Real-time analysis method and device of student status based on deep learning

Abstract

The present invention discloses a method and device for real-time analysis of student status based on deep learning. The method obtains and preprocesses student image data within the classroom, performs multi-scale face detection and facial feature extraction on the preprocessed image data, extracts expression features based on standardized facial feature data and fuses them with attention indicators and posture features, constructs a time series feature sequence and applies a heavy-tailed recurrent neural network for time series modeling, establishes evaluation criteria, and adjusts evaluation parameters through adaptive thresholds to ultimately obtain student status assessment results. By utilizing a heavy-tailed recurrent neural network and a slow transition mechanism to low-dimensional chaos, the present invention can accurately capture subtle changes and long-term trends in student status, addressing the technical issues of traditional student status monitoring methods, such as poor real-time performance, limited coverage, and insufficient personalization.

Description

Technical Field

[0001] The present invention relates to the field of educational informatization technology, and in particular to a method and device for real-time analysis of student status based on deep learning, which is suitable for monitoring and analyzing student status in smart classrooms, online education platforms, and education management systems. Background Art

[0002] In the process of educational modernization, real-time monitoring and analysis of student status has become a key link in improving teaching quality. In traditional classroom teaching, it is difficult for teachers to pay attention to the status of each student at the same time, especially in large class environments. This problem is more prominent.

[0003] Currently, common methods for monitoring student status include questionnaires and manual observation. Questionnaires collect student self-assessment data on a regular basis to understand student status, but lack real-time performance. Manual observation relies on teachers' experience and judgment, which can be subjective and have limited coverage. In recent years, computer vision-based student status analysis technologies have begun to be applied in education, such as using the OpenCV library for simple facial expression recognition or gesture detection.

[0004] Existing computer vision-based student status analysis technologies primarily utilize traditional convolutional neural network architectures to process and analyze collected student images. These technologies utilize pre-trained models to identify features such as students' facial expressions and head posture, inferring their attention levels and emotional states. These recognition results are then visualized for teachers to use as a reference for adjusting teaching schedules and methods.

[0005] However, existing technologies have obvious shortcomings in practical applications: first, traditional convolutional neural networks are not very effective when processing long-term series data, and it is difficult to accurately capture subtle changes and long-term trends in students' status; second, existing models have poor adaptability to non-standard postures, complex lighting conditions and partial occlusion, and are prone to misjudgment; in addition, existing systems generally lack the ability to adapt to individual differences among students and are unable to perform personalized analysis based on the baseline status of different students, which affects recognition accuracy. Summary of the Invention

[0006] The purpose of the present invention is to provide a real-time analysis method and device for student status based on deep learning, aiming to solve technical problems existing in the existing technology, such as difficulty in accurately capturing subtle changes and long-term trends in student status, poor adaptability to complex environments, and lack of personalized analysis capabilities.

[0007] To achieve the above-mentioned objectives, the present invention provides a real-time analysis method for student status based on deep learning, comprising: acquiring student image data in a classroom, performing image enhancement, illumination compensation and standardization on the student image data to obtain preprocessed image data; performing multi-scale face detection on the preprocessed image data, extracting facial feature points and performing standardization to obtain standardized facial feature data; based on the standardized facial feature data, extracting expression features through a convolutional neural network, combining attention indicators and posture features, performing feature fusion processing to obtain a state feature vector; constructing a time series feature sequence for the state feature vector, performing time series modeling through a heavy-tailed recursive neural network, analyzing the state change trend, and obtaining a state time series analysis result; establishing an evaluation standard based on the state time series analysis result, adjusting the parameters of the evaluation standard through an adaptive threshold, and performing student status evaluation on the evaluation standard after adjusting the parameters to obtain a student status evaluation result.

[0008] Preferably, the student image data is subjected to image enhancement, illumination compensation and standardization processing, including: calculating an image clarity score based on the student image data, and screening out valid images with an image clarity score higher than a preset threshold; performing adaptive histogram equalization processing on the valid images to obtain a brightness-balanced image; applying a bilateral filtering algorithm to the brightness-balanced image to remove image noise to obtain a denoised image; and resizing and normalizing the pixel values ​​of the denoised image to obtain the preprocessed image data.

[0009] Preferably, multi-scale face detection is performed on the pre-processed image data, facial feature points are extracted and standardized, including: applying a cascaded convolutional neural network to perform face area detection on the pre-processed image data to obtain face detection frame data; applying a non-maximum suppression algorithm to the face detection frame data to eliminate overlapping detection frames to obtain optimized face position data; extracting facial feature points based on the optimized face position data to obtain facial feature point coordinate data; performing posture angle estimation and geometric correction on the facial feature point coordinate data to obtain the standardized facial feature data.

[0010] Preferably, based on the standardized facial feature data, expression features are extracted by a convolutional neural network, and feature fusion processing is performed in combination with attention index and posture features, including: applying a convolutional neural network to the standardized facial feature data to extract expression features of the eyes, eyebrows and mouth areas to obtain an expression feature vector; calculating the eye opening degree and blinking frequency based on the standardized facial feature data to obtain an attention index feature vector; applying a posture estimation network to the preprocessed image data to extract upper body key points to obtain a posture feature vector; and performing weighted fusion on the expression feature vector, attention index feature vector and posture feature vector by calculating the attention weight coefficient to obtain the state feature vector.

[0011] Preferably, a time series feature sequence is constructed for the state feature vector, including: arranging the state feature vector in chronological order to obtain original feature time series data; setting sliding windows of multiple time scales for the original feature time series data to obtain segmented time series feature data; applying a heavy-tailed recursive neural network to process the segmented time series feature data to obtain state change feature data; identifying the inflection point of concentration change based on the state change feature data, predicting the state change trend, and obtaining the state time series analysis result.

[0012] Preferably, the segmented time series feature data is processed by applying a heavy-tailed recursive neural network, including: introducing an activation function of an α-stable distribution to the segmented time series feature data to obtain feature data with heavy-tail characteristics; applying an attention mechanism to perform long-term dependency modeling on the feature data with heavy-tail characteristics to obtain time series correlation feature data; dynamically adjusting the network parameter entropy value of the time series correlation feature data to obtain feature data with introduced randomness; calculating a state transition probability matrix based on the feature data with introduced randomness to obtain state migration feature data; applying a multi-task learning strategy to perform training optimization on the state migration feature data to obtain the state change feature data.

[0013] Preferably, an evaluation standard is established based on the state time series analysis results, including: extracting the concentration mean and standard deviation characteristics of the state time series analysis results to obtain statistical feature data; applying a clustering algorithm to the statistical feature data to perform state pattern division to obtain state distribution data; calculating the normal concentration interval and fatigue threshold based on the state distribution data to obtain state evaluation standard parameters; performing multivariate regression analysis on the state evaluation standard parameters in combination with environmental parameters to obtain environmental impact coefficient data; and correcting the state judgment standard based on the environmental impact coefficient data to obtain the evaluation standard.

