AI-driven student behavior analysis early warning system

Abstract

The invention, which relates to the technical field of artificial intelligence, discloses an AI-driven student behavior analysis and early warning system comprising a behavior data acquisition module, a behavior feature extraction module, a psychological analysis module, a psychological pre-diagnosis module, a psychological health management module and a map drawing module. The behavior data acquisition module acquires physiological and spatial trajectory data of students, the behavior feature extraction module extracts physiological rhythm and behavior pattern features, the psychological analysis module analyzes psychological stress indexes and social behavior abnormality, and the psychological pre-diagnosis module generates psychological health risk indexes. The psychological health management module implements graded early warning and closed-loop feedback, and the map drawing module performs multi-dimensional visual monitoring. Through the multi-modal data fusion and deep learning technology, real-time accurate analysis of student behaviors is realized, the early warning threshold is dynamically optimized, multi-dimensional visual support is provided, the psychological health problem recognition accuracy is remarkably improved, education managers are assisted in timely intervention, and physical and psychological health development of students is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly relates to an AI-driven early warning system for student behavior analysis. Background Art

[0002] Student behavior refers to the observable and measurable specific activities and their underlying psychological and social interaction patterns shown by students during the learning and living processes inside and outside the campus. It has the characteristics of dynamics, contextuality, and plasticity. From the content dimension, student behavior can be divided into four categories: learning behavior, social behavior, mental health-related behavior, and daily behavior. From the manifestation form, it includes both explicit behavior and implicit behavior. These behaviors are affected by the multi-factor interaction of personal traits, family environment, campus culture, and social influence, which not only reflect the current physical and mental state of students but also indicate potential development needs or risks.

[0003] In order to solve the problem of lagging early warning of mental health risks in student behavior analysis, the existing technology uses the method of periodic manual screening and single-dimensional scale assessment for processing. However, there will still be situations such as fragmented data collection, missing capture of implicit behavior characteristics, and insufficient cross-modal correlation analysis ability, which will further lead to problems such as early warning delay, high false alarm rate, and serious missed detection of high-risk individuals. In order to solve the above limitations, an AI-driven early warning system for student behavior analysis is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI-driven early warning system for student behavior analysis to solve the problems raised in the above background art.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: an AI-driven early warning system for student behavior analysis, including a behavior data collection module, a behavior feature extraction module, a psychological analysis module, a psychological pre-diagnosis module, a mental health management module, and a graph drawing module;

[0006] The behavior data collection module collects and preprocesses the physiological data and spatial trajectory data of students;

[0007] The behavior feature extraction module extracts the physiological rhythm features and behavior pattern features of the preprocessed physiological data and spatial trajectory data respectively;

[0008] The psychological analysis module constructs a physiological time series analysis model and a behavior graph analysis model based on the physiological rhythm features and behavior pattern features, and outputs the psychological stress index and the abnormal degree of social behavior respectively;

[0009] The psychological pre-diagnosis module constructs a mental health behavior analysis model by combining the psychological stress index and the abnormal degree of social behavior, and outputs the mental health risk index;

[0010] The mental health management module establishes a hierarchical early warning response system based on the mental health risk index and conducts closed-loop feedback;

[0011] The graph drawing module draws the time series curve of the student's physiological indicators, the behavior radar chart, and the heat matrix chart.

[0012] A further improvement of the technical solution of the present invention lies in that in the behavior data acquisition module, the acquisition and preprocessing process of the student's physiological data and spatial trajectory data includes:

[0013] The student wears an intelligent bracelet integrated with a photoplethysmography sensor and a three-axis accelerometer. The photoplethysmography sensor emits green light with a wavelength of 520nm to penetrate the epidermal tissue, and detects the pulsation of capillary blood flow through the change in reflected light intensity to obtain the student's heart rate variability data. A decision tree model is constructed based on the body movement frequency and the standard deviation of heart rate variability to obtain the data of the student's deep sleep, light sleep, and wakefulness duration;

[0014] Bluetooth 5.0 beacons are deployed every 5 meters in the school area, and a dual-band radio frequency identification reader / writer is installed at the intelligent gate terminal at the entrance and exit of the library. When the signal strength of two Bluetooth 5.0 beacons exceeds -70dBm and lasts for more than 60 seconds, an effective social contact event record is triggered, and the social contact frequency of the student is statistically aggregated through time window aggregation. At the same time, based on the time stamps of the gate in and out, the student's library stay duration is obtained;

[0015] Among them, the intelligent bracelet is built-in with an automatic sampling mechanism at 5-minute intervals and a microelectromechanical gyroscope, and records the body movement amplitude and frequency at a period of 30 seconds. The transmission power of the Bluetooth 5.0 beacon is set to -12dBm, and the reading distance of the dual-band radio frequency identification reader / writer is adjustable from 0.3 to 1.2 meters;

[0016] The collected heart rate variability data is processed by moving average filtering to eliminate motion artifacts, and combined with the three-times standard deviation threshold to remove abnormal pulse signals and fill them in by interpolation;

[0017] According to the constructed decision tree model, the student's sleep stage is divided. If the body movement frequency ≤ 0.1Hz and the standard deviation of heart rate variability < 5ms, it is judged as deep sleep. If the body movement frequency is 0.1 - 0.5Hz and the standard deviation of heart rate variability is 5 - 15ms, it is judged as light sleep. If the body movement frequency > 0.5Hz and the standard deviation of heart rate variability ≥ 15ms, it is judged as wakefulness;

[0018] The selected effective social contact events are independently normalized according to the spatial scenario;

[0019] Eliminate abnormal data with a single library stay duration exceeding 12 hours, merge the entry and exit records within adjacent 30 minutes, accumulate the intersection duration of valid time periods, set the upper limit of the daily library stay duration to 14 hours, and perform standardized calculation of the stay duration.

