraction of the multidimensional teaching behavior feature vector in S3 includes the following steps: S31. Construct a knowledge graph of teaching behaviors, where nodes include atomic actions such as explanation, blackboard writing, questioning, and interactive teaching, and edges represent the logical order between actions, and are associated with standard emotional states and student response models.

[0020] S32. The dynamic time warping algorithm is used to align and match the preprocessed teacher action sequence with the standard action nodes in the knowledge graph to identify the type of classroom activity to which the current teaching segment belongs.

[0021] S33. Based on the aligned activity segments, quantitatively extract the teacher's gesture amplitude and frequency, speech rate, students' collective head-up rate and average facial expression pleasure, and splice them to generate a fixed-dimensional multi-dimensional teaching behavior feature vector.

[0022] Preferably, the calculation of the original scores for the three dimensions in S4 includes the following steps: S41. Construct a multi-task neural network with a shared underlying feature extractor, introduce a domain adversarial training mechanism to eliminate the influence of class differences, and use a self-attention layer to perform global dependency modeling on the complete classroom time sequence.

[0023] S42. Input the multidimensional feature vectors into different task branches of the model in parallel, and calculate the scores for teaching method diversity, classroom atmosphere activity, and knowledge point explanation clarity, respectively.

[0024] S43. The outputs of the three task branches are fused through a multi-head attention mechanism, and a heatmap interpretation of the evaluation decision is generated by combining gradient-weighted class activation mapping, finally outputting the original scores in three dimensions.

[0025] Preferably, the construction of the multi-task neural network and the introduction of the domain adversarial training mechanism in step S41 includes the following steps: S411. Construct a multi-task neural network structure with a shared underlying feature extractor. The underlying network is used to extract general feature representations in teaching behavior, and the upper layer is divided into three task branches: diversity of teaching methods, activity level of classroom atmosphere, and clarity of knowledge point explanation.

[0026] S412. A domain adversarial training mechanism is introduced after the bottom feature extractor. The gradient inversion layer and the domain discriminator are used to adversarially align the data distribution of different classes and subjects, eliminate the interference of domain differences on feature extraction, and enhance the model's cross-scene generalization ability.

[0027] S413. In the temporal feature encoding layer, a multi-head self-attention mechanism is adopted to model the global dependency relationship of the teaching behavior sequence within the complete class time, and to capture the logical relationship between the teaching links such as the opening introduction, the middle interaction, and the ending summary.

[0028] Preferably, the generation of the comprehensive evaluation score in S5 includes the following steps: S51. Establish a course content feature database, store teaching method weight templates for different course types, and collect the post-class evaluation results of students in this class as feedback on teaching effectiveness.

[0029] S52. Retrieve the teacher's historical evaluation records and extract the evaluation trend and improvement rate as dynamic time decay factors.

[0030] S53. The Kalman filter algorithm is used to integrate the original scores with the prior course weights, and the state is corrected by combining post-class assessment feedback and historical trend factors to output the comprehensive teaching evaluation score after eliminating subject bias.

[0031] Preferably, the state correction using the Kalman filter algorithm in step S53 includes the following steps: S531. Construct a time decay weight function, assign weights based on the time sequence of historical evaluation records, so that the recent evaluation trend has a greater impact on the current evaluation than the earlier records, and the decay coefficient is dynamically adjusted based on the teacher's learning curve.

[0032] S532. The teaching method weight template in the course content feature library is used as the state estimation prior, and the original score is used as the observation. The Kalman filter state update equation is used for fusion to eliminate the systematic scoring bias caused by the course type.

[0033] S533. Use the students' after-class assessment results as a dynamic adjustment factor for observation noise. When the assessment results deviate abnormally from the teaching evaluation scores, increase the process noise covariance of the Kalman filter, reduce the fusion confidence of the observation, and prevent abnormal feedback from contaminating the assessment results.

[0034] Preferably, the generation of the personalized teaching evaluation report in S6 includes the following steps: S61. Using gradient attribution or SHAP value analysis methods, calculate the contribution of multidimensional feature vectors to the comprehensive teaching evaluation score, and locate the key teaching behavior segments that affect the score.

[0035] S62. Compare the key segments of the positioning with the teaching behavior knowledge map to identify specific types of advantageous behaviors and weak links.

[0036] S63. Use template generation technology to convert the analysis results into a personalized teaching evaluation report that includes a teaching trajectory heatmap, behavior timeline and improvement suggestions.

[0037] Preferably, the automatic updating of the evaluation threshold in S7 includes the following steps: S71. Utilize deep reinforcement learning to build an adaptive learning engine, using the type of weak link as the state, teaching skills micro-courses as actions, and improving subsequent teaching evaluation scores as rewards to dynamically generate personalized push content.

[0038] S72. Monitor the distribution changes of group evaluation scores. When the distribution shows a statistically significant shift, automatically trigger the recalibration process of the scoring threshold in the evaluation model.

[0039] S73. Synchronize the updated assessment thresholds and pushed micro-courses to the teacher's application, and collect a new round of teaching data to enter the assessment process.

[0040] The present invention has the following beneficial effects: 1. In this invention, a multimodal sensor array is deployed to achieve synchronous acquisition and hardware-level timestamp alignment of heterogeneous data from multiple sources, including teacher video streams, audio streams, student facial expression sequences, and classroom interaction events. Spectral subtraction combined with Wiener filtering is used for audio noise reduction, inter-frame difference is used to extract keyframes from the video, and minimum-maximum normalization and linear interpolation are used to achieve unified dimensional and temporal alignment of the heterogeneous data, generating a standardized multimodal teaching data stream. Compared with existing technologies that rely on a single camera or manual recording, resulting in fragmented data sources and asynchronous time synchronization, this invention achieves systematic fusion and precise alignment of multi-dimensional teaching data, providing a high-quality, highly consistent data foundation for subsequent analysis. It solves the problems of incomplete, asynchronous, and difficult-to-integrate traditional teaching evaluation data collection, improving the completeness and usability of the evaluation data.