[0014] Preferably, the parameters of the evaluation criteria are adjusted through adaptive thresholds, and the status of students is evaluated through the evaluation criteria after the parameters are adjusted to obtain student status evaluation results, including: applying a decision tree model to the environmental impact coefficient data to quantify the influence weights of each environmental factor to obtain environmental weight data; calculating the state deviation threshold under current environmental conditions based on the environmental weight data to obtain adaptive threshold data; dynamically adjusting the evaluation criteria based on the adaptive threshold data in combination with historical state evaluation data; evaluating the real-time status of students based on the evaluation criteria after the parameters are adjusted, calculating the degree of state deviation, and obtaining state evaluation data; generating a concentration and emotional state change trend chart for the state evaluation data to obtain the student status evaluation results.

[0015] Preferably, the method also includes: generating class status visualization data based on the student status assessment results, specifically including: performing statistical clustering processing on the student status assessment results to obtain class overall status distribution data; generating a concentration heat map based on the class overall status distribution data to obtain spatial distribution visualization data.

[0016] The present invention also provides a real-time analysis device for student status based on deep learning, including: an image acquisition module, used to obtain student image data in the classroom, perform image enhancement, illumination compensation and standardization on the student image data, and obtain preprocessed image data; a feature extraction module, used to perform multi-scale face detection on the preprocessed image data, extract facial feature points and perform standardization to obtain standardized facial feature data; a state recognition module, used to extract expression features based on the standardized facial feature data through a convolutional neural network, combine attention indicators and posture features, perform feature fusion processing, and obtain a state feature vector; a time series analysis module, used to construct a time series feature sequence for the state feature vector, perform time series modeling through a heavy-tailed recursive neural network, analyze the state change trend, and obtain a state time series analysis result; an evaluation module, used to establish an evaluation standard based on the state time series analysis result, adjust the parameters of the evaluation standard through an adaptive threshold, and evaluate the student's status through the evaluation standard after adjusting the parameters to obtain a student status evaluation result.

[0017] The beneficial effects of the present invention are:

[0018] 1. By acquiring student image data within the classroom and performing image enhancement, illumination compensation, and standardization, we achieved high-quality preprocessing of images under various lighting conditions and environments, improving the quality of the underlying data for subsequent analysis.

[0019] 2. Through "multi-scale face detection and facial feature point standardization processing," the system solves the adaptability issues of traditional face detection in complex classroom environments. It can accurately locate students' faces at different distances and angles, providing precise regional positioning for feature extraction.

[0020] 3. By extracting facial features using a convolutional neural network and combining them with attention metrics and posture features, we achieved effective integration of multimodal features, overcoming the limitations of single features and improving the robustness and accuracy of student status recognition.

[0021] 4. By constructing time series feature sequences and applying heavy-tailed recurrent neural networks for time series modeling, the system can accurately capture subtle changes and long-term trends in student status, enabling high-precision dynamic analysis of student status changes and effectively improving the ability to predict abnormal conditions.

[0022] 5. By "establishing evaluation criteria and using adaptive thresholds to adjust parameters," the problem of traditional fixed criteria being unable to adapt to individual differences among students is resolved. This allows for personalized status assessments tailored to different students and environments, significantly improving the accuracy and practicality of the evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A flow chart of a method for real-time analysis of student status based on deep learning provided by an embodiment of the present invention;

[0025] Figure 2 A flowchart of time series data modeling and state evolution analysis provided by an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of the structure of the real-time analysis device of student status based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0028] like Figure 1 As shown, an embodiment of the present invention provides a real-time analysis method of student status based on deep learning, comprising the following steps:

[0029] S11. Acquire student image data in the classroom, perform image enhancement, illumination compensation, and standardization processing on the student image data to obtain pre-processed image data.

[0030] In this embodiment, this step is achieved by installing a high-definition camera array at various locations within the classroom, capturing raw image data streams from multiple angles, including frontal and side views of students. Environmental parameters, such as light intensity and classroom layout, are also recorded to provide a reference for subsequent processing. The acquired raw image data often suffers from uneven lighting and noise interference, necessitating image enhancement, illumination compensation, and normalization. Specifically, image clarity assessment algorithms (such as the Laplacian operator and gradient magnitude analysis) are first applied to score the quality of each image frame to select high-quality valid images. These valid images are then subjected to illumination compensation and enhancement using adaptive histogram equalization (CLAHE) and multi-scale Retinex algorithms to balance image brightness and contrast. Bilateral filtering and non-local mean filtering are then used to remove image noise while preserving edge details. Finally, resizing, cropping, and pixel value normalization are performed, and data augmentation techniques are used to improve model generalization, ultimately resulting in high-quality pre-processed image data.

[0031] S12. Perform multi-scale face detection on the pre-processed image data, extract facial feature points and perform normalization processing to obtain standardized facial feature data.

[0032] In this embodiment, multi-scale face detection on preprocessed image data is a key step in achieving student status analysis. A feature pyramid network is constructed using an improved MTCNN (Multi-task Cascaded Convolutional Networks) or RetinaFace detection algorithm to detect face regions in images at different scales. This multi-scale detection strategy effectively handles faces at varying distances and angles, improving detection robustness. Detected face regions often overlap and contain redundancy. Therefore, a non-maximum suppression (NMS) algorithm is applied to eliminate overlapping detection frames. Temporal continuity constraints (via inter-frame tracking) are then incorporated to filter out false detections, ensuring accurate localization of each student's face. Subsequently, an improved facial landmark location model (such as the FAN-Face Alignment Network or the shape predictor from the DLib library) is applied to extract 68 standard feature points for each face, including key locations of the eyes, eyebrows, nose, mouth, and facial contour. To handle non-frontal faces, a PnP (Perspective-n-Point) pose estimation algorithm is applied based on these feature points to calculate the 3D head pose angle and generate an affine transformation matrix for geometric correction. Finally, by aligning the eye centers and normalizing the facial proportions, we eliminated individual differences and the influence of camera angles, and obtained standardized facial feature data, laying the foundation for subsequent analysis.

[0033] S13. Based on the standardized facial feature data, expression features are extracted through a convolutional neural network, and feature fusion processing is performed in combination with attention indicators and posture features to obtain a state feature vector.

