An intelligent education management method and system based on the Internet of Things

Abstract

The application relates to an intelligent education management method and system based on an Internet of Things. The method comprises the following steps: inputting a multi-channel physiological feature sequence into a hierarchical probability state space model to obtain a short-term cognitive state and a short-term cognitive state posterior probability; performing stability statistical testing on the short-term cognitive state posterior probability, and marking the short-term cognitive state as a phase-established state when the stability condition is met; constructing a cost function based on the phase-established state and a probability evolution track, and performing rolling time domain optimization based on the minimum cost function and a dynamic refractory period constraint to obtain an optimal learning action; calculating a decision oscillation index according to a current executed learning action sequence, and increasing the weight of an action change amplitude penalty and prolonging the locking time of the dynamic refractory period constraint by a set step length when the decision oscillation index exceeds a preset first oscillation threshold. The method can guarantee the continuity of a learning process and the stability of a cognitive state.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, and in particular relates to an intelligent education management method and system based on the Internet of Things. Background Technology

[0002] With the development of technologies in the fields of the Internet of Things and wearable devices, technologies have emerged that use sensors to monitor learners’ cognitive and emotional states in real time and adaptively adjust the difficulty, pace and presentation of learning content accordingly.

[0003] In traditional technologies, decision-making for learning paths mainly relies on raw, high-frequency data collected by sensors to drive decisions. This involves capturing raw physiological indicators such as specific frequency bands of EEG energy, skin conductance levels, or heart rate variability, and directly mapping them to the learner's cognitive and emotional state, thereby triggering adjustments to the learning content. However, when raw data fluctuates drastically and causes decision-making anomalies, existing approaches often resort to low-pass filtering or reducing the decision update frequency in an attempt to mitigate the impact of data fluctuations.

[0004] However, when the above methods directly drive decision-making using raw high-frequency data, the physiological signals are highly non-stationary. Even slight body movements, external noises, or momentary lapses in concentration can cause drastic instantaneous fluctuations in the data. Simple adaptive algorithms may overfit to these noises, leading to decision oscillations and frequent switching of learning content. This not only disrupts the continuity of learning and increases cognitive confusion for students, but also forms an unstable positive feedback loop, causing the system to completely lose control. Furthermore, low-pass filtering used to deal with fluctuations introduces phase delay, impairing the system's real-time response to sudden changes in the actual cognitive state. Simply reducing the decision update frequency discards the effective information contained in the high-frequency data, weakening the core value of real-time intervention. Summary of the Invention

[0005] Therefore, it is necessary to provide an IoT-based intelligent education management method and system that can solve the overfitting and decision oscillation problems caused by high-frequency physiological signals, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides an intelligent education management method based on the Internet of Things, including:

[0007] The system acquires multimodal physiological signals synchronously collected by sensors worn by students, and performs adaptive artifact suppression on the multimodal physiological signals based on individual physiological baselines to obtain multichannel physiological feature sequences.

[0008] The multi-channel physiological feature sequence is input into the hierarchical probabilistic state space model to obtain the short-term cognitive state and the short-term cognitive state posterior probability. During the inference of the short-term cognitive state posterior probability, the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequence.

[0009] A stability statistical test based on a continuous time window is performed on the posterior probability of short-term cognitive state. When the posterior probability of short-term cognitive state satisfies the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a staged established state.

[0010] The historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence are input into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a preset time period in the future.

[0011] Based on the phased establishment of states and probabilistic evolution trajectories, a cost function is constructed that includes the cost of state deviation and the penalty for the magnitude of action change. Then, based on minimizing the cost function and the dynamic refractory period constraint, rolling time-domain optimization is performed to obtain the optimal learning action for the current decision cycle.

[0012] The currently executed learning action sequence is updated by executing the optimal learning action, and the decision oscillation index is calculated based on the updated currently executed learning action sequence. When the decision oscillation index exceeds the preset first oscillation threshold, the weight of the penalty for the change amplitude of the action is increased by setting the step size, and the locking time of the dynamic refractory period constraint is extended.

[0013] Secondly, this application also provides an IoT-based intelligent education management system, including:

[0014] The physiological data module is used to acquire multimodal physiological signals synchronously collected by sensors worn by students, and to perform adaptive artifact suppression on the multimodal physiological signals based on individual physiological baselines to obtain multi-channel physiological feature sequences.

[0015] The short-term cognitive state module is used to input multi-channel physiological feature sequences into the hierarchical probabilistic state space model to obtain short-term cognitive states and short-term cognitive state posterior probabilities. During the inference of the short-term cognitive state posterior probability, the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequences.

[0016] The stage cognitive state module is used to perform a stability statistical test on the posterior probability of short-term cognitive state based on a continuous time window. When the posterior probability of short-term cognitive state meets the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a stage-established state.

[0017] The evolution prediction module is used to input the historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a preset time period in the future.

[0018] The learning action solution module is used to construct a cost function that includes the cost of state deviation and the penalty for the magnitude of action change based on the phased established state and the probabilistic evolution trajectory. Based on minimizing the cost function and the dynamic refractory period constraint, it performs rolling time-domain optimization to solve for the optimal learning action in the current decision cycle.

[0019] The oscillation suppression module is used to update the currently executed learning action sequence by executing the optimal learning action, and calculate the decision oscillation index based on the updated currently executed learning action sequence. When the decision oscillation index exceeds the preset first oscillation threshold, the weight of the penalty for the change amplitude of the action is increased by setting the step size and the locking time of the dynamic refractory period constraint is extended.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned Internet of Things-based intelligent education management methods.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described IoT-based intelligent education management methods.

[0022] The aforementioned IoT-based intelligent education management method and system actively weakens the interference of transient spikes on state estimation when inferring short-term cognitive states by inputting multimodal physiological features after adaptive artifact suppression into a hierarchical probabilistic state-space model that can dynamically adjust the observation noise scale according to the feature change amplitude. Through continuous time window stability statistical testing, a state is established as a triggerable decision only when the posterior probability of the state continuously meets the conditions, further filtering out transient fluctuations. Simultaneously, it combines an individualized recurrent neural network prediction model to anticipate the evolution of future cognitive states. Based on this, a cost function is constructed that includes state deviation costs and action change amplitude penalties, and a dynamic refractory period triggered by the previous significant action is applied. The learning path is optimized in a rolling time domain, ensuring that each adjustment takes into account both current and future cognitive needs while proactively avoiding frequent and drastic switching. By monitoring the decision oscillation index of the action sequence in real time, the penalty weight for action changes is adaptively increased and the refractory period lock-in time is extended when oscillation tendencies are observed. This forms a complete technical chain from signal acquisition and processing to state inference, decision optimization, and closed-loop stability control. While fully leveraging the ability of real-time physiological sensing to dynamically adjust the learning path, it effectively suppresses overfitting and decision oscillations caused by high-frequency physiological noise, ensuring the continuity of the learning process and the stability of the cognitive state. This avoids the risk of positive feedback runaway caused by the additional pressure on students due to the system's own oscillations. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the IoT-based intelligent education management method of the present invention.

[0025] Figure 2 This is a structural diagram of the IoT-based intelligent education management system of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] In one embodiment, such as Figure 1As shown, an IoT-based intelligent education management method is provided. This embodiment illustrates the application of this method to an intelligent education management terminal. It is understood that this method can also be applied to an intelligent education management server, and further to an intelligent education management system including both an intelligent education management terminal and an intelligent education management server, and is implemented through the interaction between the intelligent education management terminal and the intelligent education management server. In this embodiment, the method includes the following steps:

[0028] S101. Acquire multimodal physiological signals synchronously collected by the sensors worn by the student, and perform adaptive artifact suppression on the multimodal physiological signals based on the individual physiological baseline to obtain a multi-channel physiological feature sequence.

[0029] Indicatively, the intelligent education management terminal can establish a stable Internet of Things (IoT) communication connection with multimodal sensors worn by students. These sensors may include an EEG cap, a skin conductance sensor, and a photoplethysmography (PPG) sensor. Each sensor synchronously collects the student's multimodal physiological signals according to a preset sampling rate. These signals include EEG, skin conductance, and pulse wave signals. Before acquiring the multimodal physiological signals, the intelligent education management terminal pre-constructs and stores an individual physiological baseline. This baseline is formed by collecting multimodal physiological data from the student at rest, preprocessing it, and calculating the mean, variance, and higher-order statistics of each feature channel. This baseline is continuously updated online during the learning process. The intelligent education management terminal extracts features from the artifact-suppressed physiological signals, aligns the extracted features temporally, and then concatenates them to form a multi-channel physiological feature sequence.