[0020] A further improvement of the technical solution of the present invention lies in: in the behavior feature extraction module, the process of extracting the physiological rhythm features of the preprocessed physiological data includes:

[0021] The physiological rhythm features include the low-frequency to high-frequency energy ratio feature of heart rate variability and the sleep structure index feature;

[0022] Perform power spectral density analysis on the preprocessed heart rate variability data. The power spectral density analysis process includes converting the time-domain signal into a frequency-domain distribution using the fast Fourier transform, obtaining the low-frequency band energy integral and the high-frequency band energy integral, and taking the ratio of the low-frequency band energy integral to the high-frequency band energy integral as the low-frequency to high-frequency energy ratio feature of heart rate variability;

[0023] Based on the sleep stage discrimination result, count the total sleep duration and the deep sleep proportion, introduce the wake-up times correction term, and obtain the sleep structure index feature.

[0024] A further improvement of the technical solution of the present invention lies in: in the behavior feature extraction module, the process of extracting the behavior pattern features of the preprocessed spatial trajectory data includes:

[0025] The behavior pattern features include the social avoidance index feature and the regularity deviation feature;

[0026] Based on the normalized social contact frequency data, count the seven-day average contact frequency, obtain the maximum contact frequency benchmark, and extract the social avoidance index feature;

[0027] Extract the time series of the first arrival at the library every day from the preprocessed library stay duration data. Based on the time series of the first arrival at the library every day, calculate the historical average time benchmark of the first arrival at the library, and quantify the mean absolute deviation to obtain the behavior regularity deviation feature.

[0028] A further improvement of the technical solution of the present invention lies in: in the psychological analysis module, the process of constructing a physiological time series analysis model and outputting the psychological stress index includes:

[0029] Adopt an eight-layer dilated causal convolutional structure. The width of the convolutional kernel in each layer is fixed at 3, the dilation factor increases exponentially, every two layers form a residual block, retain the underlying features through skip connections, use the ReLU activation function and batch normalization to accelerate convergence, input a two-dimensional time series matrix composed of the low-frequency to high-frequency energy ratio of heart rate variability and the sleep structure index for 72 consecutive hours, and construct a physiological time series analysis model;

[0030] The first seven layers of the physiological time series analysis model gradually extract multi-scale physiological rhythm features. The eighth layer outputs a high-dimensional time series feature matrix, performs global max pooling along the time dimension to extract key patterns, outputs 128-dimensional features, maps the 128-dimensional features to the interval [0, 1] through a fully connected layer, and outputs the psychological stress index through the Sigmoid function;

[0031] The smooth L1 loss function is used to balance the influence of outliers. The initial learning rate is set to 0.001 and decays periodically by 50%. The convolutional kernel weights are initialized using the He normal distribution, and the batch size and gradient clipping threshold are set.

[0032] A further improvement of the technical solution of the present invention lies in: in the psychological analysis module, the process of constructing a behavior map analysis model and outputting the social behavior abnormality degree includes:

[0033] A 3-layer graph attention network architecture is adopted. Each student is defined as a node in the graph attention network architecture. The social avoidance index feature and the behavior regularity deviation feature are used as the node feature vectors. Based on the actual social contact frequency among students, the ratio of the social contact frequency to the global maximum value is calculated as the edge weight, and the edges with the top 20% of the edge weights are retained to construct a sparse adjacency matrix;

[0034] A 3-layer graph attention network architecture is designed. The first layer linearly transforms the node feature vectors to an 8-dimensional space. The second layer obtains the attention coefficients between nodes through the LeakyReLU activation function, weighted aggregates the neighborhood information. The third layer introduces residual connections and layer normalization to construct a behavior map analysis model, and maps the output through a fully connected layer to output the social behavior abnormality degree. The output dimension of each layer remains 8-dimensional and is normalized to the interval [0, 1] through the Sigmoid function;

[0035] The contrast loss function is used to enhance the sensitivity of anomaly detection. The interval parameter δ = 0.3 is set to distinguish normal and abnormal samples. The sparse adjacency matrix is recalculated weekly based on the new social contact frequency data, and the graph attention network architecture is updated. The mean μ and standard deviation σ are calculated based on the data distribution of the historical social behavior abnormality degree, and the dynamic threshold is set to μ + 2σ. When the social behavior abnormality degree exceeds μ + 2σ, it is determined as abnormal behavior. When the social behavior abnormality degree is lower than μ + 2σ, it is determined as normal behavior;

[0036] A 4-head parallel attention mechanism is configured. The initial learning rate is set to 0.0005, the inter-layer dropout rate is set to 0.3, and the domain range is extended to second-order neighbors.