[0034] In this embodiment, based on standardized facial feature data, the present invention uses deep learning methods to extract and fuse multimodal features to comprehensively analyze students' learning status. First, a deep convolutional neural network with a modified VGG-Face or ResNet architecture is applied to extract deep facial expression features, focusing particularly on key areas of emotional expression such as the eyes, eyebrows, and mouth. Subtle changes in these areas often reflect students' emotional and cognitive states. Simultaneously, a series of attention-related metrics are calculated based on facial feature points, including eye openness (EAR), pupil tracking, blink frequency, and head posture stability. These metrics are important for assessing students' concentration. Furthermore, a human pose estimation network (such as OpenPose or HRNet) can be applied to extract key points of the student's upper body and analyze postural features such as sitting posture, body tilt angle, and movement frequency. These features can reflect students' engagement and status through the lens of body language. Finally, the feature fusion network of the attention mechanism integrates the three modal features of expression features, attention indicators and body posture, dynamically assigns weights according to data quality and environmental conditions, and applies a multi-task learning framework to simultaneously perform concentration level scoring and emotion category judgment. The final output is a state feature vector containing the concentration score, emotion category and its confidence.

[0035] S14, constructing a time series feature sequence for the state feature vector, performing time series modeling through a heavy-tail recursive neural network, analyzing the state change trend, and obtaining a state time series analysis result.

[0036] In this embodiment, constructing a time-series feature sequence from the state feature vector is a key step in analyzing the dynamic changes in student status. The present invention first arranges each student's state feature vector in chronological order to construct raw feature time-series data. To simultaneously capture patterns of change at different time scales, sliding windows of multiple time scales (such as 30 seconds, 1 minute, and 5 minutes) are set to segment the sequence, generating multi-scale time-series feature data. To address the complexity and uncertainty of student status, a recursive neural network structure with heavy-tailed distribution characteristics is designed. The introduction of an α-stable activation function enhances the network's ability to model abnormal states. This network implements a slow transition mechanism to low-dimensional chaos. By dynamically adjusting the entropy of network parameters and introducing controlled randomness, the model maintains stability while capturing subtle changes in status. The model is trained using a multi-task learning strategy and regularization techniques (such as dropout and weight decay). Hyperparameters are tuned using a validation set to obtain an optimized student status time-series model. Based on this model, trend analysis is performed on each student's status sequence to identify concentration inflection points and emotional change patterns, and short-term (such as the next 5-10 minutes) status change trends are predicted, ultimately outputting the time series analysis results of the student status.

[0037] S15. Based on the state time series analysis result, establish an evaluation standard, adjust the parameters of the evaluation standard through an adaptive threshold, and evaluate the student's state according to the evaluation standard after adjusting the parameters to obtain the student's state evaluation result.

[0038] In this embodiment, student evaluation criteria are established based on the results of state time series analysis. Specifically, historical accumulated data is first combined to extract statistical characteristics of each student's state (such as the mean concentration level, standard deviation, and typical emotional distribution) to construct a basic student state profile. Then, an adaptive clustering algorithm and probability distribution fitting method are applied to calculate a personalized state baseline interval for each student, including parameters such as the normal concentration interval, fatigue threshold, and emotional fluctuation tolerance. Multivariate regression analysis and a decision tree model are also used to quantify the impact of environmental factors (such as lighting, course type, and time of day) on the student's state, generating an environmental impact coefficient matrix. Based on this data, an adaptive threshold adjustment algorithm is designed to dynamically adjust the evaluation criteria based on current environmental conditions and the student's real-time state. Specifically, a decision tree model is first applied to quantify the influence weight of each environmental factor and calculate the state deviation threshold under current environmental conditions. The evaluation criteria are then dynamically adjusted based on historical state assessment data. The student's real-time state is then evaluated based on the adjusted evaluation criteria, and the degree of state deviation is calculated. Finally, a comprehensive assessment report is generated, including concentration, emotional state, and their changing trends, and the student's state assessment results are output.

[0039] Optionally, the real-time analysis method of student status based on deep learning also includes: S16, generating class status visualization data according to the student status evaluation results.

[0040] In this embodiment, based on student status assessment results, the present invention generates intuitive class status visualization data to provide decision support for teachers. First, using weighted averaging and statistical clustering methods, the overall class concentration distribution, emotional hotspots, and abnormal status areas are calculated to obtain class status statistics. Then, data visualization techniques (such as heat maps, status curves, and emotional radar charts) are applied to design an intuitive teacher interface, achieving a multi-level display of class overviews and individual student details.

[0041] In addition, you can also set up status anomaly detection rules to monitor and identify key events that require teachers' attention in real time, such as states where concentration is continuously below the preset threshold and states where emotional fluctuations exceed the set range, and generate reminder information through priority sorting.

[0042] In one embodiment of the present invention, performing image enhancement, illumination compensation, and standardization processing on the student image data includes:

[0043] Based on the student image data, an image clarity score is calculated, and valid images with an image clarity score higher than a preset threshold are screened out; adaptive histogram equalization is performed on the valid images to obtain an image with balanced brightness; a bilateral filtering algorithm is applied to the image with balanced brightness to remove image noise to obtain a denoised image; the denoised image is resized and pixel values ​​are normalized to obtain the preprocessed image data.

[0044] Specifically, a high-definition camera array installed at different locations in the classroom is used to collect raw image data streams from multiple angles, including the front and side views of students. At the same time, environmental parameters (such as light intensity, classroom layout, etc.) are recorded to output a multi-perspective raw image dataset and an environmental parameter set.

[0045] Then, based on the multi-view original image dataset, image clarity assessment algorithms (such as Laplace operator and gradient amplitude analysis) are applied to perform quality scoring on each frame of the image, screen out valid images that meet the requirements of subsequent processing, and output the valid image set with quality scoring.

[0046] Then, based on the valid image set and combined with the environmental parameter set, the adaptive histogram equalization (CLAHE) and multi-scale Retinex algorithm are applied to enhance and compensate the images under different lighting conditions, balance the image brightness and contrast, and output the enhanced image set after illumination compensation.

[0047] Subsequently, based on the enhanced image set, bilateral filtering and non-local means (NLM) filtering algorithms are applied to remove image noise while retaining edge details, repairing possible occluded or blurred areas, and outputting the denoised repaired image set.

[0048] Finally, based on the repaired image set, resizing, cropping, and pixel value normalization (normalized to the range [0,1]) are performed, and the dataset is augmented through operations such as rotation, scaling, and horizontal flipping to improve the model's generalization ability and ultimately output a high-quality preprocessed image dataset.

[0049] In one embodiment of the present invention, multi-scale face detection is performed on the pre-processed image data, facial feature points are extracted and standardized, including: applying a cascaded convolutional neural network to perform face region detection on the pre-processed image data to obtain face detection frame data; applying a non-maximum suppression algorithm to the face detection frame data to eliminate overlapping detection frames to obtain optimized face position data; extracting facial feature points based on the optimized face position data to obtain facial feature point coordinate data; and performing posture angle estimation and geometric correction on the facial feature point coordinate data to obtain the standardized facial feature data.