[0030] Specifically, the resting state does not require absolute stillness, but rather a seated, calm posture consistent with the natural learning readiness state. For the acquired raw EEG signals, morphological filtering is used to identify and label blinking and muscle activity artifacts. Then, through independent component analysis based on individual data, the artifact-free EEG is decomposed into several independent components. Based on the dipole characteristics of the scalp topography of each component, neural source components are automatically retained, while residual electromyographic and electrooculographic components are removed. For example, the mathematical expression of independent component analysis can be represented as: ,in, This is the original EEG signal matrix. It is a mixed matrix. It is an independent component matrix. The noise matrix is ​​solved by the intelligent education management terminal through an optimization algorithm. and To achieve signal separation, a convex optimization separation algorithm is used to decompose the electrodermal signal into a slowly varying tension component and a rapidly varying phase component. The objective function of the convex optimization separation can be expressed as: ,in This is the original electrodermal signal. It is a tension component. As a phase component, and For regularization parameters, Let the total variation norm be . The L1 norm is used. Successive heartbeat intervals are extracted from the pulse wave signal. After ectopic heartbeat detection and interpolation correction, the time and frequency domain indices of heart rate variability are calculated. Based on the cleaned data of this resting period, the mean, variance, and higher-order statistics of each feature channel are calculated to form an individual physiological baseline vector. During the learning process, whenever the system detects that the learner is in a state of continuous focus and low workload exceeding a preset duration, it will recursively integrate the currently cleaned features into the baseline model. The fusion rate is controlled by a learning rate parameter to ensure that the baseline can track the individual's physiological drift without being contaminated by transient disturbances.

[0031] S102. Input the multi-channel physiological feature sequence into the hierarchical probabilistic state space model to obtain the short-term cognitive state and the short-term cognitive state posterior probability. During the inference of the short-term cognitive state posterior probability, the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequence.

[0032] Furthermore, the intelligent education management terminal pre-constructs a hierarchical probabilistic state space model, which is clearly divided into a bottom-level instantaneous physiological signal layer, an intermediate short-term cognitive state layer, and a top-level stable learning trait layer. The intelligent education management terminal inputs multi-channel physiological feature sequences frame by frame into the hierarchical probabilistic state space model. Each frame in the multi-channel physiological feature sequence corresponds to a multi-dimensional instantaneous physiological feature vector at a time step. Optionally, the short-term cognitive state is a discrete, persistent hidden variable whose value is composed of a set of predefined cognitive state labels. This is achieved by collecting a large amount of multimodal physiological signal data covering students with different knowledge bases, age levels, and learning traits in a standardized learning scenario, and simultaneously matching professionally labeled samples of various cognitive states to form a group benchmark dataset. A group mixed-effects model integrating fixed and random effects is used for modeling and analysis. The group mixed-effects model is a hierarchical statistical fitting architecture. Fixed effects are used to characterize the correlation between common physiological characteristics and cognitive states of all learners, while random effects are used to characterize the physiological responses of different learners. The differentiated offset feature can take into account both the general association patterns of the group and the physiological heterogeneity of individuals. After inputting the group benchmark dataset into the group mixed-effects model, the internal parameters of the model are solved iteratively by relying on maximum likelihood estimation. Based on each predefined cognitive state label, the mean vector, covariance matrix and higher-order statistical distribution parameters of the physiological features corresponding to each cognitive state are fitted and solved separately, thereby obtaining the benchmark statistical distribution of physiological features in each state. In the early stage of individual use, a small number of samples labeled with the individual's corresponding cognitive state and physiological features are used to fine-tune and calibrate the statistical distribution parameters of physiological features obtained from the group fitting through transfer learning, adapting to the differences in physiological response characteristics of individual individuals. The intelligent education management terminal realizes the inference from multi-channel physiological feature sequences to short-term cognitive state posterior probabilities through a Bayesian filter with time-dependent dependence. Its recursive update formula can be expressed as: ,in, for Short-term cognitive state of time The posterior probability, for Multi-channel physiological feature vectors at time points. To observe the likelihood, For the elements of the state transition matrix, for Short-term cognitive state of time The posterior probability. To avoid overfitting to instantaneous physiological noise, the observation likelihood adopts an adaptive adjustment method, the expression of which is: ,in, Represents a multidimensional Gaussian distribution. Short-term cognitive state The corresponding physiological characteristic mean vector, Short-term cognitive state The corresponding covariance matrix basis, An adaptive noise scale. The update is dynamically updated based on the magnitude of physiological feature changes in the multi-channel physiological feature sequence. The update formula is as follows: ,in, for Adaptive noise scaling at any given time. The learning rate parameter, It is the difference between the moving average of physiological characteristic changes over several consecutive frames and the baseline fluctuation threshold. To adjust the constant, and These are the minimum and maximum thresholds for adaptive noise scaling, respectively. The function is used to... The constraints are within a preset range. The intelligent education management terminal calculates frame by frame using the above formula to obtain the short-term cognitive state and the corresponding posterior probability of the short-term cognitive state.

[0033] S103. Perform a stability statistical test on the posterior probability of the short-term cognitive state based on a continuous time window. When the posterior probability of the short-term cognitive state meets the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a staged established state.

[0034] Optionally, the intelligent education management terminal is equipped with a state stabilization judge to perform stability statistical tests on the posterior probability of short-term cognitive states. The test process employs a cumulative sum test approach to avoid repeated state switching triggered within a short period due to slight fluctuations in the posterior probability of short-term cognitive states. The intelligent education management terminal maintains a cumulative statistic for each predefined short-term cognitive state label, and the update formula for the cumulative statistic is as follows: ,when hour, ,in, for Time of the first The cumulative statistics of short-term cognitive states for Time of the first The cumulative statistics of short-term cognitive states for Time of the first The posterior probability of a short-term cognitive state. This is the upper bound threshold for the posterior probability. This is the lower bound threshold for the posterior probability. The stability condition is set when the cumulative statistic crosses the preset decision boundary, i.e., when... At that time, the judgment of the first A short-term cognitive state satisfies a stability condition within a continuous time window, wherein... The length of the continuous time window is determined by a preset number of frames, serving as the preset decision boundary threshold. The intelligent education management terminal continuously monitors the cumulative statistics of each short-term cognitive state. When the cumulative statistics of a short-term cognitive state meet the above stability conditions, the short-term cognitive state is marked as a staged establishment state. Optionally, if after marking a staged establishment state, the intelligent education management terminal detects another short-term cognitive state establishment within a short period of time, and there is a significant cognitive demand conflict between the two staged establishment states, it is considered a potential precursor to oscillation. The intelligent education management terminal will suppress the marking of the latter staged establishment state to ensure that the marked staged establishment state is robust and supported by sufficient evidence, rather than an illusion caused by transient noise.

[0035] S104. Input the historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a preset time period in the future.