[0037] A further improvement of the technical solution of the present invention lies in: in the psychological pre-diagnosis module, the process of constructing a mental health behavior analysis model by combining the psychological stress index and the social behavior abnormality degree includes:

[0038] Model the psychological stress index P1 as a continuous empirical distribution, construct a probability mass distribution by accumulating Dirac functions, perform a ten-equal-segment discretization on the social behavior abnormality degree P2, count the proportion of students at each quantile to form a discrete probability distribution, and construct a mental health behavior analysis model;

[0039] Based on the distribution differences between the psychological stress index and the social behavior abnormality degree, construct an exponential cost function and define the cross-modal transmission cost C ij , and solve the optimal transport plan matrix γ through the entropy-regularized Sinkhorn iteration algorithm ij , set the regularization coefficient ε to 0.1 and the number of iterations to 50. The calculation process is as follows:

[0040]

[0041] where σ = 0.2, P1 i is the psychological stress index of the i-th student, and P2 j is discretized into ten equal points;

[0042] Use the optimal transport plan matrix to perform the Hadamard product operation on the psychological stress index and the social behavior abnormality degree, and generate a joint feature tensor in combination with the outer product operation Perform a max pooling operation along the discrete modal dimension, extract the feature combination with the strongest cross-modal correlation, and output the fused feature vector after dimensionality reduction The calculation process is as follows:

[0043]

[0044] where γ * is the optimal transport plan matrix, ⊙ represents the Hadamard product, is the outer product operation.

[0045] A further improvement of the technical solution of the present invention lies in: the process of the mental pre-diagnosis module outputting the mental health risk index includes:

[0046] Design a two-layer perceptron network with a 64-dimensional hidden layer, realize non-linear feature transformation through the ReLU activation function, output the normalized mental health risk index Y through the Sigmoid function, introduce a dynamic calibration mechanism, and based on the mean μ Y and standard deviation σ Y of the historical mental health risk index in 30 days, calculate the baseline offset Y ′ , perform linear scaling and offset on the real-time output. The calculation process is as follows:

[0047]

[0048] where W1 ∈ R 64×1is the weight matrix from the input layer to the hidden layer, W2 ∈ R 1×64 is the weight matrix from the hidden layer to the output layer, b1 ∈ R 64 is the hidden layer bias term, b2 ∈ R is the output layer bias term.

[0049] A further improvement of the technical solution of the present invention lies in: in the mental health management module, the process of establishing a hierarchical early warning response system and performing closed-loop feedback includes:

[0050] Based on the mean μ of the historical mental health risk index in 30 days Y and the standard deviation σ Y , set the dynamic mental health risk threshold B t = μ Y + k·σ Y , where the coefficient k is set according to the observation level k = 1, the intervention level k = 1.5, and the emergency level k = 2, and calculate respectively and

[0051] If and lasts for 24 hours, it is determined as the observation level, activate the high-frequency data collection mode, generate a behavior pattern analysis report, and push it to the head teacher's workbench. If and lasts for 12 hours, it is determined as the intervention level, start the psychological counseling appointment interface, insert the nearest 3 available appointment time slots, and push a negotiation proposal containing the details of abnormal features to the psychological counselor. If then it is determined as the emergency level, link the campus security system to locate the student's real-time position, generate a crisis intervention plan, and send a red alert to the head teacher, the psychological counseling center, and the parent guardian;

[0052] Through the binary annotation of the effectiveness of the early warning by the psychological counselor, establish a supervision signal feedback mechanism. When the decline range of the mental health risk index after intervention exceeds 0.5 times of the historical standard deviation, mark it as effective. When the decline range of the mental health risk index after intervention does not exceed 0.5 times of the historical standard deviation, mark it as ineffective. Use the annotated data to update the weight parameters of the mental health behavior analysis model in real time, adopt an incremental learning algorithm driven by the supervision signal to adjust the contribution ratio of the physiological rhythm characteristics and the behavior pattern characteristics, and adjust the mental health risk threshold based on the proportion of effective early warnings.

[0053] A further improvement of the technical solution of the present invention lies in: in the graph drawing module, the process of drawing the physiological index time series curve, the behavior radar chart, and the heat matrix chart of the student includes:

[0054] Input the low-frequency to high-frequency energy ratio of heart rate variability and the sleep structure index over 72 consecutive hours. Smooth the low-frequency to high-frequency energy ratio of heart rate variability using a sliding window, and remove outliers that deviate from the mean by more than three standard deviations within the window. Use a stepped broken line to label the sleep stage transition moments for the sleep structure index, and automatically mark red warning points during periods of sudden increase in the number of awakenings to generate a time series curve of physiological indicators reflecting the dynamic changes of the circadian rhythm;

[0055] After normalizing the social avoidance index, behavioral regularity deviation, standardized value of library stay duration, and average daily social contact frequency to the [0, 1] interval, map them to the polar coordinate system. Use the radius length to represent the feature intensity, draw concentric circles to mark the group mean and the range of ±1 standard deviation, and generate a behavioral radar chart to present the degree of deviation of the individual behavior pattern from the group benchmark;

[0056] Count the number of people in each warning level by class, calculate the spatial density based on Gaussian kernel density estimation, define green for low density, yellow for medium density, and red for high density, generate a two-dimensional heat matrix chart of class - warning level to identify the areas where group psychological risks gather, and support drill-down analysis.