[0041] 2. In this invention, a knowledge graph of teaching behavior is constructed and combined with a dynamic time warping algorithm to semantically align teacher action sequences with standard action nodes, achieving structured analysis of teaching behavior. Based on this, a multi-task deep learning model with a shared underlying feature extractor is adopted, introducing a domain adversarial training mechanism to eliminate data distribution interference caused by differences in class and subject. Furthermore, a multi-head self-attention mechanism is used to model the global temporal sequence of the classroom, jointly evaluating three dimensions: the diversity of teaching methods, the activity level of the classroom atmosphere, and the clarity of knowledge point explanation. Compared with existing technologies that rely on subjective scoring by supervisors or single-indicator quantification, making it difficult to characterize the deep semantics and multi-dimensional relationships of teaching behavior, this invention achieves objective quantification and multi-dimensional joint evaluation of the teaching process, solving the problems of strong subjectivity and single-dimensional evaluation, and improving the objectivity, comprehensiveness, and precision of the evaluation.

[0042] 3. In this invention, a course content feature database is established and combined with the Kalman filter algorithm to dynamically integrate prior course weights with student post-class assessment feedback and teacher historical evaluation records. Simultaneously, a time decay weight function is introduced to ensure that recent performance has a greater impact than earlier records, eliminating systematic evaluation biases caused by subject differences and class foundations. Furthermore, attribution analysis is used to accurately locate key behavioral segments and weak points in the teaching process. An adaptive learning engine is constructed using deep reinforcement learning, with weak points as states, micro-courses as actions, and improved evaluation scores as rewards, dynamically pushing personalized improvement content. At the same time, the evaluation threshold is automatically calibrated based on changes in the distribution of group evaluation scores. Compared with existing technologies that have rigid evaluation standards, vague feedback, and no improvement loop, this invention achieves personalized correction of evaluation results, precise location of weak points, and dynamic delivery of teaching skills, forming a data-driven continuous improvement loop for teaching quality. This solves the problems of insufficient guidance from evaluation results and inability to effectively promote teacher professional development, enhancing the practicality, relevance, and long-term effectiveness of evaluations. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the intelligent teaching evaluation method based on multi-dimensional feature fusion according to the present invention. Detailed Implementation

[0044] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0045] Example 1, please refer to Figure 1 As shown: The intelligent teaching evaluation method based on multi-dimensional feature fusion includes the following steps: Data collection: S1. Collect real-time multi-source data during the teaching process through a multimodal sensor array deployed in the classroom, including teacher video stream, teacher voice stream, student facial expression sequence, and classroom interaction event log.

[0046] S2. Perform timeline alignment and preprocessing on real-time multi-source data, including audio noise reduction, video keyframe extraction, and heterogeneous data normalization, to generate a standardized multimodal teaching data stream.

[0047] Analysis and evaluation: S3. Based on the knowledge graph of teaching behavior, perform semantic parsing on the multimodal teaching data stream to extract multidimensional teaching behavior feature vectors, including teacher gesture trajectory, voice emotion features, student head-up rate and interaction frequency.

[0048] S4. Utilize the constructed multi-task deep learning model to jointly evaluate the multi-dimensional teaching behavior feature vectors, and calculate the original scores for three dimensions: teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation.

[0049] Feedback enhances the closed loop: S5. Combining course content characteristics, student post-class assessment feedback, and historical teaching evaluation records, the original scores are corrected through a dynamic weighted calibration algorithm to eliminate assessment biases caused by subject differences and class foundation, and a comprehensive teaching evaluation score is generated.

[0050] S6. Based on the comprehensive evaluation scores and feature vectors, the attribution analysis model is used to locate the teacher's strengths and weaknesses in the teaching process, and generate a personalized evaluation report that includes a visualized teaching trajectory.

[0051] S7. Input the analysis results of weak links into the adaptive learning engine to push customized micro-courses to improve teachers' teaching skills, and automatically update the evaluation threshold based on the distribution changes of group evaluation data.

[0052] The acquisition of real-time multi-source data in S1 includes the following steps: S11. Using a high-definition camera array combined with skeletal point recognition technology, collect data on the teacher's podium positioning trajectory, gestures, and blackboard writing area location. Specifically, the implementation includes: The OpenPose skeletal point recognition algorithm is used to detect the joints of the teacher in the video stream in real time, and outputs a two-dimensional coordinate sequence containing key points such as the shoulder, elbow, wrist, and hip. Let the first... The first frame of the image The image coordinates of the key points are: By performing coordinate transformation using the camera's intrinsic parameter matrix K and extrinsic parameter matrix R, the image coordinates are mapped to the classroom's three-dimensional spatial coordinate system, thus obtaining the corresponding spatial position. .

[0053] S12. A microphone array consisting of 8 microphones is used to collect the teacher's voice signal and ambient sound. A delay-summation beamforming algorithm is used to spatially filter the array signal, enhancing the sound signal from the teacher's direction and suppressing ambient noise in the student area. Let the steering vector of the microphone array be... The beamformer output is ,in Let m be the signal collected by the m-th microphone. The time-domain signal acquired by this microphone. The delay time is relative to the reference microphone. The data is weighted by coefficients and collected through student area cameras and terminal devices, including student facial orientation, expression changes, hand raising, and classroom interaction events.