[0036] Specifically, the intelligent education management terminal pre-deploys an individualized recurrent neural network prediction model. This individualized recurrent neural network prediction model can adopt an encoder-decoder architecture, with the basic computing unit of the architecture being a gated recurrent unit (GRU), which balances the ability to capture temporal features with the need for lightweight models, and is suitable for the edge computing deployment scenarios of the intelligent education management terminal. The encoder employs a three-layer stacked structure, with 128 GRU units in each layer. The ReLU activation function is uniformly used in all hidden layers. The input layer dimension strictly matches the high-dimensional input vector dimension concatenated from the multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence, and the currently executed learning action sequence. After each GRU unit output, it passes through a layer normalization module with the normalization axis set to the hidden layer feature dimension and a regularization dropout module with a fixed dropout probability of 0.2, effectively suppressing model overfitting and accelerating training convergence. The decoder also employs a three-layer stacked GRU structure, with 128 GRU units in each layer. The ReLU activation function is used in all hidden layers, and the Softmax activation function is used in the output layer to map the hidden states of the last layer of the decoder to the posterior probability distribution of each short-term cognitive state. The encoder, through three layers of temporal coding operations, progressively compresses the input historical multimodal long sequence into a 128-dimensional global context vector. This vector contains all the key information about historical physiological characteristics, cognitive states, and learning actions. The decoder uses this global context vector as its initial input and employs an autoregressive, time-series generation method to iteratively output the short-term cognitive state probability distribution for each future time step, forming a complete probability evolution trajectory. The overall encoder-decoder architecture first completes global pre-training based on temporal data of physiological characteristics, labeled data of cognitive states, and sample data of learning actions from a large-scale student population to obtain general prototype network weights. Then, through a meta-learning algorithm, it uses a small number of individual cognitive state and physiological characteristic labeled samples from the initial stage of use to fine-tune and calibrate the network weights, quickly adapting to the correlation between the physiological characteristic fluctuation patterns and cognitive state evolution patterns of specific students, achieving accurate prediction of individualized cognitive state probability evolution trajectories. The intelligent education management terminal extracts the historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence within a preset time window. The above three types of sequences are used as inputs to the individualized recurrent neural network prediction model. Among them, the historical multi-channel physiological feature sequence is a set of multi-channel physiological feature vectors collected frame by frame in the past period, the historical short-term cognitive state posterior probability sequence is a set of short-term cognitive state posterior probabilities calculated frame by frame in the corresponding time period, and the currently executed learning action sequence is a sequence of learning actions executed before the current decision cycle.The encoder of the individualized recurrent neural network prediction model compresses the input three types of multimodal sequences into a context vector, and the decoder generates the short-term cognitive state posterior probability of each time step within a preset future time period in an autoregressive manner, forming a probability evolution trajectory, the expression of which can be: ,in, For individualized recurrent neural network prediction models, These are the model weight parameters. Preset time period for the future The probabilistic evolution trajectory of short-term cognitive states within the period. This is a historical multi-channel physiological characteristic sequence. This is a posterior probability sequence of historical short-term cognitive states. This is the currently executed sequence of learning actions. The length of the historical sequence. Predict the time domain length for the future.

[0037] The intelligent education management terminal can use the cross-entropy loss function to train the individualized recurrent neural network prediction model online. The loss function formula is as follows: ,in, for Time of the first The true posterior probability of a short-term cognitive state For model predictions Time of the first The posterior probability of a short-term cognitive state. The number of short-term cognitive states. For the weight decay regularization term coefficient, represents the L2 norm of the model weights. The intelligent education management terminal uses the Adam optimizer to incrementally update the model weights via online stochastic gradient descent, ensuring that the model can continuously capture the unique state change patterns of students and ultimately output the probability evolution trajectory of each short-term cognitive state within a preset future time period.

[0038] S105. Based on the phased established state and probabilistic evolution trajectory, construct a cost function that includes the cost of state deviation and the penalty for the magnitude of action change. Then, perform rolling time-domain optimization based on minimizing the cost function and dynamic refractory period constraints to obtain the optimal learning action for the current decision cycle.

[0039] For example, the intelligent education management terminal uses a phased established state as the benchmark for the current cognitive state, and combines it with a probabilistic evolution trajectory to construct a cost function to quantify the decision costs corresponding to different learning action sequences. Its expression is: ,in, The cost of overall decision-making To predict the length of the time domain, KL divergence is used to calculate... Probability distribution of short-term cognitive states predicted at any time Distribution with ideal state The degree of deviation between them constitutes the state deviation cost; The weighting coefficient for the penalty of the magnitude of the change in motion. and They are respectively Time and Learning actions at all times The penalty for the magnitude of the change in adjacent learned actions is the square of the change in the action's magnitude. Weighting coefficients for matching rewards to traits. For students' stable learning trait vectors, Let $\frac{ ...

[0040] Furthermore, when solving for the minimum cost function, the intelligent education management terminal introduces a dynamic refractory period constraint, which can be expressed as follows: ,in, This is a marker of the dynamic refractory period. This indicates that it is in the dynamic refractory period. This indicates that the period is outside the dynamic refractory period. The learning action for the current decision-making cycle. The learning actions performed in the previous decision-making cycle. The threshold for action difference is defined. The intelligent education management terminal adopts a rolling time-domain optimization strategy, treating the adjustment of the learning path as a finite-time-domain optimal control problem with stability constraints. It solves the above-mentioned constrained cost function minimization problem by using the direct shooting method or Monte Carlo tree search method. After discretizing the action space, it finds the current optimal first step of learning action as the optimal learning action for the current decision cycle.

[0041] S106. Update the currently executed learning action sequence by executing the optimal learning action, and calculate the decision oscillation index based on the updated currently executed learning action sequence. When the decision oscillation index exceeds the preset first oscillation threshold, increase the weight of the action change amplitude penalty by setting the step size and extend the locking time of the dynamic refractory period constraint.

[0042] Indicatively, the intelligent education management terminal controls the intelligent learning terminal to execute the optimal learning action for the current decision-making cycle. Simultaneously, it adds this optimal learning action to the end of the currently executed learning action sequence, updating the sequence. The terminal also records the type of the optimal learning action, the short-term cognitive state posterior probability before and after triggering, the probability evolution trajectory, and the decision cost. The intelligent education management terminal has an independent stability monitoring module used to calculate the decision oscillation index based on the updated sequence of currently executed learning actions. The decision oscillation index quantifies the degree of oscillation in recent decision-making actions, and its calculation formula is as follows: ,in, Let be the decision oscillation index at time t. The length of the sliding window. The attenuation coefficient is... for Time and Continuously learn the measurement of differences in movements. for Learning actions performed at all times for The learning actions are executed continuously. The intelligent education management terminal compares the calculated decision oscillation index with a preset first oscillation threshold, which is a pre-defined critical value used to determine whether slight oscillations have occurred. When the decision oscillation index exceeds the preset first oscillation threshold, it indicates that the current decision action has shown a slight oscillation trend. The intelligent education management terminal then increases the weighting coefficient of the action change amplitude penalty by setting a step size. Increasing the weighting coefficient of the penalty for significant action changes can suppress frequent and large-scale action switching. Simultaneously, extending the lock-in time of the dynamic refractory period constraint by setting a step size avoids executing learning actions with excessively large differences within a short period, thereby enhancing the system's inertia and suppressing further exacerbation of oscillation trends. After adjusting the weighting coefficient of the penalty for significant action changes and the lock-in time of the dynamic refractory period constraint, the intelligent education management terminal continues to monitor changes in the decision oscillation index to ensure the stability of decision actions and cut off the positive feedback vicious cycle where high-frequency noise distorts students' physiological states through oscillations, which in turn exacerbates the oscillations.

[0043] In the aforementioned IoT-based intelligent education management method, multimodal physiological signals synchronously collected by sensors worn by students are acquired. Adaptive artifact suppression is then applied based on individual physiological baselines to obtain multi-channel physiological feature sequences. These sequences are input into a hierarchical probabilistic state-space model that dynamically adjusts the observation noise scale according to the magnitude of physiological feature changes, yielding short-term cognitive states and posterior probabilities. A continuous-time-window stability statistical test is performed on the posterior probabilities of the short-term cognitive states to determine the staged establishment state. Finally, historical multi-channel physiological feature sequences, corresponding short-term cognitive state posterior probability sequences, and currently executed learning action sequences are input into an individualized recurrent neural network prediction model to generate the future short-term cognitive state probability evolution. The trajectory is constructed based on the established state and probabilistic evolution trajectory, which includes the cost of state deviation and the penalty for the magnitude of action change. Combined with the dynamic refractory period constraint, the optimal learning action for the current decision cycle is solved through rolling time domain optimization. The optimal learning action is executed to update the action sequence and calculate the decision oscillation index. When the index exceeds the preset first oscillation threshold, the penalty weight for the magnitude of action change is increased and the dynamic refractory period lock time is extended. Ultimately, this achieves the goal of eliminating decision oscillation, ensuring a smooth and continuous learning process, balancing the system's rapid response capability to changes in cognitive state with decision stability, improving learning efficiency and learning experience, resolving the contradiction between real-time performance and stability, and realizing accurate and robust assessment of students' cognitive state and adaptive adjustment of learning paths.