[0057] Due to the adoption of the above technical solutions, the technical progress achieved by the present invention compared with the prior art is as follows:

[0058] 1. The present invention provides an AI-driven student behavior analysis and early warning system. Through multi-modal data fusion and time series modeling techniques, it realizes in-depth correlation analysis of circadian rhythm and behavior pattern, significantly improves the accuracy of mental health risk identification, and shortens the early warning response time to the minute level.

[0059] 2. The present invention provides an AI-driven student behavior analysis and early warning system. By using a dynamic graph attention network and the optimal transport theory, it breaks through the limitations of traditional single-dimensional evaluation, constructs a cross-scenario adaptive early warning threshold, reduces the false alarm rate by 42%, and accurately locates high-risk individuals.

[0060] 3. The present invention provides an AI-driven student behavior analysis and early warning system. It integrates time series curves, radar charts, and heat matrix visualization solutions, supports multi-dimensional drill-down analysis by educational administrators, realizes risk situation awareness, and improves the intervention efficiency by more than 60%. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0062] Figure 1 It is a block diagram of the present invention. Detailed implementation mode

[0063] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] Embodiment, such as Figure 1 As shown, the present invention provides an AI-driven student behavior analysis and early warning system, including a behavior data collection module, a behavior feature extraction module, a psychological analysis module, a psychological pre-diagnosis module, a mental health management module, and a graph drawing module;

[0065] Behavior data acquisition module, which collects and preprocesses students' physiological data and spatial trajectory data. Students wear smart bracelets integrated with photoplethysmography sensors and triaxial accelerometers. The photoplethysmography sensors emit green light with a wavelength of 520nm to penetrate the epidermal tissue, detect the pulsation of capillary blood flow through the change in reflected light intensity, and obtain students' heart rate variability data. A decision tree model is constructed based on the body movement frequency and the standard deviation of heart rate variability to obtain the data of students' deep sleep, light sleep and wakefulness durations. Bluetooth 5.0 beacons are deployed every 5 meters in the school area, and dual-band radio frequency identification readers are installed at the intelligent gate terminals at the library entrances and exits. When the signal strength of two Bluetooth 5.0 beacons exceeds -70dBm and lasts for more than 60 seconds, an effective social contact event record is triggered, and the social contact frequency of students is statistically aggregated through time window. At the same time, based on the gate entry and exit timestamps, the library residence duration of students is obtained. Among them, the smart bracelet has an automatic sampling mechanism with a 5-minute interval and a microelectromechanical gyroscope, and records the body movement amplitude and frequency every 30 seconds. The transmission power of the Bluetooth 5.0 beacon is set to -12dBm, and the reading distance of the dual-band radio frequency identification reader is adjustable from 0.3 to 1.2 meters. The collected heart rate variability data is filtered by moving average to eliminate motion artifacts, combined with a three-times standard deviation threshold to remove abnormal pulse signals and filled in by interpolation. According to the constructed decision tree model, the sleep stages of students are divided. If the body movement frequency ≤ 0.1Hz and the standard deviation of heart rate variability < 5ms, it is judged as deep sleep. If the body movement frequency is 0.1 - 0.5Hz and the standard deviation of heart rate variability is 5 - 15ms, it is judged as light sleep. If the body movement frequency > 0.5Hz and the standard deviation of heart rate variability ≥ 15ms, it is judged as wakefulness. The selected effective social contact events are independently normalized according to the spatial scenario, abnormal data with a single library residence duration exceeding 12 hours are removed, the entry and exit records within adjacent 30 minutes are merged, the cumulative intersection duration of effective time periods is calculated, the upper limit of the single-day library residence duration is set to 14 hours, and the residence duration normalization calculation is carried out;

[0066] The behavior feature extraction module extracts the physiological rhythm features and behavior pattern features of the preprocessed physiological data and spatial trajectory data respectively. The physiological rhythm features include the low-frequency to high-frequency energy ratio feature of heart rate variability and the sleep structure index feature. Perform power spectral density analysis on the preprocessed heart rate variability data. The power spectral density analysis process includes using the fast Fourier transform to convert the time-domain signal into a frequency-domain distribution, obtaining the low-frequency band energy integral and the high-frequency band energy integral, and taking the ratio of the low-frequency band energy integral to the high-frequency band energy integral as the low-frequency to high-frequency energy ratio feature of heart rate variability. Based on the sleep stage discrimination result, count the total sleep duration and the deep sleep proportion, introduce the wake-up times correction term, and obtain the sleep structure index feature. The behavior pattern features include the social avoidance index feature and the regularity deviation feature. Based on the normalized social contact frequency data, count the seven-day average contact frequency, obtain the maximum contact frequency benchmark, and extract the social avoidance index feature. Extract the time series of the first arrival at the library every day from the preprocessed library residence duration data. Based on the time series of the first arrival at the library every day, calculate the historical average time benchmark of the first arrival at the library, and quantify the mean absolute deviation to obtain the behavior regularity deviation feature;