[0054] S13. The collected teacher skeleton data, speech waveforms, student facial features, and interaction events are uniformly timestamped with hardware timestamps, and aggregated via a data bus to generate a time-synchronized raw multimodal data matrix. In specific implementation: The collected teacher skeletal point data, teacher voice waveform data, student facial expression features, and interaction event data are uniformly stamped with a hardware timestamp generated by a GPS timing module, with time synchronization accuracy controlled within 1 millisecond. All data is aggregated to a local server via a high-speed data bus, and data is associated using the timestamp as the primary key to construct a time-synchronized original multimodal data matrix M. The rows of this matrix correspond to the sampling points of the time series, and the columns include dimensions for teacher behavior features, voice features, student facial expression features, and interaction events, specifically represented as follows: ,in For the spatial coordinate sequence of teacher skeleton points, This is the amplitude sequence of the denoised speech signal. Encode the coordinates of key facial features and expression categories for students. To determine the interaction event type and timestamp, a linear interpolation method is used to unify data with different sampling rates to the same time scale, ensuring strict alignment of each modality's data in the time dimension, and outputting the original multimodal data matrix for subsequent preprocessing.

[0055] Generating a standardized multimodal instructional data stream in S2 includes the following steps: S21. Background noise suppression of the teacher's speech stream is achieved by combining spectral subtraction with Wiener filtering. The Mel-frequency cepstral coefficients after noise reduction are extracted as the basis for speech features. Specific implementation details are as follows: The collected teacher's voice signal Frame segmentation and windowing are performed, with each frame having a length of 512 sampling points and a frame shift of 256 sampling points. The spectrum of each frame signal is obtained by performing a Fast Fourier Transform. Two seconds of ambient noise were collected before the start of class as a noise sample, and the average power spectrum of the noise was calculated. The spectral subtraction method obtains the enhanced speech power spectrum by subtracting the noise power spectrum estimate from the noisy speech power spectrum. ,in The over-subtraction factor ranges from 1.2 to 2.0; the spectrum after spectral subtraction is used as the input to the Wiener filter, and the gain function of the Wiener filter is... ,in The noise suppression coefficient is used. After multiplying the gain function with the original spectrum, the time-domain speech signal is reconstructed by inverse Fourier transform. Based on the denoised speech signal, the denoised speech signal is passed through a Mel filter bank composed of 24 triangular filters. The logarithmic energy of the output of each filter is calculated, and then a 24-dimensional MFCC feature vector is obtained by discrete cosine transform. This feature vector is used for subsequent speech emotion analysis and teaching state recognition.

[0056] S22. Keyframe extraction is performed on the teacher and student video streams using inter-frame difference and object detection algorithms, retaining effective frames containing teaching actions and facial expression changes. Specifically, the implementation involves: For the teacher's video stream, the three-frame difference method is used to calculate the motion amplitude between adjacent frames. Let the first frame be... Frame image is Calculate the difference image and After binarizing the difference image, the motion region is obtained through a logical AND operation. Calculate the pixel ratio of the moving region ,in and These are the width and height of the image, respectively; when When the value exceeds a set threshold of 0.05, the frame is determined to be a motion keyframe. For student video streams, the YOLOv5 object detection algorithm is used to detect the student facial region in the image, and the displacement vector of the student's facial position between adjacent frames is calculated. When more than 80% of the student's facial displacement vector magnitude is greater than 10 pixels, the frame is determined to be an expression change keyframe. At the same time, the teacher's motion keyframes and the student's expression change keyframes are associated according to the timestamp, duplicate frames are removed, and the set of effective frames containing significant teaching behaviors and expression changes is retained.

[0057] S23. Using minimum-maximum normalization, data such as speech decibels, movement amplitude, and facial expression intensity with different dimensions are mapped to a unified numerical range, and then aligned to the same time scale through linear interpolation to generate a standardized multimodal teaching data stream. In specific implementation: For speech features, extract the decibel value of the denoised speech signal. ,in The amplitude of the speech signal is mapped to the interval [0, 1] using min-max normalization: , The normalized speech decibel value. and These represent the maximum and minimum decibel levels of speech, respectively. For the teacher's range of motion, the spatial distance between the wrist key point and the shoulder key point is considered. As an indicator of the range of motion, among which , , and , , The x, y, and z coordinates of the wrist and shoulder key points in the classroom's three-dimensional coordinate system are given, and normalized using the same method. For student facial expression intensity, the emotion category output by the expression classification model is mapped to a pleasure score, ranging from -1 (negative) to 1 (positive), and then normalized to the interval [0, 1]. After normalizing the data for each modality, due to the different sampling rates of different sensors (camera 25fps, microphone 16kHz), linear interpolation is used to unify all data to the same time scale of 25Hz. Let the original data sequence be... The interpolated time point is Interpolation results The normalized and time-aligned teacher speech features, teacher action features, student facial expression features, and interaction event data are concatenated to form a standardized multimodal teaching data stream arranged in time sequence, which serves as the input for subsequent semantic parsing and feature extraction.