[0044] In one embodiment, adaptive artifact suppression is performed on multimodal physiological signals based on an individual's physiological baseline to obtain a multi-channel physiological feature sequence, including:

[0045] S21. Based on a sliding window, the temporal amplitude kurtosis and specific frequency band power spectral density of the EEG signal of multimodal physiological signals are calculated, and the comprehensive deviation of the temporal amplitude kurtosis and specific frequency band power spectral density relative to the individual physiological baseline is calculated. When the comprehensive deviation exceeds the set deviation threshold, the EEG signal is subjected to instantaneous artifact subtraction based on a time-varying Kalman smoother within the sliding window to obtain artifact-suppressed EEG signal; the process noise covariance of the time-varying Kalman smoother is adaptively adjusted with the comprehensive deviation.

[0046] In illustrative terms, the intelligent education management terminal uses a fixed-length sliding window to traverse the EEG signals of multimodal physiological signals. The length of the sliding window can be flexibly set according to the actual sampling rate of the acquired EEG signals to ensure that the complete waveform of transient artifacts can be captured. The intelligent education management terminal calculates the time-domain amplitude kurtosis of the EEG signal within each sliding window. The formula for calculating the time-domain amplitude kurtosis is as follows: ,in, For time-domain amplitude kurtosis, The EEG signal data within the sliding window, The mean value of the EEG signal within the sliding window. The time-domain amplitude kurtosis represents the mathematical expectation operation. It is used to quantify the steepness of the EEG signal waveform. The presence of artifacts can lead to significant abnormalities in the kurtosis value.

[0047] Furthermore, the intelligent education management terminal performs a Fourier transform on the EEG signals within the sliding window to calculate the power spectral density of a specific frequency band. This specific frequency band may include delta, theta, alpha, beta, and the low gamma band. The formula for calculating the power spectral density is as follows: ,in, For the power spectral density of a specific frequency band, This represents the number of sampling points for the EEG signal within the sliding window. The Fourier transform result of the EEG signal. This refers to the frequency value for a specific frequency band.

[0048] For example, the intelligent education management terminal calculates the combined deviation of time-domain amplitude kurtosis and power spectral density in a specific frequency band relative to an individual's physiological baseline. The formula for calculating the combined deviation is as follows: ,in, For the overall deviation, and These are the weighting coefficients for the time-domain amplitude kurtosis and the power spectral density in a specific frequency band, respectively, and satisfying the following conditions: , This is the time-domain amplitude kurtosis baseline value in an individual's physiological baseline. A set of frequencies within a specific frequency band. The corresponding frequency in the individual's physiological baseline The power spectral density reference value. The intelligent education management terminal compares the overall deviation with the set deviation threshold. When the overall deviation exceeds the set deviation threshold, it determines that there is a high probability of transient artifacts within the sliding window. At this time, a time-varying Kalman smoother is activated within the sliding window to subtract transient artifacts from the EEG signal. The observation equation is ,in, for The true state vector of the EEG signal at any given time. for The true state vector of the EEG signal at any given time. Here is the state transition matrix. This is the process noise vector. for Observation of EEG signal vectors at different times For the observation matrix, The observed noise vector. The process noise covariance matrix. The adjustment is adaptive based on the overall deviation, and the adjustment formula is as follows: ,in, for The process noise covariance matrix at time step 1. The process noise covariance reference matrix is... To adjust the coefficient, The overall deviation of the current sliding window is the largest. The larger the overall deviation, the larger the process noise covariance matrix, which makes the time-varying Kalman smoother more confident in the observation residuals, thereby quickly tracking and subtracting instantaneous artifacts, and finally obtaining the artifact-suppressed EEG signal.

[0049] S22. Based on the acceleration signal of sudden changes in body movement, identify and compress the amplitude of the skin conductance phase peak caused by sudden changes in body movement in the skin conductance signal of multimodal physiological signals to obtain the artifact-suppressed skin conductance signal.

[0050] Optionally, the intelligent education management terminal synchronously collects abrupt acceleration signals of body movements from an accelerometer worn by the student. This accelerometer can be integrated with a multimodal sensor or deployed independently to capture instantaneous changes in the student's body movements. The intelligent education management terminal performs threshold detection on the abrupt acceleration signals to determine if abrupt changes in body movements exist. The detection formula is as follows: ,in, for The acceleration signal value of sudden changes in body movement at any given moment. The acceleration threshold for sudden changes in body movement is defined as follows: when this formula is satisfied, a sudden change in body movement is determined to exist, and the timestamp of the change is recorded. Simultaneously, the intelligent education management terminal extracts the phase components of the electrodermal signal from multimodal physiological signals. The extraction method can employ a convex optimization separation algorithm to obtain the phase components of the electrodermal signal. Then, peak detection is performed on the phase components to identify the phase peaks of skin conductance. The formula for peak detection is... ,in, This serves as the timestamp for the appearance of the skin conductance phase peak. For peak detection time window, This represents the phase component of the skin conductance signal. The intelligent education management terminal matches the timestamp of the skin conductance phase peak with the timestamp of abrupt changes in body movement. If the difference between the two timestamps is within a preset time range, the skin conductance phase peak is determined to be an artifact peak caused by abrupt changes in body movement. For this artifact peak, the intelligent education management terminal uses an amplitude compression algorithm to compress its amplitude. The compression formula is as follows: ,in, The amplitude of the compressed skin conductance phase peak. The amplitude of the phase peak of the original skin conductance. The compression factor is 1. The absolute value of acceleration corresponds to the sudden change in body movement. The larger the absolute value of acceleration, the greater the compression amplitude, ensuring effective suppression of skin conductance artifacts caused by sudden changes in body movement, and finally obtaining the artifact-suppressed skin conductance signal.

[0051] S23. Extract successive heartbeat interval sequences from the photoplethysmography pulse wave signals of multimodal physiological signals, and detect and remove ectopic beats in the successive heartbeat interval sequences based on the dynamic threshold of median absolute deviation to obtain artifact-suppressed heart rate variability.

[0052] Furthermore, the intelligent education management terminal performs waveform analysis on the photoplethysmography (PPG) wave signal of multimodal physiological signals to identify feature points of each heartbeat cycle. These feature points can include peak or trough points of the pulse wave. By calculating the time interval between two adjacent feature points, a successive heartbeat interval sequence is extracted. The expression for the successive heartbeat interval sequence is as follows: ,in, This is a sequence of successive heartbeat intervals. For the first The first heartbeat cycle and the first The time interval between heartbeat cycles This represents the length of the successive heartbeat interval sequence.

[0053] The intelligent education management terminal calculates a dynamic threshold based on the absolute deviation of the median to detect ectopic beats in successive heartbeat interval sequences. Specifically, it calculates the median of the successive heartbeat interval sequence using the following formula: ,in, The median of the successive heartbeat interval sequence. This represents the function for calculating the median. Further, the absolute deviation of each heartbeat interval from the median is calculated, expressed as: ,in, For the first The absolute deviation of each heartbeat interval can be calculated, optionally, by calculating the median of the absolute deviations, i.e., the median absolute deviation, using the following formula: ,in, This represents the absolute deviation of the median. It is an absolute deviation sequence.

[0054] For example, the formula for calculating the dynamic threshold is as follows: ,in, For dynamic thresholds, This is a threshold adjustment coefficient, which can be flexibly set according to actual needs to control the sensitivity of ectopic pulsation detection. The intelligent education management terminal compares each heartbeat interval in the successive heartbeat interval sequence with the dynamic threshold. If... or If the heartbeat interval is not found to be an ectopic beat, it is determined to be removed from the successive heartbeat interval sequence. Optionally, for missing values ​​after removing ectopic beats, the intelligent education management terminal uses cubic spline interpolation to fill in the missing values. The interpolation formula is as follows: ,in, For cubic spline basis functions, These are the coefficients of the basis functions. The number of basis functions. The time points corresponding to the missing values ​​are filled in to obtain a complete successive heartbeat interval sequence. Then, based on this sequence, the time domain and frequency domain indices of heart rate variability are calculated to obtain the artifact-suppressed heart rate variability.