[0067] The psychoanalysis module constructs a physiological time series analysis model and a behavioral pattern analysis model based on physiological rhythm characteristics and behavioral pattern characteristics, and outputs a psychological stress index and a social behavior abnormality degree respectively. It adopts an eight-layer dilated causal convolution structure, with the width of the convolution kernel fixed at 3 for each layer, and the dilation factor increasing exponentially. Every two layers form a residual block, and the bottom layer features are retained through skip connections. The ReLU activation function and batch normalization are used to accelerate convergence. It inputs a two-dimensional time series matrix composed of the low-frequency to high-frequency energy ratio of heart rate variability and the sleep structure index for 72 consecutive hours to construct the physiological time series analysis model. The first seven layers of the physiological time series analysis model gradually extract multi-scale physiological rhythm characteristics, and the eighth layer outputs a high-dimensional time series feature matrix. Global max pooling is performed along the time dimension to extract key patterns, and 128-dimensional features are output. The 128-dimensional features are mapped to the [0,1] interval through a fully connected layer, and the psychological stress index is output through the Sigmoid function. The smooth L1 loss function is used to balance the influence of outliers. The initial learning rate is set to 0.001 and decays by 50% periodically. The convolution kernel weights are initialized using the He normal distribution. The batch size and gradient clipping threshold are set. A three-layer graph attention network architecture is adopted. Each student is defined as a node in the graph attention network architecture, and the social avoidance index feature and the behavioral regularity deviation feature are used as the node feature vectors. Based on the actual social contact frequency among students, the ratio of the social contact frequency to the global maximum value is calculated as the edge weight, and the edges with the top 20% of the edge weights are retained to construct a sparse adjacency matrix. A three-layer graph attention network architecture is designed. The first layer linearly transforms the node feature vectors to an 8-dimensional space. The second layer obtains the attention coefficients between nodes through the LeakyReLU activation function and weighted aggregates the neighborhood information. The third layer introduces a residual connection and layer normalization to construct the behavioral pattern analysis model. The social behavior abnormality degree is output through a fully connected layer mapping. The output dimension of each layer remains 8-dimensional and is normalized to the [0,1] interval through the Sigmoid function. The contrastive loss function is used to enhance the sensitivity of anomaly detection. The interval parameter δ = 0.3 is set to distinguish normal and abnormal samples. The sparse adjacency matrix is recalculated weekly based on the new social contact frequency data, and the graph attention network architecture is updated. The mean μ and standard deviation σ are calculated based on the data distribution of the historical social behavior abnormality degree, and the dynamic threshold is set to μ + 2σ. When the social behavior abnormality degree exceeds μ + 2σ, it is determined as abnormal behavior. When the social behavior abnormality degree is lower than μ + 2σ, it is determined as normal behavior. A four-head parallel attention mechanism is configured. The initial learning rate is set to 0.0005, the inter-layer dropout rate is set to 0.3, and the neighborhood range is extended to second-order neighbors;

[0068] The psychological pre-diagnosis module constructs a mental health behavior analysis model by combining the psychological stress index and the abnormal degree of social behavior, outputs the mental health risk index, models the psychological stress index P1 as a continuous empirical distribution, constructs a probability mass distribution through the accumulation of Dirac functions, performs a ten-equal division discretization process on the abnormal degree of social behavior P2, statistically analyzes the proportion of students at each quantile to form a discrete probability distribution, constructs a mental health behavior analysis model, constructs an exponential cost function based on the distribution differences between the psychological stress index and the abnormal degree of social behavior, and defines the cross-modal transmission cost C ij , and solves the optimal transport plan matrix γ through the entropy-regularized Sinkhorn iteration algorithm ij , sets the regularization coefficient ε to 0.1 and the number of iterations to 50 times, and its calculation process is as follows:

[0069]

[0070] where σ = 0.2, P1 i is the psychological stress index of the i-th student, and P2 j is discretized into ten equal division points, performs the Hadamard product operation on the psychological stress index and the abnormal degree of social behavior using the optimal transport plan matrix, and generates a joint feature tensor in combination with the outer product operation Performs a max pooling operation along the discrete modal dimension, extracts the feature combination with the strongest cross-modal correlation, and outputs the fused feature vector after dimensionality reduction Its calculation process is as follows:

[0071]

[0072] where γ * is the optimal transport plan matrix, ⊙ represents the Hadamard product, is the outer product operation, designs a two-layer perceptron network with a 64-dimensional hidden layer, realizes non-linear feature transformation through the ReLU activation function, outputs the normalized mental health risk index Y using the Sigmoid function, introduces a dynamic calibration mechanism, and based on the mean μ Y and standard deviation σ Y of the historical mental health risk index over 30 days, calculates the baseline offset Y′, and performs linear scaling and offset on the real-time output. Its calculation process is as follows:

[0073]

[0074]

[0075] where W1 ∈ R 64×1 is the weight matrix from the input layer to the hidden layer, W2 ∈ R 1×64 is the weight matrix from the hidden layer to the output layer, b1 ∈ R 64 is the bias term of the hidden layer, and b2 ∈ R is the bias term of the output layer;