[0050] Specifically, based on the preprocessed image dataset, the improved MTCNN (Multi-task Cascaded Convolutional Networks) or RetinaFace detection algorithm is applied to build a feature pyramid network to detect face areas in the image at different scales, and output a set of face detection boxes containing location coordinates and confidence levels.

[0051] Then, based on the face detection frame set, the non-maximum suppression (NMS) algorithm is applied to eliminate overlapping detection frames, and the false detection results are filtered out in combination with the time continuity constraint (through inter-frame tracking) to output an accurate set of student face positions.

[0052] Next, based on the student face position set, an improved facial landmark location model (such as FAN-FaceAlignment Network or the shape predictor of the DLib library) is applied to extract 68 standard feature points of each face, including the key positions of the eyes, eyebrows, nose, mouth, and facial contours, and output the original facial landmark coordinate set.

[0053] Subsequently, based on the original facial feature point coordinate set, a PnP (Perspective-n-Point)-based posture estimation algorithm is applied to calculate the pitch, yaw, and roll angles of the head. An affine transformation matrix is ​​generated based on the estimated posture angles to perform geometric correction on non-frontal faces and output a set of facial images after posture correction.

[0054] Finally, based on the facial image set after posture correction, the positions of facial landmarks are re-extracted and fine-tuned. Through eye center alignment and facial proportion normalization, individual differences and the influence of camera angles are eliminated, and finally a standardized facial landmark dataset is output.

[0055] In one embodiment of the present invention, based on the standardized facial feature data, expression features are extracted by a convolutional neural network, combined with attention index and posture features, including: applying a convolutional neural network to the standardized facial feature data to extract expression features of the eye, eyebrow and mouth areas to obtain an expression feature vector; calculating the eye openness and blinking frequency based on the standardized facial feature data to obtain an attention index feature vector; applying a posture estimation network to the preprocessed image data to extract upper body key points to obtain a posture feature vector; and performing weighted fusion of the expression feature vector, attention index feature vector and posture feature vector by calculating the attention weight coefficient to obtain the state feature vector.

[0056] Specifically, based on the standardized facial feature point dataset, the deep convolutional neural network with the improved VGG-Face or ResNet architecture is applied to extract the deep expression features of the facial area, focusing on key areas of emotional expression such as eyes, eyebrows and mouth, and outputting a high-dimensional expression feature vector.

[0057] Then, based on the standardized facial feature point dataset, attention-related indicators such as eye openness (EAR), pupil tracking, blinking frequency, and head posture stability are calculated, and these indicators are normalized to output the attention indicator feature vector.

[0058] Next, based on the preprocessed image dataset, a human posture estimation network (such as OpenPose or HRNet) is applied to extract the key points of the student's upper body, analyze posture features such as sitting posture, body tilt angle and movement frequency, and output the body posture feature vector.

[0059] Finally, based on the facial expression feature vector, attention index feature vector and body posture feature vector, the three features are integrated through the feature fusion network of the attention mechanism, dynamic weights are assigned to different features, and a comprehensive feature representation vector is output. The multi-task learning framework is applied to simultaneously perform concentration level scoring (1-5 points) and emotion category (focus, confusion, fatigue, anxiety, etc.) judgment, and calculate the confidence of each result. Finally, the state feature vector containing the concentration score, emotion category and its confidence is output.

[0060] Specifically, multimodal feature fusion is a key step in identifying student states. Expression feature extraction utilizes a modified ResNet-50 architecture, pre-trained on a large-scale facial expression dataset and fine-tuned for the student state recognition task. The network input is a 224×224×3 RGB image. After five residual blocks and global average pooling, the output is a 2048-dimensional feature vector. To enhance the expressiveness of features, a spatial attention module is incorporated into the network, focusing on key areas of emotional expression such as the eyes, eyebrows, and mouth. This is achieved by calculating a spatial attention weight matrix on the feature map and weighting the original feature map to highlight the features of important regions. The network was fine-tuned using 10,000 student facial images labeled with seven basic emotions (neutral, focused, confused, tired, bored, excited, and anxious). A cross-entropy loss function was used, with a learning rate of 0.0001 and a batch size of 32, and training was performed for 50 epochs until convergence.

[0061] In terms of attention index feature extraction, the eye openness (EAR) is first obtained by calculating the ratio of the distance between the vertical and horizontal feature points of the eyes, where the vertical feature point distance is represented by the distance between the midpoints of the upper and lower eyelids, and the horizontal feature point distance is represented by the distance between the corners of the eyes. This ratio can effectively reflect the opening and closing state of the eyes. Secondly, the blinking frequency is calculated by the number of times the EAR value is lower than the threshold (usually 0.2) within the sliding time window (3 seconds). In addition, the present invention also calculates the standard deviation of the head pitch angle, yaw angle and roll angle through the head posture estimation algorithm as a posture stability index; and calculates the pupil movement trajectory length and line of sight dispersion through the pupil position tracking algorithm as a line of sight concentration index. After normalization, these indicators form a 12-dimensional attention index feature vector.

[0062] Body posture feature extraction utilizes the HRNet network, which accurately locates 17 key points on the human body. Particular attention is paid to upper body posture, including the position and motion characteristics of the shoulders, arms, and upper torso. By calculating the relative position and motion characteristics of key points, features such as sitting angle, body tilt, and movement frequency and amplitude are extracted. These features are then analyzed over time to generate a 20-dimensional posture feature vector reflecting student engagement and focus.

[0063] Multimodal feature fusion employs a dynamic weighting method based on an attention mechanism. First, the three modal features (2048-dimensional expression features, 12-dimensional attention index features, and 20-dimensional posture features) are mapped to the same latent space dimension (128 dimensions) through a fully connected layer. Then, a multi-head attention mechanism is designed to calculate the relevance and importance weights of the three features in different representation spaces. Specifically, each modal feature is passed through three attention heads, each of which independently learns the representation of that modality in a specific semantic space. A soft attention mechanism is then used to calculate the fusion weights of the three modalities in each semantic space. This design dynamically adjusts the importance of different modalities based on real-time data quality, environmental conditions, and task requirements, improving the robustness and adaptability of feature fusion.

[0064] The fused features are processed by a multi-task learning head, which simultaneously outputs a focus score (a continuous value ranging from 1 to 5) and an emotion category (a probability distribution of seven basic emotions). The focus score task uses a mean squared error loss function, while the emotion classification task uses a cross-entropy loss function. The two losses are weighted together in a 3:2 ratio. In actual tests, this fusion method has improved recognition accuracy by an average of 18.5% compared to single-modal feature methods. The performance improvement is particularly significant in complex environments such as low lighting or partial occlusion, reaching over 27%.