[0055] S24. Align artifact-suppressed EEG signals, artifact-suppressed EEG signals, and artifact-suppressed heart rate variability according to a unified timestamp to obtain a multimodal physiological signal time series. Extract multidimensional physiological features frame by frame from the multimodal physiological signal time series to obtain a multi-channel physiological feature sequence.

[0056] The intelligent education management terminal extracts timestamp information for artifact-suppressed EEG signals, artifact-suppressed TE signals, and artifact-suppressed heart rate variability, respectively. These timestamps are recorded synchronously by each sensor during signal acquisition to ensure consistency of the time reference. The intelligent education management terminal uses a linear interpolation method to align the three signals in time; the alignment formula is as follows: ,in, To align at a certain moment The signal value, and To align two adjacent sampling times before, and satisfy the following conditions: , and Given the signal values ​​at corresponding times, this formula unifies the three signals onto the same time axis, yielding a multimodal physiological signal time series. The expression for the multimodal physiological signal time series is: ,in, for Time series vector of multimodal physiological signals at any given time. for Artifact suppression of EEG signal values ​​at different times for Artifact suppression of skin conductance signal values ​​at specific times. for The artifact-suppressed heart rate variability value at each moment. The intelligent education management terminal uses a fixed-length frame window to divide the multimodal physiological signal time series frame by frame. The frame shift can be set to a time unit slightly longer than the sensor sampling period to ensure sufficient information density without causing redundancy. Multidimensional physiological features are extracted from the multimodal physiological signal time series within each frame window. For artifact-suppressed EEG signals, the relative power of five frequency bands (delta, theta, alpha, beta, and low gamma), frontal alpha asymmetry, frontal theta / beta ratio, sample entropy, and Lempel-Ziv complexity can be extracted. For artifact-suppressed EEG signals, the current value of the tension component, the cumulative amplitude of the phase component exceeding the threshold peak in the past few seconds, and the peak density can be extracted. For artifact-suppressed heart rate variability, the root mean square of the continuous difference and the power ratio of low frequency to high frequency can be extracted. All extracted multidimensional physiological features are aligned temporally and then concatenated to obtain a multi-channel physiological feature sequence. The expression for the multi-channel physiological feature sequence is as follows: ,in, It is a multi-channel physiological characteristic sequence. For the first Multidimensional physiological feature vector of a frame The total number of frames.

[0057] In one embodiment, a multi-channel physiological feature sequence is input into a hierarchical probabilistic state-space model to obtain short-term cognitive states and posterior probabilities of short-term cognitive states, including:

[0058] S31. Calculate the difference between the feature vector of the current frame and the feature vector of the previous frame in the multi-channel physiological feature sequence, and maintain the moving average of the difference over multiple consecutive frames.

[0059] Indicative, for the current processing frame ( The intelligent education management terminal calculates the difference between the feature vector of the current frame and the feature vector of the previous frame. The formula for calculating the difference is as follows: ,in, The difference between the feature vectors of the current frame and the previous frame. This is the multidimensional physiological feature vector of the current frame. This is the multidimensional physiological feature vector from the previous frame. The L2 norm is used to quantify the Euclidean distance between feature vectors of two frames, intuitively reflecting the magnitude of changes in physiological characteristics. To avoid the instantaneous fluctuations in feature differences within a single frame affecting subsequent adaptive noise scaling adjustments, the intelligent education management terminal maintains a moving average of the differences over multiple consecutive frames. The formula for calculating the moving average is as follows: ,in, This is the moving average of the feature differences corresponding to the current frame. The moving average window length can be flexibly set according to the frame shift and sampling rate of the multi-channel physiological feature sequence. The first frame before the current frame The difference in feature vectors between a frame and the corresponding previous frame. The intelligent education management terminal updates the moving average of feature differences frame by frame using this formula to ensure that it can stably reflect the continuous changing trend of physiological features in the multi-channel physiological feature sequence.

[0060] S32. Compare the moving average and the baseline fluctuation threshold in the individual physiological baseline to obtain the deviation difference, and calculate the adaptive noise scale of the current frame by integral update based on the deviation difference.

[0061] Furthermore, the intelligent education management terminal calls upon pre-stored individual physiological baselines and extracts the baseline fluctuation threshold from these baselines. This baseline fluctuation threshold is calculated based on multi-channel physiological characteristic data of the student at rest and is used to characterize the normal fluctuation range of the student's physiological characteristics. Its calculation formula is as follows: ,in, The baseline fluctuation threshold, This is a set of inter-frame differences in multi-channel physiological feature sequences under resting conditions. This represents the mean calculation function. This represents the function for calculating standard deviation. This is the threshold adjustment coefficient, which can be set according to actual needs to control the coverage of the normal fluctuation range. The intelligent education management terminal compares the moving average of the feature differences of the current frame with the baseline fluctuation threshold to obtain the deviation difference. The formula for calculating the deviation difference is: ,in, The deviation difference, This is the moving average of the feature differences in the current frame. The baseline fluctuation threshold is defined as follows: a positive deviation indicates that the current feature change exceeds the normal fluctuation range, while a negative deviation indicates that the current feature change is within the normal fluctuation range. Based on the deviation value, the intelligent education management terminal calculates the adaptive noise scale of the current frame through integral updates. The update formula is as follows: ,in, The adaptive noise scale for the current frame. The adaptive noise scale is the same as the previous frame. The integral learning rate parameter controls the update speed of the adaptive noise scale. This represents the deviation difference for the current frame. and These are the minimum and maximum thresholds for adaptive noise scaling, respectively. This function constrains the adaptive noise scale of the current frame within a preset range, preventing it from being too large or too small and affecting the accuracy of the observation likelihood. When the deviation is positive, the adaptive noise scale increases with the integral update, enhancing the response of the observation likelihood to changes in current features; when the deviation is negative, the adaptive noise scale decreases with the integral update, weakening the sensitivity of the observation likelihood to instantaneous noise.

[0062] S33. Using the state transition probability matrix of the hierarchical probabilistic state space model, the posterior probability of the short-term cognitive state of the previous frame, the feature vector of the current frame, and the adaptive noise scale of the current frame, the posterior probability distribution of each short-term cognitive state in the current frame is calculated by the Bayesian forward recursive algorithm, thus obtaining the short-term cognitive state and the posterior probability of the short-term cognitive state.

[0063] Optionally, the intelligent education management terminal pre-constructs the state transition probability matrix of a hierarchical probabilistic state space model, and the expression of the state transition probability matrix is ​​as follows: ,in, Here is the state transition probability matrix. The number of short-term cognitive states. Let be the state transition probability, representing the probability from the previous frame to the th . The short-term cognitive state transitions to the current frame. The probability of a short-term cognitive state is represented by the diagonal elements of the matrix, which are dominant, reflecting the inherent inertia of short-term cognitive states. This also allows for legitimate rapid state transitions and satisfies... That is, the sum of the probabilities of each row vector is 1. The intelligent education management terminal calls the short-term cognitive state posterior probability of the previous frame, and the expression for the short-term cognitive state posterior probability of the previous frame is: ,in, This is the posterior probability vector of the short-term cognitive state of the previous frame. For the previous frame The posterior probability of a short-term cognitive state, and satisfying The intelligent education management terminal calculates the observation likelihood based on the feature vector of the current frame and the adaptive noise scale of the current frame. The expression for the observation likelihood is: ,in, Let be the observation likelihood of the current frame feature vector in the j-th short-term cognitive state. Represents a multidimensional Gaussian distribution. The feature vector of the current frame. For the first The mean vector of physiological features corresponding to each short-term cognitive state The adaptive noise scale for the current frame. For the first The basis of the covariance matrix corresponding to each short-term cognitive state. The intelligent education management terminal calculates the posterior probability distribution of each short-term cognitive state in the current frame using a Bayesian forward recursive algorithm, with the recursive formula as follows: ,in, For the current frame number The posterior probability of a short-term cognitive state. The observation likelihood for the current frame. Let be the state transition probability. For the previous frame The posterior probability of a short-term cognitive state, denoted by . This indicates a proportional relationship. The intelligent education management terminal normalizes the calculated posterior probabilities of each short-term cognitive state using the following normalization formula: To ensure that the sum of the normalized posterior probabilities is 1, the posterior probability distribution of the short-term cognitive state in the current frame is obtained. The intelligent education management terminal determines the short-term cognitive state with the highest posterior probability as the short-term cognitive state of the current frame, ultimately obtaining the short-term cognitive state and its posterior probability.