[0058] Example 2: Extraction of multidimensional teaching behavior feature vectors in S3 includes the following steps: S31. Construct a knowledge graph of teaching behaviors, where nodes include atomic actions such as lecturing, writing on the blackboard, asking questions, and interactive teaching, and edges represent the logical sequence between actions. This graph is associated with standard emotional states and student response models. In practice: Knowledge graphs are defined using a directed graph structure. , where the node set For each teaching atomic action, an attribute vector is attached to each node. , representing the standard duration of the action, the expected teacher emotional state code (e.g., 1 for positive, 0 for neutral), and the expected student response type (e.g., interaction response ratio), respectively; edge set This indicates the logical dependencies between actions. To transfer weights, this is reflected in the actions within the standard teaching process. Follow-up action The probability of a question node is calculated by taking the question node as an example. Its standard duration is 5 seconds, the teacher's emotional state is expected to be positive, and the students' response is expected to be a head-raising rate of more than 80% and at least 3 students raising their hands to respond. Its outgoing edges point to interactive discussion or explanation nodes, with transition weights of 0.7 and 0.3 respectively. After the graph is constructed, it is stored in JSON format for subsequent action sequence alignment and matching.

[0059] S32. The dynamic time warping algorithm is used to align and match the preprocessed teacher action sequence with the standard action nodes in the knowledge graph to identify the type of classroom activity to which the current teaching segment belongs. Specifically, the implementation involves: Let the teacher's actual action sequence be... Each of them Let the teacher's action type and corresponding timestamp be defined, and let the standard action sequence in the knowledge graph be . Each of them Construct a cost matrix corresponding to the atomic action nodes in the knowledge graph. Its elements express and The matching cost between them is defined as the weighted sum of the Euclidean distance and the time deviation of the action type encoding: ,in Encode the action type. For actual time points, For the desired time point, The time deviation weighting coefficient is set to 0.5; the regularized path with the minimum cumulative cost is found through dynamic programming. ,in It satisfies the boundary conditions and continuity constraints; the cumulative cost formula is: When the cumulative cost is less than the set threshold of 0.3, it is determined that the current teaching segment is successfully matched with the standard action sequence. Based on the dominant atomic action type in the matching path, the type of classroom activity to which the teaching segment belongs (such as the introduction of a new lesson or the explanation of exercises) is identified, and the activity type label is output for subsequent feature quantification. S33. Based on the aligned activity segments, quantitatively extract the teacher's gesture amplitude and frequency, speech rate, student collective head-up rate, and average facial expression pleasure, and concatenate them to generate a fixed-dimensional multi-dimensional teaching behavior feature vector. In specific implementation: For the teacher's gesture amplitude, the gesture amplitude is calculated within the activity segment time window. mean , The total number of sampling frames for gesture amplitude data within the activity segment; gesture frequency. ,in For gesture amplitude sequences exceeding a threshold The number of peaks, The duration of the activity segment; for speech rate, the number of characters in the teacher's speech within the activity segment. (Obtained through automatic speech recognition and transcription), calculate speech rate Unit: words / minute; For the collective head-up rate of students, the ratio of the number of students who looked up at each sampling moment within the activity segment to the total number of students is calculated, and the average value is calculated. ,in For a moment The number of students looking up The total number of students in the class; for the average facial expression pleasure level, the facial expression recognition model outputs the students' facial pleasure scores (ranging from -1 to 1), and the mean of all students and all moments within the activity segment is calculated. , The total number of sampling moments within the activity segment; the above quantitative features are concatenated in a fixed order to form a fixed-dimensional multidimensional teaching behavior feature vector. This vector serves as the input to subsequent multi-task deep learning models for calculating the raw scores across the three dimensions.

[0060] The calculation of the raw scores for the three dimensions in S4 includes the following steps: S41. Construct a multi-task neural network with a shared underlying feature extractor, introduce a domain adversarial training mechanism to eliminate the influence of class differences, and use a self-attention layer to perform global dependency modeling of the complete classroom time sequence. In specific implementation: The construction of a multi-task neural network and the introduction of a domain adversarial training mechanism in S41 includes the following steps: S411. Construct a multi-task neural network structure with a shared underlying feature extractor. The underlying network is used to extract general feature representations in teaching behavior, and the upper layer has three task branches: teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation. In specific implementation: The low-level feature extractor uses a three-layer temporal convolutional network, with 64 convolutional kernels per layer. The kernel size is 3×1, and the stride is 1. The input is a standardized sequence of multidimensional teaching behavior feature vectors. Output feature map ,in To determine the sequence time step, a temporal feature encoding layer is added after the bottom layer network. This layer consists of two bidirectional long short-term memory networks with 128 hidden units, and outputs temporal features. Based on this, three task branches are set up: the teaching method diversity branch adopts global average pooling compression. Then, two fully connected network layers (128 and 64 neurons) are connected, and the output score is generated after Sigmoid activation; the classroom atmosphere activity branch is then global pooled and compared with the extracted gesture frequency. Speech speed Student head-up rate Average facial pleasure The data is concatenated and then output as a score through a fully connected layer. The knowledge point explanation clarity branch uses a convolutional attention module to perform channel weighting on temporal features and then outputs a score through a fully connected layer. The three branches share the parameters of the underlying feature extractor and temporal coding layer.

[0061] S412. A domain adversarial training mechanism is introduced after the low-level feature extractor. Through a gradient inversion layer and a domain discriminator, the data distribution of different classes and subjects is adversarially aligned to eliminate the interference of domain differences on feature extraction and enhance the model's cross-scenario generalization ability. Specifically, the implementation is as follows: The domain discriminator is a three-layer fully connected network (128 and 64 neurons), with the number of neurons in the output layer equal to the number of domain categories. The softmax output domain probability distribution is used. The domain discriminator loss employs cross-entropy loss. Where N is the number of samples, For real domain tags, To predict probabilities, the gradient inversion layer performs an identity transformation during forward propagation and inverts the gradient of the domain discriminator before passing it to the feature extractor during backward propagation. ,in These are the parameters for the feature extractor. For domain discriminator parameters, To enhance the adversarial strength coefficient, the feature extractor learns a domain-independent general feature representation through adversarial training.