[0065] like Figure 2As shown, in one embodiment of the present invention, constructing a time series feature sequence for the state feature vector includes:

[0066] S21. Arrange the state feature vectors in chronological order to obtain original feature time series data.

[0067] S22: Setting sliding windows of multiple time scales for the original characteristic time series data to obtain segmented time series characteristic data.

[0068] S23: Apply a heavy-tail recursive neural network to the segmented time series feature data to obtain state change feature data.

[0069] S24. Identify the concentration change inflection point based on the state change characteristic data, predict the state change trend, and obtain the state time series analysis result.

[0070] In this embodiment, time series data modeling and state evolution analysis are key links in understanding the dynamic changes in student status. First, the state feature vectors are arranged in chronological order to construct the original feature time series data. This step arranges the state feature vectors (including concentration scores, emotion categories, confidence levels, etc.) obtained for each student at consecutive time points in sequence according to timestamps to form a time series data structure. This time series arrangement preserves the temporal continuity of student status changes and provides a basic data framework for subsequent analysis. The original feature time series data not only records the state characteristics of the students at each time point, but also implies the rate and pattern information of state changes, which is crucial for understanding students' learning engagement and cognitive processes.

[0071] To comprehensively capture state-changing features at different time scales, sliding windows at multiple time scales are set on the original feature time series data to obtain segmented time series feature data. Specifically, the original time series data is segmented using short-term windows (e.g., 30 seconds), medium-term windows (e.g., 1 minute), and long-term windows (e.g., 5 minutes). Short-term windows capture instantaneous changes and minor fluctuations in student states, making them suitable for identifying sudden shifts in attention or mood swings. Medium-term windows smooth short-term fluctuations and reflect relatively stable state trends, making them suitable for evaluating student responses to specific teaching content. Long-term windows reveal long-term state-changing trends, making them suitable for analyzing learning outcomes across entire courses or teaching units. Multi-scale window segmentation enables simultaneous acquisition of state-changing features at different time granularities, providing rich input data for subsequent deep time series modeling. The data within each window retains the detailed information of the original features while achieving smooth temporal transitions through a sliding mechanism, avoiding the information loss that can occur with hard segmentation.

[0072] The segmented time series feature data is processed using an innovatively designed heavy-tailed recurrent neural network to obtain state change feature data. Traditional recurrent neural networks often perform poorly when processing time series data with uncertainties and mutations, such as student status. Therefore, a recurrent neural network architecture with heavy-tailed distribution characteristics is designed. The network first introduces an α-stable activation function, enabling it to handle outliers and mutations. Unlike the Gaussian distribution, the α-stable distribution has a "heavy tail" property, making it better able to model common non-normal distribution phenomena in the real world, particularly situations such as sudden shifts in students' attention or sudden changes in their emotions. Secondly, an attention mechanism is incorporated into the network architecture to enhance its ability to model long-term dependencies, enabling it to "remember" important state information from earlier moments and leverage it in current state predictions. Furthermore, a slow transition mechanism to low-dimensional chaos is implemented. By dynamically adjusting the entropy of network parameters and introducing controlled randomness, the model maintains overall stability while being sensitive to subtle changes in state. This design ensures that the network neither overfits to noise (such as meaningless student movements) nor ignores important state change signals (such as subtle expressions when experiencing difficulty understanding). Through a multi-task learning strategy, the network simultaneously predicts multiple targets such as changes in concentration, emotion transitions, and state duration, further improving the model's generalization ability and prediction accuracy.

[0073] Based on the state change feature data obtained after processing with a heavy-tailed recurrent neural network, we further identify inflection points in concentration and predict state change trends, ultimately generating state time series analysis results. Inflection points in concentration refer to moments when students' attention levels undergo significant shifts, typically associated with changes in the difficulty of the instructional content, shifts in instructional activities, or changes in students' cognitive load. By analyzing the derivative information (rate of change) and second-order derivative information (acceleration of change) in the state change feature data and combining them with statistical significance testing, we accurately identify these key inflection points. Furthermore, we apply time series forecasting techniques to predict short-term (e.g., 5-10 minutes) future trends in student state based on historical state change patterns. This forecasting strategy employs a hybrid forecasting strategy, combining deterministic components (such as regular changes due to course progress) with stochastic components (such as uncertainty due to individual differences) to improve prediction accuracy and reliability. Ultimately, we generate comprehensive analysis results that include state change trajectories, key inflection point markers, state stability intervals, and future trend forecasts.

[0074] Through this complete set of time series data modeling and analysis processes, discrete student state observation data can be converted into continuous state change curves. Compared with traditional methods, the time series analysis method of the present invention has three significant advantages: first, the multi-scale time window design can simultaneously capture the state change characteristics of different time granularities, providing a more comprehensive view of the learning process; second, the innovative design of the heavy-tailed recurrent neural network greatly improves the sensitivity to abnormal states and subtle changes, making the prediction results more accurate and reliable; finally, the concentration inflection point identification and trend prediction functions provide teachers with forward-looking decision-making support, making teaching interventions more timely and effective.

[0075] Specifically, based on the state feature vector, a feature sequence is first constructed for each student in chronological order, and a sliding window (such as 30 seconds, 1 minute, 5 minutes, and other multi-scales) is set to segment the sequence, and a multi-scale time series feature sequence set is output.

[0076] In one embodiment of the present invention, the segmented time series feature data is processed by applying a heavy-tailed recursive neural network, including: introducing an activation function of an α-stable distribution to the segmented time series feature data to obtain feature data with a heavy-tailed characteristic; applying an attention mechanism to perform long-term dependency modeling on the feature data with a heavy-tailed characteristic to obtain time series correlation feature data; dynamically adjusting the network parameter entropy value of the time series correlation feature data to obtain feature data with introduced randomness; calculating a state transition probability matrix based on the feature data with introduced randomness to obtain state migration feature data; applying a multi-task learning strategy to perform training optimization on the state migration feature data to obtain the state change feature data.

[0077] Specifically, based on statistical learning theory and dynamic system theory, a recursive neural network structure with heavy-tailed distribution characteristics is designed. By introducing the activation function and attention mechanism of α-stable distribution, the network's modeling ability of abnormal states and long-term dependencies is enhanced, and the heavy-tailed RNN model structure is output.

[0078] Then, based on the heavy-tailed RNN model structure, a slow transition mechanism to low-dimensional chaos is realized. By dynamically adjusting the entropy of network parameters and introducing controlled randomness, the model can capture subtle changes in state while maintaining stability, and output an enhanced RNN model with low-dimensional chaotic characteristics.