[0064] In one embodiment, a stability statistical test based on a continuous time window is performed on the posterior probability of the short-term cognitive state. When the posterior probability of the short-term cognitive state satisfies the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a staged established state, including:

[0065] S41. For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state corresponding to the short-term cognitive state is greater than the preset upper threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the upper threshold of the posterior probability is added to the cumulative sum statistic corresponding to the short-term cognitive state.

[0066] Schematic illustration: The intelligent education management terminal maintains an independent cumulative sum statistic for each predefined short-term cognitive state label. The cumulative sum statistic quantifies the stability of the posterior probability of the short-term cognitive state within a continuous time window, and its initial value is set to 0. For each frame of the multi-channel physiological feature sequence, the intelligent education management terminal iterates through all short-term cognitive states, obtaining the posterior probability of each short-term cognitive state. The expression for the posterior probability of the short-term cognitive state is: ,in Let be the posterior probability of the k-th short-term cognitive state at time t. For the time step index of the current frame, This serves as an index for short-term cognitive states. The intelligent education management terminal pre-sets a threshold for the posterior probability, the expression for which is: , The high-confidence range used to define the posterior probability of short-term cognitive states can be set based on individual physiological baselines and empirical data from educational psychology, and can be dynamically adjusted according to the actual situation during the learning process. The intelligent education management terminal compares the posterior probability of each short-term cognitive state with a preset posterior probability threshold. When the threshold is met... When this occurs, it indicates that the short-term cognitive state has a high confidence level in the current frame. The intelligent education management terminal incrementally updates the cumulative sum and statistics corresponding to this short-term cognitive state using the following formula: ,in, for Time of the first The cumulative sum and statistics of short-term cognitive states for Time of the first The cumulative sum and statistics of short-term cognitive states The increment representing the portion of the short-term cognitive state posterior probability exceeding a preset upper threshold is used to accumulate stable evidence for the short-term cognitive state. A larger cumulative sum and statistic indicate stronger stability of the short-term cognitive state. The intelligent education management terminal continuously accumulates stable evidence for each short-term cognitive state by updating the cumulative sum and statistic frame by frame, providing data support for subsequent stability assessments.

[0067] S42. For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state corresponding to the short-term cognitive state is less than the preset threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the threshold of the posterior probability is subtracted from the cumulative sum statistic corresponding to the short-term cognitive state.

[0068] The intelligent education management terminal pre-sets a threshold value under a preset posterior probability. The expression for the threshold value under the preset posterior probability is: , Used to define the low-confidence range of the posterior probability of short-term cognitive states, and satisfying the following conditions: , It can be set in conjunction with a preset posterior probability threshold, and can also be adjusted based on individual physiological baselines and empirical data from educational psychology to ensure effective differentiation of confidence levels for short-term cognitive states. For each short-term cognitive state in each frame, the intelligent education management terminal sets its short-term cognitive state posterior probability... With a preset posterior probability threshold Compare, when satisfied When this occurs, it indicates that the confidence level of the short-term cognitive state in the current frame is low and there is a lack of evidence of its stable existence. The intelligent education management terminal then updates the cumulative sum and statistics corresponding to this short-term cognitive state by reducing the amount of data. The update formula is as follows: ,in, for Time of the first The cumulative sum and statistics of short-term cognitive states for Time of the first The cumulative sum and statistics of short-term cognitive states The portion of the short-term cognitive state posterior probability that is lower than a preset lower threshold is reduced to weaken the stable evidence of the short-term cognitive state, thus avoiding misjudgments caused by transiently high confidence. Simultaneously, the intelligent education management terminal sets the lower limit of the cumulative sum statistic to 0; that is, when the cumulative sum statistic is less than 0 after the reduction update, it is automatically reset to 0. The formula is as follows: This ensures that accumulation and statistics can only accumulate positive, stable evidence, and that there is no judgment bias caused by negative accumulation.

[0069] S43. When the cumulative sum of any short-term cognitive state is greater than or equal to the decision boundary for the first time, the short-term cognitive state is determined to be a stage-established state, and the cumulative sum of the short-term cognitive state is reset to zero.

[0070] The intelligent education management terminal pre-sets decision boundaries, and the expression for the decision boundaries is: , It is used to quantify the minimum amount of accumulated evidence required for short-term cognitive states to stabilize, and can be flexibly set according to the preset number of frames in a continuous time window, the type of short-term cognitive state, and individual physiological characteristics. A larger value indicates that more continuous and stable evidence is needed to determine a stage of established cognitive state, effectively reducing the probability of misjudgment caused by transient noise. The intelligent education management terminal monitors the cumulative sum and statistics of all short-term cognitive states frame by frame. Each cumulative sum statistic is compared with the decision boundary. Comparison, when any short-term cognitive state satisfies Furthermore, when this short-term cognitive state first appears within the current continuous time window, the intelligent education management terminal determines that the short-term cognitive state is a phased establishment state. This indicates that the short-term cognitive state has maintained a high confidence level over multiple consecutive frames, possesses sufficient stability, and can serve as a reliable basis for adjusting the learning path. After determining it to be a phased establishment state, the intelligent education management terminal immediately resets the cumulative sum statistics corresponding to the short-term cognitive state to zero, using the following reset formula: This approach avoids the accumulation of statistics and subsequent repeated judgments, while also accumulating evidence for the next stage of stability testing, ensuring that the judgment of each stage of established state is based on independent continuous time window evidence, thus improving the reliability of the judgment results.

[0071] S44. During the lock-in period of the dynamic refractory period constraint activated after the previous major learning action, if there is a conflict between the cognitive load level of the staged established state and the currently executed learning action sequence, then the staged established state is suppressed.

[0072] After executing each significant learning action, the intelligent education management terminal immediately activates the dynamic refractory period constraint and starts the lock-in timer. Significant learning actions refer to actions that have a substantial impact on the learning path, including adjustments to the difficulty level of learning content exceeding a preset range, fundamental changes in the type of learning activity, etc. The lock-in time expression for the dynamic refractory period constraint is as follows: , It can adaptively adjust based on the learner's individual response speed, the nature of the current learning task, and the oscillation tendency detected by the system. The intelligent education management terminal monitors in real time whether the locking time of the dynamic refractory period constraint has ended. Before the learning process concludes, if a new short-term cognitive state is identified as a stage of establishment, the intelligent education management terminal will make a conflict judgment between this stage of establishment and the cognitive load level of the currently executed learning action sequence. The quantitative formula for cognitive load level is as follows: ,in, For cognitive load level, The length of the currently executed learning action sequence. For the first Cognitive load weights for each learned action already performed. For the first The cognitive load quantification value of each performed learning action. The formula for conflict judgment is: ,in, To establish the cognitive load requirement value corresponding to each stage, The cognitive load conflict threshold is defined by this formula. When this threshold is met, it indicates a significant conflict between the established stage state and the cognitive load level of the currently executed learning action sequence. For example, the cognitive load level corresponding to the currently executed learning action sequence may be high, while the newly determined established stage state is "too low cognitive load," creating a clear contradiction. In this case, the intelligent education management terminal will suppress this established stage state, meaning it will not use it as a basis for adjusting the learning path. This avoids frequent switching of learning actions due to cognitive load conflict during the dynamic refractory period, further enhancing the stability of system decisions and preventing oscillations.

[0073] In one embodiment, a cost function is constructed based on the phased established state and probabilistic evolution trajectory, including the cost of state deviation and the penalty for the magnitude of action change. Then, rolling time-domain optimization is performed based on minimizing the cost function and dynamic refractory period constraints to obtain the optimal learning action for the current decision cycle, including:

[0074] The cost function is obtained using the following formula:

[0075]

[0076] in, Pre-set time periods for the future, For the probabilistic evolution trajectory, the first The short-term cognitive state probability vector of a frame. For the probability distribution of the ideal cognitive state, Let KL divergence be the KL divergence. The sequence of learning actions to be optimized is... For the previously executed sequence of learned actions, The penalty coefficient for the range of motion change. The cost coefficient for state deviation. For learner stable trait vectors, This is a function of the degree of matching between the learned action sequence and the stable trait vector to be optimized; The L2 norm of the change in the learned action sequence is used to activate the refractory period timer when the L2 norm of the change in the learned action sequence exceeds a preset threshold for significant actions. The base duration of the refractory period timer is determined by the product of the magnitude of the action change and the learner's individual information processing speed coefficient, and is superimposed with an extension that is positively correlated with the current decision oscillation index. During the lockout time period, the refractory period timer is used to mark the learned actions whose L2 norm of the change in the learned action sequence exceeds the maximum allowable change threshold as infeasible.