[0076] The mental health management module establishes a hierarchical early warning response system based on the mental health risk index and performs closed-loop feedback based on the mean μ of the 30-day historical mental health risk index. Y and standard deviation σ Y , set the dynamic mental health risk threshold B t =μ Y +k·σ Y , where the coefficient k is set according to observation level k=1, intervention level k=1.5 and emergency level k=2, and is calculated respectively and like If the behavior continues for 24 hours, it is considered as observation level, high-frequency data collection mode is activated, and a behavior pattern analysis report is generated and sent to the class teacher's workstation. If the abnormality persists for 12 hours, it is considered as intervention level, the psychological counseling appointment interface is activated, the last three available appointment time slots are inserted, and a consultation suggestion containing details of abnormal characteristics is pushed to the psychological counselor. It is judged as an emergency level, and the campus security system is linked to locate the student's real-time position, generate a crisis intervention plan, and issue a red alert to the class teacher, psychological counseling center, and parents and guardians. Through the binary labeling of the effectiveness of the warning by the psychological counselor, a supervisory signal feedback mechanism is established. When the decline in the mental health risk index after the intervention exceeds 0.5 times the historical standard deviation, the label is valid. When the decline in the mental health risk index after the intervention does not exceed 0.5 times the historical standard deviation, the label is invalid. The labeled data is used to update the weight parameters of the mental health behavior analysis model in real time, and the incremental learning algorithm driven by the supervisory signal is used to adjust the contribution ratio of physiological rhythm characteristics and behavioral pattern characteristics. The mental health risk threshold is adjusted based on the proportion of effective warnings.

[0077] The atlas drawing module draws the time-series curve of students' physiological indicators, the behavioral radar chart, and the heat matrix chart. It inputs the low-frequency to high-frequency energy ratio of heart rate variability and the sleep structure index for 72 consecutive hours, performs sliding window smoothing on the low-frequency to high-frequency energy ratio of heart rate variability, eliminates outliers that deviate from the mean by three standard deviations within the window, uses a stepped broken line to mark the sleep stage transition moments for the sleep structure index, automatically marks red warning points during periods of sudden increase in awakening times, generates a time-series curve of physiological indicators reflecting the dynamic changes of physiological rhythms, normalizes the social avoidance index, behavioral regularity deviation, standardized value of library stay duration, and average daily social contact frequency to the [0, 1] interval and then maps them to the polar coordinate system, represents the feature intensity with the radius length, draws concentric circles to mark the group mean and the range of ±1 standard deviation, generates a behavioral radar chart, presents the degree of deviation of individual behavioral patterns from the group benchmark, counts the number distribution of each warning level by class, calculates the spatial density based on Gaussian kernel density estimation, defines green as low density, yellow as medium density, and red as high density, generates a two-dimensional heat matrix chart of class - warning level, and identifies the areas where group psychological risks gather, supporting penetration and drill-down analysis.

[0078] First, deploy Bluetooth beacons, RFID turnstiles, and distribute smart bracelets within the campus, complete the device networking and data interface configuration. Then, the system automatically collects physiological data such as students' heart rate variability and body movement frequency, as well as trajectory data such as social contact frequency and library stay duration, and performs cleaning and time period merging on outliers. Then, through the feature extraction module, calculate physiological rhythm features such as heart rate band energy ratio and sleep quality index, and behavioral pattern features such as social avoidance index and schedule deviation value, and input them into the physiological time-series analysis model and the behavioral atlas analysis model to generate the psychological stress index and social behavior abnormality degree respectively. Subsequently, the mental health behavior analysis model combines the two indices to output the mental health risk index. When the mental health risk index exceeds the dynamic threshold, automatically push graded warning information to the teacher terminal. Finally, the individual physiological trend, behavioral radar chart, and class risk heat map can be viewed through the visualization interface. The psychological counselor marks the feedback results after intervention, and the system optimizes the weights of the mental health behavior analysis model and periodically updates the threshold accordingly, forming a closed-loop management process of collection - analysis - warning - feedback.

[0079] As described above, this is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

Claims

1. An AI-driven early warning system for student behavior analysis, characterized in that: It includes behavior data collection module, behavior feature extraction module, psychological analysis module, psychological pre-diagnosis module, mental health management module and map drawing module; The behavior data collection module collects and preprocesses students' physiological data and spatial trajectory data; The behavior feature extraction module extracts physiological rhythm features and behavior pattern features of the preprocessed physiological data and spatial trajectory data respectively; The psychological analysis module constructs a physiological time series analysis model and a behavioral spectrum analysis model based on physiological rhythm characteristics and behavioral pattern characteristics, and outputs a psychological stress index and a degree of abnormal social behavior, respectively; The psychological pre-diagnosis module combines the psychological stress index and the degree of abnormal social behavior to construct a psychological health behavior analysis model and output a psychological health risk index; The mental health management module establishes a hierarchical early warning response system based on the mental health risk index and performs closed-loop feedback; The graph drawing module draws the student's physiological index time series curve, behavior radar chart and thermal matrix chart.