[0079] Next, based on the multi-scale time series feature sequence set and the enhanced RNN model, multi-task learning strategies and regularization techniques (such as Dropout and weight decay) are applied to train the model. The hyperparameters are tuned through the validation set to output the optimized student state time series model.

[0080] Finally, based on the optimized student status time series model, trend analysis is performed on each student's student status sequence to identify concentration inflection points and emotional change patterns, and predict short-term (such as the next 5-10 minutes) status change trends, and output the student status time series analysis results.

[0081] Specifically, the designed heavy-tailed recursive neural network is a deep learning model customized for the special needs of student status analysis. The heavy-tailed distribution refers to a distribution in which the tail of the probability distribution decays slower than the exponential distribution. This characteristic makes the model more sensitive to abnormal events or extreme values. In this embodiment, the α-stable distribution is used as the basis of the activation function, where the α parameter controls the thickness of the tail of the distribution, and the value range is (0,2]. When α=2, the distribution degenerates into a Gaussian distribution; when α is close to 0, the tail of the distribution becomes thicker. According to the characteristics of the student status data, the α value is selected between 1.5 and 1.8. This range can improve the ability to capture abnormal states while maintaining the stability of the model.

[0082] The basic structure of a heavy-tailed recurrent neural network consists of an input layer, hidden layers, and an output layer. The input layer receives the state feature vector, with a dimension of 32. The hidden layer consists of three layers of bidirectional LSTM units, each containing 128 neurons. An activation layer with α-stable distribution is added after each LSTM layer. The output layer has a dimension set according to the task requirements. The model is trained using the Adam optimizer, with an initial learning rate of 0.001 and a learning rate decay strategy that decreases the learning rate by 0.9 every 50 epochs. To prevent overfitting, the model uses dropout, with a dropout probability of 0.3 after each LSTM layer and an L2 regularization coefficient of 0.0001. The training data is derived from real-world student status data collected in classrooms, consisting of 5,000 samples, 80% of which are used for training and 20% for validation. A multi-task loss function is used during training, simultaneously optimizing both the focus prediction and emotion classification tasks, with a loss weight ratio of 3:2 between the two tasks.

[0083] The slow transition mechanism to low-dimensional chaos is an innovative point of the present invention, which is achieved by dynamically adjusting the entropy of network parameters. Specifically, in the initial stage of model training, the dimension of the network parameter space is high and the system behavior is relatively regular; as the training progresses, the dimension of the parameter space is gradually reduced, and controlled randomness is introduced, showing low-dimensional chaotic characteristics. The implementation method is to calculate the entropy of the network parameters after each training epoch and adjust it according to the preset entropy decay curve. When a small change in the student state sequence is detected, the randomness is appropriately increased to enhance the sensitivity to such changes; when the state is relatively stable, the randomness is reduced to maintain predictive stability. This mechanism enables the model to balance stability and sensitivity and adapt to the dynamic changes in the student state.

[0084] In practical applications, the heavy-tailed recurrent neural network (RNN) improved student status data processing accuracy by approximately 24% compared to traditional RNNs. In particular, it provided an average of 5.7 minutes of early warning time for detecting sudden changes in concentration, providing a sufficient window for teacher intervention. The model demonstrated good generalization in tests across diverse environmental conditions and student populations, and was significantly more robust to interference factors (such as lighting changes and partial occlusion) than the baseline model.

[0085] In one embodiment of the present invention, an evaluation standard is established based on the state timing analysis results, including: extracting the concentration mean and standard deviation characteristics of the state timing analysis results to obtain statistical feature data; applying a clustering algorithm to the statistical feature data to perform state pattern division to obtain state distribution data; calculating the normal concentration interval and fatigue threshold based on the state distribution data to obtain state evaluation standard parameters; performing multivariate regression analysis on the state evaluation standard parameters in combination with environmental parameters to obtain environmental impact coefficient data; and correcting the state judgment standard based on the environmental impact coefficient data to obtain the evaluation standard.

[0086] Specifically, based on the results of the student status time series analysis and combined with historical accumulated data, we extract the statistical characteristics of each student's status (such as the mean concentration, standard deviation, typical emotional distribution, etc.), construct a basic status portrait of the student, and output the student's individual characteristic model.

[0087] Then, based on the student's individual characteristic model, an adaptive clustering algorithm and probability distribution fitting method are applied to calculate a personalized state benchmark interval for each student, including parameters such as normal concentration interval, fatigue threshold, and emotional fluctuation tolerance, and output a personalized state benchmark model.

[0088] Then, based on the environmental parameter set and the results of the time series analysis of student status, multivariate regression analysis and decision tree model are applied to quantify the impact of environmental factors (such as lighting, course type, time period, etc.) on student status and output the environmental impact coefficient matrix.

[0089] Finally, based on the personalized state benchmark model and the environmental impact coefficient matrix, an adaptive threshold adjustment algorithm is designed to dynamically adjust the evaluation criteria according to the current environmental conditions and the real-time status of the students, improve the accuracy of state judgment, and output an adaptive threshold system.

[0090] Specifically, a baseline profile of each student's state is first established. Through long-term observation (data from at least five class hours), the mean, standard deviation, fluctuation period, and emotional distribution characteristics of the student's concentration are calculated. These statistical characteristics are modeled using a Gaussian mixture model (GMM), typically using 3-5 Gaussian components. The model parameters are estimated using the EM algorithm to obtain a probabilistic model that describes the distribution characteristics of the student's state. After model training is complete, key parameters such as the normal concentration range (typically the mean ± 1.5 standard deviations), fatigue threshold (typically the mean - 2 standard deviations), and emotional fluctuation tolerance (based on a 95% confidence interval of historical emotional fluctuations) are automatically determined for each student.

[0091] The impact of environmental factors on student status cannot be ignored, and these impacts are quantified through multivariate regression analysis and decision tree models. The recorded environmental factors include light intensity (lux value), ambient noise level (dB value), classroom temperature (°C), humidity (%), course type (liberal arts or science), class time (morning, afternoon, evening), and course duration. The relationship between these factors and student status deviations is analyzed using the gradient boosting decision tree (GBDT) model to construct an environmental impact coefficient matrix. This matrix records the weight of the impact of each environmental factor on concentration and emotional state within different value ranges. For example, it may be found that in a high temperature environment above 30°C, students' concentration decreases by an average of 0.7 points and their emotional fluctuations increase by 25%. These quantitative results provide a basis for environmental compensation for subsequent status assessments.