[0077] Indicative, The total decision cost is used to comprehensively measure the rationality of the decision corresponding to the learning action sequence. The smaller the total decision cost, the more the learning action sequence fits the student's current cognitive state and stable learning characteristics, and the more it can avoid decision oscillation. To predict the time domain length, i.e. the number of frames contained in the future preset time period corresponding to the probability evolution trajectory, it can be flexibly set according to the frame shift of the multi-channel physiological feature sequence and the complexity of the learning task, to ensure that it can cover enough future cognitive state change trends. KL divergence, also known as relative entropy, is used to calculate the k-th ... The degree of deviation between the short-term cognitive state probability distribution of a frame and the ideal cognitive state probability distribution constitutes the state deviation cost. The larger the KL divergence value, the more serious the deviation between the predicted short-term cognitive state and the ideal state, and the higher the corresponding state deviation cost. For the probabilistic evolution trajectory, the first The short-term cognitive state probability vector of the frame, whose elements are the first... The posterior probabilities corresponding to each short-term cognitive state in a frame, with the sum of all elements equal to 1, are obtained from the output of an individualized recurrent neural network prediction model. The ideal cognitive state probability distribution can be based on educational goals, students' stable learning trait vectors, and phased establishment of states. The ideal cognitive state probability distribution can be a unimodal distribution, that is, the probability of a certain optimal short-term cognitive state is 1 and the probability of the other states is 0, or it can be a multimodal distribution, that is, multiple suitable short-term cognitive states are assigned a certain probability to adapt to the needs of different learning scenarios. This is the penalty coefficient for the magnitude of action change, used to adjust the weight of the penalty term for the magnitude of action change. It can be dynamically adjusted based on the decision oscillation trend detected by the system; the more pronounced the decision oscillation, the more significant the impact. The larger the value can be, the stronger the suppression effect on frequent action switching. The learning action sequence to be optimized is defined by the learning action vectors corresponding to each time step in the future prediction time domain. The learning action vectors may include multi-dimensional decision parameters such as the difficulty of the learning content, the type of learning activity, and the learning pace. In the sequence of learning actions to be optimized, and The vector of the preceding adjacent learned action is used to calculate the magnitude of change in the adjacent learned action. The square of the L2 norm of the change in the learned action sequence is used to quantify the degree of difference between two adjacent learned actions, constituting the penalty for the magnitude of action change. The larger the value, the more significant the difference between adjacent learned actions, and the higher the penalty for the magnitude of action change. This is the state deviation cost coefficient, used to adjust the weight of the state deviation cost term. It can be flexibly set based on the importance of the learning task and the student's learning sensitivity. For core learning tasks, the weight can be increased. Prioritize ensuring that the cognitive state matches the ideal state. The learner stable trait vector contains long-term stable individual trait parameters such as learner's cognitive ability, learning preferences, and information processing speed. It is obtained from the top stable learning trait layer of the hierarchical probabilistic state-space model and can be continuously calibrated through long-term learning data. This is a function that measures the degree of matching between the sequence of learning actions to be optimized and the stable trait vector. It quantifies the fit between the learning actions and the learner's stable traits. The smaller the function value, the higher the degree of matching between the learning actions and the stable traits, and the smaller the corresponding mismatch cost. It can be constructed using various forms such as the reciprocal of cosine similarity and Euclidean distance. The intelligent education management terminal can choose the appropriate function form according to actual adaptation needs.

[0078] Furthermore, the intelligent education management terminal pre-sets a threshold for critical actions, the expression for which is: , This method is used to determine whether the variation in a learning sequence of actions meets the criteria for a significant movement. It can be flexibly set based on the learner's individual information processing speed and the continuity requirements of the learning task. Different learners and different learning tasks can be assigned different criteria. The intelligent education management terminal calculates in real time the L2 norm of the change in adjacent learning actions in the sequence of learning actions to be optimized, i.e. and compare it with a preset threshold for major actions. Compare, when satisfied When the change in the current adjacent learning action reaches the threshold for a significant action, the intelligent education management terminal immediately activates the refractory period timer to prevent another large-scale switching of learning actions within a short period and to prevent decision oscillation. The base duration of the refractory period timer is determined by the product of the magnitude of the action change and the learner's individual information processing speed coefficient, plus an extension positively correlated with the current decision oscillation index. The calculation formula is as follows: , ,in, This is the base duration of the refractory period timer. This is a duration adjustment factor used to adjust the overall magnitude of the base duration. To trigger the learning action change amplitude of the refractory period timer, The learner's individual information processing speed coefficient is derived from the learner's stable trait vector. The faster the information is extracted and processed, the better. The larger the value, the shorter the base duration. The learner's stable trait vector is the reciprocal of their individual information processing speed coefficient. For example, the learner's stable trait vector includes multi-dimensional quantitative components such as working memory capacity, information processing speed, knowledge acceptance sensitivity, learning preferences, and fatigue susceptibility. The learner's stable trait vector can be represented as... In the formula, For learner stable trait vectors, To stabilize the overall dimension of the trait vector, This refers to the original quantized component in the stable trait vector that specifically corresponds to the information processing speed dimension. The intelligent education management terminal works by jointly solving the learner's individual information processing speed coefficient through an initial assignment phase and an online iterative correction phase. The initial assignment phase relies on the collection of standardized cognitive assessment tasks and learning response behavior data from students. The collected data includes the time delay for understanding and responding to standardized knowledge points, the time span for consecutive answers to similar question types, and the reaction time for cognitive acceptance of unfamiliar learning materials. The intelligent education management terminal performs mean statistics and dimensionless normalization on multiple sets of time-series response data to obtain the raw quantitative component of information processing speed. Based on this, the coefficient conversion is completed using a monotonically increasing mapping function. The conversion formula is as follows: In the formula, The learner's individual information processing speed coefficient. This is a pre-defined monotonic linear mapping function that maintains the characteristic that the input and output change monotonically in the same direction. The faster the information processing speed, the more likely it is to correspond to this mapping function. The larger the value, the faster the learner's individual information processing speed coefficient is after mapping. The larger the value, the better. During the online iterative correction phase, the intelligent education management terminal continuously accumulates behavioral and physiological correlation data such as the rate of change of multi-channel physiological feature sequences, the response latency of short-term cognitive state switching, and the adaptation and acceptance time of learning content difficulty during long-term learning. Relying on the top-level stable trait layer of the hierarchical probabilistic state space model, it continuously fine-tunes the internal values ​​of the stable trait vector using a Bayesian posterior update mechanism. The component, synchronously updated with the learner's individual information processing speed coefficient ,make It can adapt to the slow drift and changes in the learner's physiological state and cognitive ability over a long period of time. This is the final locking time of the refractory period timer. This is the oscillation extension coefficient, used to adjust the degree of influence of the decision oscillation index on the lock-in time. The current decision oscillation index is defined as follows: the larger the decision oscillation index, the greater the superimposed extension and the longer the lock-in time, thus enhancing the suppression effect on oscillations. After the refractory period timer starts and enters the lock-in time counting period, the intelligent education management terminal pre-sets a maximum allowable change threshold. The expression for the maximum allowable change threshold is: , Less than the preset threshold for major actions This is used to define the maximum permissible range of changes in learning actions within the locked time period. During the locked time, the intelligent education management terminal filters all learning actions to be optimized, marking learning actions whose L2 norm of the change in the learning action sequence exceeds the maximum permissible change threshold as infeasible. If the learning action is deemed infeasible, it will not be included in the optimization solution scope to ensure the stability of the learning action within the locked time.