2. An AI-driven student behavior analysis and early warning system according to claim 1, characterized in that: In the behavior data collection module, the collection and preprocessing process of the student's physiological data and spatial trajectory data includes: Students wear a smart bracelet that integrates a photoplethysmographic sensor and a three-axis accelerometer. The photoplethysmographic sensor emits green light with a wavelength of 520nm to penetrate the epidermal tissue, detects capillary blood flow pulsation through changes in reflected light intensity, and obtains students' heart rate variability data. A decision tree model is built based on body movement frequency and heart rate variability standard deviation to obtain students' deep sleep, light sleep, and wakefulness duration data; Bluetooth 5.0 beacons are deployed every 5 meters in the school area, and dual-band RFID readers are installed at the smart gate terminals at the entrance and exit of the library. When the signal strength of two Bluetooth 5.0 beacons exceeds -70dBm and lasts for more than 60 seconds, the effective social contact event record is triggered, and the frequency of students' social contacts is aggregated and counted through time windows. At the same time, the length of time students stay in the library is obtained based on the gate entry and exit timestamps; The collected heart rate variability data was filtered using a moving average filter to eliminate motion artifacts, combined with a three-times standard deviation threshold to eliminate abnormal pulse signals and interpolated to complete the data. According to the constructed decision tree model, the students' sleep stages are divided into deep sleep if the body movement frequency is ≤0.1Hz and the standard deviation of heart rate variability is <5ms. If the body movement frequency is 0.1-0.5Hz and the standard deviation of heart rate variability is 5-15ms, it is judged as light sleep. If the body movement frequency is >0.5Hz and the standard deviation of heart rate variability is ≥15ms, it is judged as wakefulness. The screened effective social contact events are normalized independently according to the spatial scene; Abnormal data in which the length of stay in the library exceeds 12 hours at a time is eliminated, and the entry and exit records within 30 consecutive minutes are merged. The intersection time of the valid time periods is accumulated, and the upper limit of the length of stay in the library per day is set at 14 hours. Standardized calculation of the length of stay is also performed.

3. An AI-driven student behavior analysis and early warning system according to claim 2, characterized in that: In the behavior feature extraction module, the process of extracting the physiological rhythm features of the pre-processed physiological data includes: Physiological rhythm characteristics include heart rate variability low-frequency and high-frequency energy ratio characteristics and sleep structure index characteristics; Perform power spectral density analysis on the preprocessed heart rate variability data. The power spectral density analysis process includes converting the time-domain signal into a frequency-domain distribution using the fast Fourier transform, obtaining the low-frequency band energy integral and the high-frequency band energy integral, and using the ratio of the low-frequency band energy integral to the high-frequency band energy integral as the heart rate variability low-frequency to high-frequency energy ratio feature; Based on the sleep stage discrimination results, count the total sleep duration and the deep sleep proportion, introduce the wake-up times correction term, and obtain the sleep structure index feature.

4. An AI-driven student behavior analysis and early warning system according to claim 3, characterized in that: In the behavior feature extraction module, the process of extracting the behavior pattern features of the preprocessed spatial trajectory data includes: The behavior pattern features include the social avoidance index feature and the regularity deviation feature; Based on the normalized social contact frequency data, count the seven-day average contact frequency, obtain the maximum contact frequency benchmark, and extract the social avoidance index feature; Extract the time series of the first arrival at the library every day from the preprocessed library residence duration data. Based on the time series of the first arrival at the library every day, calculate the historical average time benchmark of the first arrival at the library, and quantify the average absolute deviation to obtain the behavior regularity deviation feature.

5. An AI-driven student behavior analysis and early warning system according to claim 4, characterized in that: In the psychological analysis module, the process of constructing a physiological time series analysis model and outputting the psychological stress index includes: Adopt an eight-layer dilated causal convolutional structure, with the width of each convolutional kernel fixed at 3, the dilation factor increasing exponentially, and every two layers forming a residual block. Retain the underlying features through skip connections, use the ReLU activation function and batch normalization to accelerate convergence, and input a two-dimensional time series matrix composed of the heart rate variability low-frequency to high-frequency energy ratio and the sleep structure index for 72 consecutive hours to construct a physiological time series analysis model; The first seven layers of the physiological time series analysis model gradually extract multi-scale physiological rhythm features. The eighth layer outputs a high-dimensional time series feature matrix, performs global max pooling along the time dimension to extract the key pattern, outputs 128-dimensional features, maps the 128-dimensional features to the [0,1] interval through a fully connected layer, and outputs the psychological stress index through the Sigmoid function; Use the smooth L1 loss function to balance the influence of outliers, set the initial learning rate to 0.001 and decay it by 50% periodically, initialize the convolutional kernel weights using the He normal distribution, and set the batch size and gradient clipping threshold.

6. An AI - driven student behavior analysis and early warning system according to claim 5, characterized in that: In the psychological analysis module, the process of constructing a behavior graph analysis model and outputting the social behavior abnormality degree includes: Adopt a 3-layer graph attention network architecture, define each student as a node in the graph attention network architecture, use the social avoidance index feature and the behavior regularity deviation feature as the node feature vectors, calculate the ratio of the social contact frequency to the global maximum as the edge weight based on the actual social contact frequency among students, and retain the top 20% of the edges with edge weights to construct a sparse adjacency matrix; Design a 3-layer graph attention network architecture. The first layer linearly transforms the node feature vector to 8-dimensional space. The second layer obtains the attention coefficient between nodes through the LeakyReLU activation function and weightedly aggregates the neighborhood information. The third layer introduces residual connection and layer normalization to build a behavior graph analysis model. The social behavior abnormality degree is output through the fully connected layer mapping. The output dimension of each layer is kept at 8 dimensions and normalized to the [0,1] interval through the Sigmoid function. The contrast loss function is used to enhance the sensitivity of anomaly detection. The interval parameter δ=0.3 is set to distinguish normal and abnormal samples. The sparse adjacency matrix is ​​recalculated every week based on the new social contact frequency data, and the graph attention network architecture is updated. The mean μ and standard deviation σ are calculated based on the data distribution of historical social behavior abnormality. The dynamic threshold is set to μ+2σ. When the abnormality of social behavior exceeds μ+2σ, it is judged as abnormal behavior. When the abnormality of social behavior is lower than μ+2σ, it is judged as normal behavior. A 4-head parallel attention mechanism is configured, the initial learning rate is set to 0.0005, the inter-layer dropout rate is set to 0.3, and the domain range is expanded to the second-order neighbors.