[0092] The adaptive threshold adjustment algorithm dynamically adjusts evaluation criteria based on current environmental conditions and a student's historical performance. The algorithm first calculates the expected state deviation under current environmental conditions using an environmental impact coefficient matrix. It then makes personalized adjustments based on the student's individual environmental sensitivity coefficient (learned from historical data). For example, for students who are particularly sensitive to noise, the concentration threshold is lowered accordingly in noisy environments. The algorithm also considers time factors, such as the cumulative effect of fatigue and the natural fluctuations in attention, dynamically adjusting the threshold using a time-weighted function. Experiments show that compared to a fixed threshold, the adaptive threshold improves the accuracy of state assessment by 21.3%, with the effect being particularly pronounced in scenarios with significant environmental fluctuations.

[0093] During the student status assessment process, the degree of deviation between the student's current status and their personal baseline is calculated and its statistical significance is determined. The assessment results include not only concentration scores and emotional categories, but also information such as status change trends, abnormal duration, and severity. The generated assessment report is presented in a visual format, including status time-series curves, current status markers, warning information, and recommended measures, providing teachers with intuitive and comprehensive information on student status. Over the long term, the student's basic status profile and environmental sensitivity model will be continuously updated to achieve continuous optimization and self-evolution of the assessment criteria, adapting to the long-term changes in student status.

[0094] In one embodiment of the present invention, the parameters of the evaluation criteria are adjusted through adaptive thresholds, and the student status is evaluated according to the evaluation criteria after the parameters are adjusted to obtain a student status evaluation result, including: applying a decision tree model to the environmental impact coefficient data to quantify the influence weight of each environmental factor to obtain environmental weight data; calculating the state deviation threshold under the current environmental conditions based on the environmental weight data to obtain adaptive threshold data; dynamically adjusting the evaluation criteria based on the adaptive threshold data in combination with historical state evaluation data; evaluating the real-time state of the student based on the evaluation criteria after the parameters are adjusted, calculating the degree of state deviation, and obtaining state evaluation data; generating a concentration and emotional state change trend chart for the state evaluation data to obtain the student status evaluation result.

[0095] In one embodiment of the present invention, class status visualization data is generated based on the student status assessment results, including: performing statistical clustering processing on the student status assessment results to obtain class overall status distribution data; generating a concentration heat map based on the class overall status distribution data to obtain spatial distribution visualization data.

[0096] Optionally, generating class status visualization data based on the student status assessment results also includes: performing anomaly detection on the student status change trend chart to identify the state where concentration is continuously lower than a preset threshold; and the state where the amplitude of emotional fluctuation exceeds a set range, to obtain key event reminder data.

[0097] Specifically, class status visualization data is generated using a multi-level, multi-dimensional approach. First, the class's student status assessment results are aggregated and analyzed to calculate the overall class focus distribution. Using the K-means clustering algorithm (typically with K=3 or K=4), students are divided into groups with high, medium, and low focus, and the proportion and spatial distribution characteristics of each group are calculated. Furthermore, affective computing models can be applied to analyze the overall emotional atmosphere of the class, identifying dominant emotion types and indicators of emotional synchronization.

[0098] The focus heat map, a core visualization tool, correlates students' physical location in the classroom with their focus status, visually demonstrating the spatial distribution of focus. The heat map uses a gradient color scale, typically mapping from cool tones (low focus) to warm tones (high focus), with color saturation indicating the intensity of focus. A time dimension can also be overlaid on the heat map, allowing dynamic playback to display the evolution of class status over time, helping teachers identify state contagion and key turning points. The heat map is generated using a bilinear interpolation algorithm, creating smooth transitions between discrete student locations, enhancing the aesthetics and readability of the visualization.

[0099] Critical event detection automatically identifies abnormal states requiring teacher attention and intervention. This application incorporates various abnormal state detection rules, including: sustained low concentration, where concentration levels continuously fall below an individual's baseline threshold for a predetermined period (typically 5 minutes); a sharp decline in concentration, where concentration levels drop by more than a predetermined threshold (typically 1.5 points) within a short period (typically 2 minutes); emotional fluctuations, where emotional states undergo significant changes within a short period (typically 3 minutes); and group effects, where multiple students exhibit abnormal states simultaneously. These rules, combined with statistical significance testing, ensure that detected anomalies are pedagogically meaningful and not random fluctuations. Detected abnormal events are prioritized by importance and urgency, generating a priority list and directing the teacher's attention through visual cues (e.g., flashing markers, color coding).

[0100] like Figure 3 As shown, the present invention also provides a real-time analysis device for student status based on deep learning, comprising:

[0101] An image acquisition module 301 is used to acquire student image data in the classroom, perform image enhancement, illumination compensation, and standardization on the student image data to obtain pre-processed image data;

[0102] A feature extraction module 302 is used to perform multi-scale face detection on the pre-processed image data, extract facial feature points and perform normalization processing to obtain standardized facial feature data;

[0103] A state recognition module 303 is configured to extract expression features based on the standardized facial feature data through a convolutional neural network, combine the attention index and the posture feature, perform feature fusion processing, and obtain a state feature vector;

[0104] The time series analysis module 304 is used to construct a time series feature sequence for the state feature vector, perform time series modeling through a heavy-tail recursive neural network, analyze the state change trend, and obtain a state time series analysis result;

[0105] The evaluation module 305 is used to establish an evaluation standard based on the state time series analysis result, and adjust the parameters of the evaluation standard through an adaptive threshold to obtain a student state evaluation result.

[0106] Through the heavy-tailed recursive neural network and the slow transition mechanism to low-dimensional chaos, the present invention can accurately capture the subtle changes and long-term trends of students' states, improve the sensitivity to abnormal states and the prediction accuracy; adopt multimodal feature fusion and dynamic weight distribution technology to integrate the three modal features of expression features, attention indicators and body posture, and improve the robustness and adaptability of student state recognition; establish evaluation standards and adaptive threshold adjustment mechanism, and make adaptive adjustments based on individual differences of students to achieve more accurate student state assessment; through multi-scale time series feature analysis, simultaneously analyze short-term fluctuations and long-term trends, identify concentration inflection points and emotion change patterns, and improve the timeliness and accuracy of state prediction; realize the quantification and compensation of environmental influencing factors, reduce the interference of external factors on recognition results, and improve adaptability in complex environments.