[0062] S413. In the temporal feature encoding layer, a multi-head self-attention mechanism is used to model the global dependency relationship of the teaching behavior sequence within the complete class duration, capturing the logical connections between teaching segments such as the opening introduction, the middle interaction, and the closing summary. Specifically, in implementation: Time series features As input, the query matrix is ​​obtained through linear transformation. Key matrix Value matrix ,in For learnable weight matrix, Calculate the scaled dot product attention weights , As a scaling factor, eight attention heads are computed in parallel. The outputs of each head are concatenated along the feature dimension and then passed through a linear layer. Fusion yields a global temporal feature representation. It is used to capture the logical connections between teaching segments. As input for subsequent task branches.

[0063] S42. Input the multi-dimensional feature vectors into different task branches of the model in parallel, and calculate the scores for teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation respectively. In specific implementation: After sharing the underlying feature extractor and self-attention layer, three independent task branches are set up, with the teaching method diversity branch using global temporal feature representation. The input is passed through a two-layer fully connected network and then activated by a Sigmoid function, outputting a score. The classroom atmosphere activity branch will After passing through a fully connected layer and the gesture frequency Speech speed Student head-up rate Average facial pleasure The data is then concatenated and output to obtain the score. The clarity of the explanation of knowledge points is improved by using a convolutional attention module. After channel weighting and extraction of key features, the score is output through a fully connected layer. The three branches are connected via a joint loss function. End-to-end training, with mean squared error loss used for each task. The domain adversarial loss weight coefficient.

[0064] S43. The outputs of the three task branches are fused through a multi-head attention mechanism, and a heatmap interpretation of the evaluation decision is generated by combining gradient-weighted class activation mapping. Finally, the raw scores in three dimensions are output. In specific implementation: Let the intermediate feature vectors of the three fully connected layers be respectively splicing to form a feature matrix After multi-head attention calculates the cross-branch association weights, a fused feature vector is output. Then, the scoring fusion layer generates the original scores in three dimensions. Simultaneously, a gradient-weighted class activation mapping is used to generate a heatmap interpretation, and the scores of each dimension are calculated for their impact on the output feature map of the self-attention layer. The gradients of each channel are used to perform global average pooling to obtain the channel weights. The channel weights are weighted and summed with the feature map, then activated by ReLU to generate a heatmap. The highlighted areas correspond to key teaching behavior periods that affect the score of this dimension, providing interpretable evidence for assessment decisions. Finally, the raw scores of the three dimensions are output for subsequent dynamic weighted calibration.

[0065] Example 3: The generation of the comprehensive evaluation score in S5 includes the following steps: S51. Establish a course content feature database, store teaching method weight templates for different course types, and collect post-class assessment results from students as feedback on teaching effectiveness. Specific implementation details include: The course content feature database is stored using a relational database. Each record contains a course identifier, course type classification (including new lectures, exercise lectures, review lectures, and experimental lectures), and a corresponding teaching method weight template vector. These values ​​represent the relative importance of teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation in the assessment of this lesson type. For example, the weighting for clarity of knowledge point explanation in a new lesson is set at 0.5, for teaching method diversity at 0.3, and for classroom atmosphere activity at 0.2; for problem-solving lessons, the weighting for classroom atmosphere activity is increased. The weighting template is derived through expert scoring and statistical analysis of historical data. Simultaneously, after the lesson, students' post-lesson test scores are collected through an online assessment system to calculate the class average accuracy rate. As a feedback indicator for teaching effectiveness, this indicator is used to measure students' mastery of the course content and reflect the actual teaching effect.

[0066] S52. Retrieve the teacher's historical evaluation records, extract the evaluation trend and the rate of improvement as a dynamic time decay factor. In practice: Retrieve the historical sequence of the teacher's most recent M (M=6) evaluation scores from the database. Sort by time from oldest to most recent, and calculate the evaluation trend factor. The slope of the historical scores was fitted using a linear regression method, i.e. ,in The average of the time sequence numbers. The slope represents the historical average scores, with a positive value indicating an upward trend in teaching quality; calculate the progress factor. , representing the recent improvement ratio relative to the earlier period; constructing a time decay weighting function. ,in For time sequence number ( (Indicates the most recent time) The attenuation coefficient is used to ensure that recent evaluations have a greater impact on the current assessment than earlier ones. This coefficient is dynamically adjusted based on historical evaluation trends and the extent of improvement. When trend factor When the value is positive and the increase is significant, reduce it appropriately. This makes the near-term impact more significant; when the trend is stable or declining, increase the appropriate amount. This reduces the impact of early outliers.

[0067] S53. The Kalman filter algorithm is used to fuse the original scores with prior course weights, and post-class assessment feedback and historical trend factors are combined for state correction to output a comprehensive evaluation score after eliminating subject bias. In specific implementation: S531. Construct a time decay weighting function, allocating weights based on the chronological order of historical evaluation records, ensuring that recent evaluation trends have a greater impact on the current assessment than earlier records. The decay coefficient is dynamically adjusted based on the teacher's learning curve. In specific implementation: For the teacher's most recent M (M=6) historical evaluation records, sorted chronologically from oldest to most recent, the time weight of the i-th record is defined as follows: ,in For the current assessment time, For the i-th evaluation period, For time intervals, This refers to the attenuation coefficient. Dynamically adjust based on teachers' learning curves: Calculate the first-order difference sequence of teachers' historical evaluation scores. If three consecutive differences are positive and the mean is greater than 0.05, the teacher is considered to be in a period of rapid growth. In this case, [the value is taken as...]. This gives higher weight to recent records; if the difference fluctuates between -0.02 and 0.02, the teacher is considered to be in a stable period, and the score is taken accordingly. If the consecutive difference values ​​are negative, the teacher is considered to be in a bottleneck period. To appropriately mitigate the impact of early adverse records.