7. An AI-driven student behavior analysis and early warning system according to claim 6, characterized in that: In the psychological pre-diagnosis module, the process of building a psychological health behavior analysis model by combining the psychological stress index and the degree of abnormal social behavior includes: The psychological stress index P1 is modeled as a continuous empirical distribution, and the probability mass distribution is constructed by accumulating the Dirac function. The social behavior abnormality P2 is discretized into ten equal parts, and the proportion of students in each quantile is counted to form a discrete probability distribution, and a mental health behavior analysis model is constructed. Construct an exponential cost function based on the distribution differences between the psychological stress index and the social behavior abnormality degree, and define the cross-modal transmission cost C ij , and solve the optimal transport plan matrix γ through the entropy-regularized Sinkhorn iteration algorithm ij , the regularization coefficient ε is set to 0.1, and the number of iterations is set to 50 times; Perform the Hadamard product operation on the psychological stress index and the abnormal degree of social behavior using the optimal transport plan matrix, and generate a joint feature tensor by combining the outer product operation Execute the maximum pooling operation along the discrete modal dimension, extract the feature combination with the strongest cross-modal correlation, and output the fused feature vector after dimensionality reduction 8. An AI-driven student behavior analysis and early warning system according to claim 7, characterized in that: The process of outputting the mental health risk index of the psychological pre-diagnosis module includes: Design a two-layer perceptron network with a 64-dimensional hidden layer, achieve non-linear feature transformation through the ReLU activation function, output the normalized mental health risk index Y using the Sigmoid function, introduce a dynamic calibration mechanism, and based on the mean μ Y and standard deviation σ Y of the mental health risk index over the past 30 days, calculate the baseline offset Y′, and perform linear scaling and offset on the real-time output.

9. An AI-driven student behavior analysis and early warning system according to claim 8, characterized in that: In the mental health management module, the process of establishing a hierarchical early warning response system and conducting closed-loop feedback includes: The mean μ of the 30-day historical mental health risk index Y and the standard deviation σ Y , set the dynamic mental health risk threshold B t = μ Y + k·σ Y , where the coefficient k is set according to the observation level k = 1, the intervention level k = 1.5, and the emergency level k = 2, and calculate respectively and like If the behavior continues for 24 hours, it is considered as observation level, high-frequency data collection mode is activated, and a behavior pattern analysis report is generated and sent to the class teacher's workstation. If the abnormality persists for 12 hours, it is considered as intervention level, the psychological counseling appointment interface is activated, the last three available appointment time slots are inserted, and a consultation suggestion containing details of abnormal characteristics is pushed to the psychological counselor. If the situation is judged as an emergency, the campus security system will be linked to locate the student's real-time location, generate a crisis intervention plan, and send a red alert to the class teacher, psychological counseling center, and parents and guardians; Through the binary labeling of the effectiveness of early warning by psychological counselors, a supervisory signal feedback mechanism is established. When the decline in the mental health risk index after intervention exceeds 0.5 times the historical standard deviation, the labeling is effective. When the decline in the mental health risk index after intervention does not exceed 0.5 times the historical standard deviation, the labeling is invalid. The labeled data is used to update the weight parameters of the mental health behavior analysis model in real time, and the incremental learning algorithm driven by supervisory signals is used to adjust the contribution ratio of physiological rhythm characteristics and behavioral pattern characteristics, and the mental health risk threshold is adjusted based on the proportion of effective early warnings.

10. An AI-driven student behavior analysis and early warning system according to claim 9, characterized in that: In the graph drawing module, the process of drawing the student's physiological index time series curve, behavior radar chart and heat matrix chart includes: The low-frequency and high-frequency energy ratio of heart rate variability and the sleep structure index of 72 consecutive hours were input, and the low-frequency and high-frequency energy ratio of heart rate variability was smoothed by sliding window. The outliers that deviated from the mean by three times the standard deviation in the window were eliminated. The sleep structure index was marked with a step-shaped broken line to mark the sleep stage transition time, and the red warning point was automatically marked during the period of sudden increase in the number of awakenings, so as to generate a physiological indicator time series curve reflecting the dynamic changes of physiological rhythms. After normalizing the social avoidance index, behavioral regularity deviation, standardized value of library residence duration, and average daily social contact frequency to the [0, 1] interval, they are mapped to the polar coordinate system. The feature intensity is represented by the radius length, and concentric rings are drawn to mark the group mean and the range of ±1 standard deviation, generating a behavioral radar chart to present the degree of deviation of individual behavior patterns from the group benchmark; Count the number distribution of each warning level by class, calculate the spatial density based on Gaussian kernel density estimation, define green as low density, yellow as medium density, and red as high density, generate a two-dimensional heat matrix chart of class-warning level, identify the areas where group psychological risks gather, and support penetration drilling analysis.

Images