[0107] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A real-time analysis method of student status based on deep learning, characterized in that: include: Acquire student image data in the classroom, perform image enhancement, illumination compensation, and standardization on the student image data to obtain preprocessed image data; Performing multi-scale face detection on the pre-processed image data, extracting facial feature points and performing standardization processing to obtain standardized facial feature data; Based on the standardized facial feature data, expression features are extracted through a convolutional neural network, and feature fusion processing is performed by combining attention indicators and posture features to obtain a state feature vector; Constructing a time series feature sequence for the state feature vector, performing time series modeling through a heavy-tail recursive neural network, analyzing the state change trend, and obtaining a state time series analysis result; Establishing an evaluation standard based on the state time series analysis result, adjusting the parameters of the evaluation standard through an adaptive threshold, and performing a state evaluation on the student according to the evaluation standard after adjusting the parameters to obtain a student state evaluation result; Constructing a time series feature sequence for the state feature vector, including: Arranging the state feature vectors in chronological order to obtain original feature time series data; Setting sliding windows of multiple time scales for the original characteristic time series data to obtain segmented time series characteristic data; Applying a heavy-tail recursive neural network to process the segmented time series feature data to obtain state change feature data; Identify the inflection point of concentration change based on the state change characteristic data, predict the state change trend, and obtain the state time series analysis result; Applying a heavy-tailed recursive neural network to process the segmented time series feature data includes: Introducing an α-stable distribution activation function into the segmented time series feature data to obtain feature data with a heavy-tail characteristic; Applying an attention mechanism to the feature data with heavy-tail characteristics to perform long-term dependency modeling to obtain time-series correlation feature data; Dynamically adjusting the network parameter entropy value of the time series correlation feature data to obtain feature data with randomness introduced; Calculating a state transition probability matrix based on the randomness-introduced feature data to obtain state transition feature data; A multi-task learning strategy is applied to the state transition feature data for training optimization to obtain the state change feature data.

2. The method according to claim 1, characterized in that Performing image enhancement, illumination compensation and standardization processing on the student image data, including: Calculating image clarity scores based on the student image data, and screening out valid images with image clarity scores higher than a preset threshold; Performing adaptive histogram equalization processing on the valid image to obtain an image with balanced brightness; Applying a bilateral filtering algorithm to the brightness-balanced image to remove image noise to obtain a noise-reduced image; The denoised image is resized and pixel values ​​are normalized to obtain the preprocessed image data.

3. The method according to claim 2, characterized in that Performing multi-scale face detection on the pre-processed image data, extracting facial feature points and performing normalization processing, including: Applying a cascade convolutional neural network to the preprocessed image data to perform face region detection to obtain face detection frame data; Applying a non-maximum suppression algorithm to the face detection frame data to eliminate overlapping detection frames to obtain optimized face position data; Extract facial feature points based on the optimized face position data to obtain facial feature point coordinate data; Performing posture angle estimation and geometric correction on the facial feature point coordinate data to obtain the standardized facial feature data.

4. The method according to claim 3, characterized in that Based on the standardized facial feature data, expression features are extracted through a convolutional neural network, and feature fusion processing is performed by combining attention indicators and posture features, including: Applying a convolutional neural network to the standardized facial feature data to extract expression features of the eye, eyebrow, and mouth regions to obtain an expression feature vector; Calculating eye opening and closing degree and blinking frequency based on the standardized facial feature data to obtain an attention index feature vector; Applying a posture estimation network to the preprocessed image data to extract upper body key points to obtain a posture feature vector; The expression feature vector, the attention index feature vector and the posture feature vector are weightedly fused by calculating the attention weight coefficient to obtain the state feature vector.

5. The method according to claim 1, wherein Based on the state timing analysis results, establish evaluation criteria, including: Extracting the concentration mean and standard deviation features from the state time series analysis results to obtain statistical feature data; Applying a clustering algorithm to the statistical feature data to perform state mode division to obtain state distribution data; Calculate the normal concentration interval and fatigue threshold based on the state distribution data to obtain state assessment standard parameters; Performing multivariate regression analysis on the state assessment standard parameters in combination with environmental parameters to obtain environmental impact coefficient data; The state judgment standard is modified based on the environmental impact coefficient data to obtain the evaluation standard.

6. The method according to claim 5, characterized in that Adjusting the parameters of the evaluation criteria by an adaptive threshold, and evaluating the status of the student by the evaluation criteria after the adjustment of the parameters, to obtain the student status evaluation result, including: Applying a decision tree model to the environmental impact coefficient data to quantify the impact weight of each environmental factor to obtain environmental weight data; Calculating a state deviation threshold under current environmental conditions based on the environmental weight data to obtain adaptive threshold data; Dynamically adjusting the evaluation criteria by combining the adaptive threshold data with historical status evaluation data; Evaluate the student's real-time status based on the evaluation criteria after adjusting the parameters, calculate the degree of status deviation, and obtain status evaluation data; A concentration and emotional state change trend graph is generated for the status assessment data to obtain the student status assessment result.

7. The method according to claim 1, characterized in that Also includes: Generate class status visualization data based on the student status assessment results, specifically including: Performing statistical clustering processing on the student status assessment results to obtain overall class status distribution data; A concentration heat map is generated based on the overall state distribution data of the class to obtain spatial distribution visualization data.

8. A real-time analysis device for student status based on deep learning, characterized in that: include: An image acquisition module is used to obtain student image data in the classroom, perform image enhancement, illumination compensation and standardization on the student image data to obtain pre-processed image data; A feature extraction module is used to perform multi-scale face detection on the pre-processed image data, extract facial feature points and perform normalization processing to obtain standardized facial feature data; A state recognition module is used to extract expression features based on the standardized facial feature data through a convolutional neural network, combine attention indicators and posture features, perform feature fusion processing, and obtain a state feature vector; A time series analysis module is used to construct a time series feature sequence for the state feature vector, perform time series modeling through a heavy-tail recursive neural network, analyze the state change trend, and obtain a state time series analysis result; An evaluation module is configured to establish an evaluation standard based on the state time series analysis result, adjust the parameters of the evaluation standard through an adaptive threshold, and evaluate the student's state based on the evaluation standard after adjusting the parameters to obtain a student state evaluation result; Constructing a time series feature sequence for the state feature vector, including: Arranging the state feature vectors in chronological order to obtain original feature time series data; Setting sliding windows of multiple time scales for the original characteristic time series data to obtain segmented time series characteristic data; Applying a heavy-tail recursive neural network to process the segmented time series feature data to obtain state change feature data; Identify the inflection point of concentration change based on the state change characteristic data, predict the state change trend, and obtain the state time series analysis result; Applying a heavy-tailed recursive neural network to process the segmented time series feature data includes: Introducing an α-stable distribution activation function into the segmented time series feature data to obtain feature data with a heavy-tail characteristic; Applying an attention mechanism to the feature data with heavy-tail characteristics to perform long-term dependency modeling to obtain time-series correlation feature data; Dynamically adjusting the network parameter entropy value of the time series correlation feature data to obtain feature data with randomness introduced; Calculating a state transition probability matrix based on the randomness-introduced feature data to obtain state transition feature data; A multi-task learning strategy is applied to the state transition feature data for training optimization to obtain the state change feature data.

Images