[0068] S532. Using the teaching method weight template in the course content feature library as the state estimation prior and the original scores as the observations, the Kalman filter state update equation is used for fusion to eliminate the systematic scoring bias caused by course type. Specifically, in implementation: The teaching method weight template output from the course content feature library in S51 As the state prior information for Kalman filtering, the state variables are defined. To assess the true teaching level after eliminating curriculum bias in the k-th evaluation, we first utilize the historical trend factor in S52. Perform state prediction: ,in The trend influence coefficient is set to 0.1; the original score vector output by S43 is multiplied by the course weight template to obtain the weighted observations. This observation reflects the initial evaluation after considering the characteristics of the course; through Kalman filtering observation equations, prediction residuals, Kalman gain calculations, and state update equations, the output state estimate is obtained. This is a comprehensive evaluation score that integrates prior knowledge of the course, effectively eliminating systematic scoring biases caused by different course types.

[0069] S533. Use student after-class assessment results as a dynamic adjustment factor for observation noise. When the assessment results deviate abnormally from the teaching evaluation scores, increase the process noise covariance of the Kalman filter and decrease the fusion confidence of the observation to prevent abnormal feedback from contaminating the assessment results. Specifically, in implementation: Define abnormal deviation judgment indicators ,in The weighted observations of the original score. The average accuracy rate of student after-class assessments collected for S51 was normalized to the interval [0, 1]. When an abnormal deviation occurs, indicating a significant discrepancy between the model's evaluation results and the students' actual understanding, it may be due to factors such as unexpected classroom situations or unusually difficult assessments. In this case, the noise covariance of the Kalman filter is dynamically adjusted. : ,in The baseline process noise covariance is set to 0.01. As the adjustment coefficient (set to 2), the increased process noise covariance leads to increased uncertainty in state prediction. This causes the Kalman filter to reduce the fusion weight of the current observation during state updates, thus mitigating the impact of outlier observations on state estimation. Simultaneously, a confidence factor is introduced into the update of the observation noise covariance RR. ,in The baseline observation noise covariance is set to 0.05. The deviation threshold is set to 0.3. To adjust the slope (take 10); when the degree of divergence Exceeding the threshold At this time, the confidence factor decays rapidly, increasing the observation noise covariance and further reducing the fusion confidence of abnormal observations. Through the above dynamic adjustment mechanism, Kalman filtering can effectively identify and suppress the contamination of evaluation results by abnormal feedback, and output a more robust comprehensive evaluation score.

[0070] The generation of personalized teaching evaluation reports in S6 includes the following steps: S61. Using gradient attribution or SHAP value analysis methods, calculate the contribution of multidimensional feature vectors to the comprehensive teaching evaluation score, and locate the key teaching behavior segments that affect the score. In specific implementation: The SHAP value analysis method is used to calculate the contribution of each teaching behavior characteristic to the overall teaching evaluation score. Let the model function be... The input is a multidimensional teaching behavior feature vector. The output is a comprehensive teaching evaluation score. The SHAP value is based on the Shapley value theory, calculating the contribution of the j-th feature as follows: ,in For the set of all features, For a subset of features that does not contain feature j, This is the model output when only features from the subset SS are used; iterates through all possible feature combinations, calculates the average marginal contribution of each feature, aligns the SHAP values ​​by time window to obtain the contribution distribution of each feature in each teaching behavior segment, sorts the SHAP values, selects the top 3 segments in terms of contribution as key teaching behavior segments, and identifies the key time periods that affect the score and their corresponding dominant feature types.

[0071] S62. Compare the key segments of the positioning with the teaching behavior knowledge map to identify specific types of advantageous behaviors and weak links. In specific implementation: For each key teaching behavior segment located by S61, extract the corresponding teaching action type sequence. ,in For a moment The teaching atomic actions are encoded (explaining, writing on the blackboard, asking questions, interacting, etc.); the corresponding standard action nodes and their attributes are retrieved from the teaching behavior knowledge graph constructed by S31, and the cosine similarity is used to calculate the similarity between the actual action sequence and the feature vectors of each atomic action node in the graph: ,in This is the feature vector of the standard action node in the graph. When the similarity is greater than the threshold of 0.7, the segment is determined to match the atomic action node. If the actual features of the matching node (such as gesture amplitude, student head-up rate) are significantly higher than the upper limit of the standard range of the node in the graph, it is identified as a dominant behavior type; if the actual features are significantly lower than the lower limit of the standard range, it is identified as a weak link type. At the same time, based on the temporal position and the relationship between the preceding and following actions of the key segments, combined with the edge logic in the graph, the specific manifestations of the weak links are identified, such as the waiting time after asking a question being too short, not giving students enough time to think, or the blackboard writing process being disconnected from student interaction.

[0072] S63. Utilize template-based generation technology to convert the analysis results into a personalized teaching evaluation report that includes a teaching trajectory heatmap, a behavior timeline, and improvement suggestions. Specifically, in implementation: An HTML template-based report generation engine is used to mark the key segments located by S61 on the classroom timeline and draw a teaching trajectory heatmap, where the horizontal axis represents classroom time (minutes) and the vertical axis represents the SHAP contribution value. The highlighted areas correspond to key teaching behavior segments. The behavior timeline uses a Gantt chart to display the start and end times and type distribution of each teaching atom action, and compares it with the standard teaching process to intuitively present deviations in time sequence. The improvement suggestion section retrieves corresponding suggestion texts from a pre-set improvement suggestion library based on the types of weaknesses identified by S62.

[0073] Automatically updating the evaluation threshold in S7 includes the following steps: S71. Utilize deep reinforcement learning to build an adaptive learning engine, using the type of weak points as the state, teaching skills micro-courses as actions, and improved subsequent evaluation scores as rewards to dynamically generate personalized content. In specific implementation: An adaptive learning engine is constructed using a deep Q-network, and a state space is defined. The S62 identifies a set of weak points, including categories such as insufficient questioning and interaction, excessively fast speaking speed, disorganized blackboard writing, and lax classroom management; and defines the action space. This is a collection of micro-courses on teaching skills, each consisting of a 5-10 minute short video covering explanations and demonstrations of specific teaching skills, such as classroom questioning techniques, methods for controlling speaking speed, and principles of blackboard design; a reward function is defined. This refers to the improvement in the overall teaching evaluation score in the next assessment after the micro-course is pushed out, compared to the current assessment. The Deep Q-Network uses a three-layer fully connected network structure. The input layer dimension is equal to the size of the state space, each hidden layer has 128 neurons, and the output layer dimension is equal to the size of the action space, outputting the Q-value for each action. Q-value updates use the Bellman equation. ,in The learning rate is set to 0.01. The discount factor is 0.9. The next state after performing an action; after training is completed, for the current type of weak link. Choose the action with the highest Q value. The system retrieves corresponding micro-courses on teaching skills from the micro-course library and pushes them to teachers through the application, enabling precise intervention for weak areas.

[0074] S72. Monitor the distribution changes of group evaluation scores. When a statistically significant shift occurs in the distribution, automatically trigger the recalibration process of the scoring threshold in the evaluation model. Specifically, the implementation includes: Establish a database of group evaluation scores, collecting comprehensive evaluation scores from all teachers across the school or grade level on a weekly basis, constructing a histogram of score distribution, and using the moving average method to monitor changes in the distribution center. Define the first... The weekly average score is ,in For the first The total number of teachers participating in the assessment each week; calculate the difference between the mean values ​​of adjacent weeks. The z-score method was used to determine the significance of the offset: ,in The standard deviation of historical scores; when At a 95% confidence level, a statistically significant shift in the distribution is determined. After triggering the recalibration process, the threshold parameters of the scoring model are recalculated. Specifically, based on the quantiles of the current score distribution, the scoring thresholds for Excellent (top 20%), Good (20%-60%), Pass (60%-90%), and Need for Improvement (bottom 10%) are reset. At the same time, the mapping function parameters between the original scores output by the multi-task model in S43 and the comprehensive evaluation scores in S53 are adjusted to match the score distribution with the preset proportion. The recalibration process uses rolling data from the past three months and is automatically executed once per quarter, or triggered immediately when a significant shift is detected.

[0075] S73. Synchronize the updated assessment thresholds and pushed micro-courses to the teacher's application, and collect a new round of teaching data to enter the assessment process.

[0076] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent teaching evaluation method based on multi-dimensional feature fusion, characterized in that, Includes the following steps: S1. Collect real-time multi-source data during the teaching process through a multimodal sensor array deployed in the classroom, including teacher video stream, teacher voice stream, student facial expression sequence, and classroom interaction event log; S2. Perform timeline alignment and preprocessing on real-time multi-source data, including audio noise reduction, video keyframe extraction, and heterogeneous data normalization, to generate a standardized multimodal teaching data stream; S3. Based on the knowledge graph of teaching behavior, perform semantic parsing on the multimodal teaching data stream to extract multidimensional teaching behavior feature vectors, including teacher gesture trajectory, voice emotion features, student head-up rate and interaction frequency. S4. Utilize the constructed multi-task deep learning model to jointly evaluate the multi-dimensional teaching behavior feature vectors and calculate the original scores for three dimensions: teaching method diversity, classroom atmosphere activity, and clarity of knowledge point explanation. S5. Combining the characteristics of course content, student after-class assessment feedback and historical teaching evaluation records, the original scores are corrected through a dynamic weighted calibration algorithm to eliminate assessment bias caused by subject differences and class foundation, and to generate a comprehensive teaching evaluation score. S6. Based on the comprehensive evaluation scores and feature vectors, the attribution analysis model is used to locate the teacher's strengths and weaknesses in the teaching process, and generate a personalized evaluation report containing a visualized teaching trajectory. S7. Input the analysis results of weak links into the adaptive learning engine to push customized micro-courses to improve teachers' teaching skills, and automatically update the evaluation threshold based on the distribution changes of group evaluation data.

2. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The real-time multi-source data acquisition in S1 includes the following steps: S11. By combining a high-definition camera array with skeletal point recognition technology, the teacher's podium positioning trajectory, gestures, and blackboard writing area location data are collected. S12. Collect teacher's voice signals and ambient sounds through a microphone array, and collect students' facial orientation, expression changes, hand raising, and classroom interaction events through student area cameras and terminal devices; S13. The collected teacher skeleton data, voice waveforms, student facial features and interactive events are uniformly added with hardware timestamps and aggregated through the data bus to generate a time-synchronized original multimodal data matrix.

3. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The generation of standardized multimodal teaching data stream in S2 includes the following steps: S21. Background noise suppression of the teacher's speech stream is achieved by using spectral subtraction combined with Wiener filtering, and the noise-reduced Mel frequency cepstral coefficients are extracted as the basis of speech features. S22. Keyframes are extracted from the video streams of teachers and students using the inter-frame difference method and target detection algorithm to retain effective images containing teaching actions and facial expression changes; S23. Using minimum-maximum normalization, data such as voice decibels, movement amplitude, and facial expression intensity of different dimensions are mapped to a unified numerical range, and then aligned to the same time scale through linear interpolation to generate a standardized multimodal teaching data stream.

4. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The extraction of the multidimensional teaching behavior feature vector in S3 includes the following steps: S31. Construct a knowledge graph of teaching behaviors, where nodes include atomic actions such as explanation, blackboard writing, questioning, and interactive teaching, and edges represent the logical order between actions, and are associated with standard emotional states and student response models. S32. The dynamic time warping algorithm is used to align and match the preprocessed teacher action sequence with the standard action nodes in the knowledge graph to identify the type of classroom activity to which the current teaching segment belongs. S33. Based on the aligned activity segments, quantitatively extract the teacher's gesture amplitude and frequency, speech rate, students' collective head-up rate and average facial expression pleasure, and splice them to generate a fixed-dimensional multi-dimensional teaching behavior feature vector.

5. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The calculation of the original scores for the three dimensions in S4 includes the following steps: S41. Construct a multi-task neural network with a shared underlying feature extractor, introduce a domain adversarial training mechanism to eliminate the influence of class differences, and use a self-attention layer to perform global dependency modeling on the complete classroom time sequence. S42. Input the multidimensional feature vectors into different task branches of the model in parallel, and calculate the scores for teaching method diversity, classroom atmosphere activity, and knowledge point explanation clarity, respectively. S43. The outputs of the three task branches are fused through a multi-head attention mechanism, and a heatmap interpretation of the evaluation decision is generated by combining gradient-weighted class activation mapping, finally outputting the original scores in three dimensions.

6. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 5, characterized in that, The construction of the multi-task neural network and the introduction of the domain adversarial training mechanism in S41 include the following steps: S411. Construct a multi-task neural network structure with a shared underlying feature extractor. The underlying network is used to extract general feature representations in teaching behavior, and the upper layer is divided into three task branches: teaching method diversity, classroom atmosphere activity, and knowledge point explanation clarity. S412. A domain adversarial training mechanism is introduced after the bottom feature extractor. The gradient inversion layer and the domain discriminator are used to adversarially align the data distribution of different classes and subjects, eliminate the interference of domain differences on feature extraction, and enhance the model's cross-scene generalization ability. S413. In the temporal feature encoding layer, a multi-head self-attention mechanism is adopted to model the global dependency relationship of the teaching behavior sequence within the complete class time, and to capture the logical relationship between the teaching links such as the opening introduction, the middle interaction, and the ending summary.

7. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The generation of the comprehensive evaluation score in S5 includes the following steps: S51. Establish a course content feature database, store teaching method weight templates for different course types, and collect the post-class evaluation results of students in this class as feedback on teaching effectiveness. S52. Retrieve the teacher's historical evaluation records and extract the evaluation trend and the extent of improvement as dynamic time decay factors. S53. The Kalman filter algorithm is used to integrate the original scores with the prior course weights, and the state is corrected by combining post-class assessment feedback and historical trend factors to output the comprehensive teaching evaluation score after eliminating subject bias.

8. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 7, characterized in that, The state correction using the Kalman filter algorithm in S53 includes the following steps: S531. Construct a time decay weight function, assign weights based on the time sequence of historical evaluation records, so that the recent evaluation trend has a greater impact on the current evaluation than the earlier records, and the decay coefficient is dynamically adjusted based on the teacher's learning curve. S532. The teaching method weight template in the course content feature library is used as the state estimation prior, and the original score is used as the observation. The Kalman filter state update equation is used for fusion to eliminate the systematic scoring bias caused by the course type. S533. Use the students' after-class assessment results as a dynamic adjustment factor for observation noise. When the assessment results deviate abnormally from the teaching evaluation scores, increase the process noise covariance of the Kalman filter, reduce the fusion confidence of the observation, and prevent abnormal feedback from contaminating the assessment results.

9. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The generation of personalized teaching evaluation reports in S6 includes the following steps: S61. Using gradient attribution or SHAP value analysis methods, calculate the contribution of multidimensional feature vectors to the comprehensive teaching evaluation score, and locate the key teaching behavior segments that affect the score. S62. Compare the key segments of the positioning with the teaching behavior knowledge graph to identify specific types of advantageous behaviors and types of weak links; S63. Use template generation technology to convert the analysis results into a personalized teaching evaluation report that includes a teaching trajectory heatmap, behavior timeline and improvement suggestions.

10. The intelligent teaching evaluation method based on multi-dimensional feature fusion according to claim 1, characterized in that, The automatic updating of the evaluation threshold in S7 includes the following steps: S71. Utilize deep reinforcement learning to build an adaptive learning engine, with the type of weak link as the state, teaching skills micro-courses as actions, and improved subsequent teaching evaluation scores as rewards, to dynamically generate personalized push content. S72. Monitor the distribution changes of group evaluation scores. When the distribution changes significantly, automatically trigger the scoring threshold recalibration process in the evaluation model. S73. Synchronize the updated assessment thresholds and pushed micro-courses to the teacher's application, and collect a new round of teaching data to enter the assessment process.

Images