[0079] Optionally, the intelligent education management terminal incorporates dynamic refractory period constraints into the rolling time-domain optimization process. These constraints include locking time constraints for the refractory period timer and filtering constraints for infeasible learning actions, ensuring the stability of learning actions during optimization and avoiding decision oscillations. The intelligent education management terminal employs a rolling time-domain optimization strategy, treating learning path adjustment as a finite-time-domain optimal control problem with stability constraints. The optimization time domain is a preset future time period. This means maintaining consistency with the predicted time domain length of the probabilistic evolution trajectory. During the optimization process, the intelligent education management terminal filters out all feasible learning action sequences that satisfy the dynamic refractory period constraint, eliminating those marked as infeasible, ensuring that all candidate learning action sequences meet stability requirements. Furthermore, the intelligent education management terminal calculates the total decision cost corresponding to each feasible learning action sequence. The optimal learning action sequence is solved by minimizing the total decision cost, and the objective function is: ,in, The learning action sequence to be optimized is shown. The intelligent education management terminal can use various optimization algorithms, such as direct target method, Monte Carlo tree search method, or gradient descent method, to solve this constrained minimization problem. The appropriate optimization algorithm can be flexibly selected based on the discreteness and dimensionality of the learning action space. For discrete action spaces, Monte Carlo tree search method is preferred; for continuous action spaces, direct target method is preferred. After obtaining the optimal learning action sequence, the intelligent education management terminal extracts the first learning action of the sequence, i.e., the learning action corresponding to the current decision cycle, as the optimal learning action for the current decision cycle. Subsequent decision cycles will re-execute the optimization process based on new multi-channel physiological feature sequences, short-term cognitive state posterior probabilities, and probability evolution trajectories, achieving real-time adaptive adjustment of the learning actions while ensuring decision stability.

[0080] In one embodiment, when the decision oscillation index exceeds a preset first oscillation threshold, the method further includes: increasing the weight of the penalty for the change in action by setting a step size and extending the lock-in time of the dynamic refractory period constraint;

[0081] S51. When the decision oscillation index exceeds the preset second threshold, a protective degradation mechanism is triggered until the decision oscillation index falls back to the safe range and remains within the safe range for the target duration. The protective degradation mechanism is used to switch from adaptive learning path adjustment to a predefined fixed lesson plan mode and send a prompt message to the teacher.

[0082] As an illustration, the intelligent education management terminal is pre-set with a second threshold, which is a critical value used to determine whether severe oscillations have occurred. Its expression is: , The value is greater than the preset first oscillation threshold. The decision oscillation index can be flexibly set based on the system runaway risk threshold and the continuity requirements of the learning task, and can be calibrated in conjunction with a preset first oscillation threshold to ensure the accuracy of severe oscillation judgment. The intelligent education management terminal displays the calculated decision oscillation index in real time. With the preset second threshold Compare, when satisfied When this occurs, it indicates that the current decision-making oscillation has reached a severe level. If the adaptive learning path adjustment continues, it may lead to frequent switching of learning content, confusion in students' cognition, or even loss of control of system decision-making. At this time, the intelligent education management terminal immediately triggers the protective degradation mechanism.

[0083] Specifically, the protective degradation mechanism involves switching from an adaptive learning path to a predefined fixed lesson plan pattern. This predefined pattern is designed by the teacher in advance based on students' learning objectives, prior knowledge, and the curriculum syllabus. It may include a fixed sequence of learning content, difficulty level, learning pace, and practice arrangements. The fixed lesson plan pattern can be expressed as follows: ,in, For a predefined fixed lesson plan pattern, For the first lesson in the fixed lesson plan model Each learning segment To ensure the total number of learning steps in the fixed lesson plan mode, after the intelligent education management terminal switches to this mode, it will stop adaptive decision-making based on multi-channel physiological feature sequences, short-term cognitive state posterior probability, and probability evolution trajectory, and strictly execute learning actions according to the predefined fixed lesson plan mode to avoid further exacerbation of decision oscillation.

[0084] Furthermore, the intelligent education management terminal sends prompts to teachers, which may include the specific value of the decision oscillation index, the severity of the oscillation, the trigger time, and the currently executed fixed lesson plan mode. This allows teachers to monitor the system status and students' learning progress in real time. Teachers can intervene based on the prompts to adjust the fixed lesson plan mode or investigate the cause of the oscillation.

[0085] After triggering the protective degradation mechanism, the intelligent education management terminal continuously monitors changes in the decision oscillation index, while simultaneously setting a safe range and a target duration. The expression for the safe range is: ,in, This is the lower limit of the safe range. It is the upper limit of the safe interval, and satisfies... This ensures that the decision oscillation index within the safe interval remains within a safe range of no oscillation or slight oscillation. The expression for the target duration is: , The minimum time that the decision oscillation index needs to remain within a safe range is used to define the minimum time it can be flexibly set according to the duration of the learning task and the student's cognitive recovery speed. Its calculation formula is as follows: ,in, The target duration is the baseline value. This is a duration adjustment factor used to adjust the target duration according to the severity of oscillations. The decision oscillation index when the downgrade mechanism is triggered. The more severe the oscillation, the more likely it is to occur. The larger the value, the longer the target duration, ensuring that decision oscillations are completely quelled.

[0086] The intelligent education management terminal determines in real time whether the decision oscillation index has fallen back to a safe range, i.e., meets the requirements. And record the timestamp when the decision oscillation index enters the safe zone. When the decision oscillation index remains within the safe range for a period of time that meets the target duration, it satisfies the condition. ,in, At the current timestamp, the intelligent education management terminal deactivates the protective degradation mechanism, restores the adaptive learning path adjustment mode, and re-determines the optimal learning action based on the multi-channel physiological feature sequence, short-term cognitive state posterior probability, and probability evolution trajectory. For example, the safety interval is set to a decision oscillation index within the range [0, 0.20], the target duration baseline is set to 120 seconds, and the duration adjustment coefficient is set to an oscillation index of 80 seconds per unit.

[0087] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0088] Based on the same inventive concept, this application also provides an IoT-based intelligent education management system for implementing the IoT-based intelligent education management method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more IoT-based intelligent education management system embodiments provided below can be found in the limitations of the IoT-based intelligent education management method described above, and will not be repeated here.

[0089] In one exemplary embodiment, such as Figure 2 As shown, an IoT-based intelligent education management system 200 is provided, comprising:

[0090] The physiological data module 201 is used to acquire multimodal physiological signals synchronously collected by the sensors worn by students, and to perform adaptive artifact suppression on the multimodal physiological signals based on individual physiological baselines to obtain multi-channel physiological feature sequences.

[0091] The short-term cognitive state module 202 is used to input multi-channel physiological feature sequences into a hierarchical probabilistic state space model to obtain short-term cognitive states and short-term cognitive state posterior probabilities. During the inference of the short-term cognitive state posterior probability, the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequences.

[0092] The stage cognitive state module 203 is used to perform a stability statistical test on the posterior probability of the short-term cognitive state based on a continuous time window. When the posterior probability of the short-term cognitive state meets the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a stage-established state.

[0093] The evolution prediction module 204 is used to input the historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a preset time period in the future.

[0094] The learning action solution module 205 is used to construct a cost function containing state deviation costs and action change magnitude penalties based on the phased established state and probability evolution trajectory, and to perform rolling time-domain optimization based on minimizing the cost function and dynamic refractory period constraints to solve for the optimal learning action in the current decision cycle.

[0095] The oscillation suppression module 206 is used to update the currently executed learning action sequence by executing the optimal learning action, and calculate the decision oscillation index based on the updated currently executed learning action sequence. When the decision oscillation index exceeds the preset first oscillation threshold, the weight of the penalty for the change amplitude of the action is increased by setting the step size and the locking time of the dynamic refractory period constraint is extended.

[0096] In one embodiment, the physiological data module 201 is further configured to:

[0097] Based on a sliding window, the temporal amplitude kurtosis and specific frequency band power spectral density of EEG signals from multimodal physiological signals are calculated, and the comprehensive deviation of the temporal amplitude kurtosis and specific frequency band power spectral density relative to the individual physiological baseline is calculated. When the comprehensive deviation exceeds a set deviation threshold, instantaneous artifact subtraction is performed on the EEG signal based on a time-varying Kalman smoother within the sliding window to obtain an artifact-suppressed EEG signal; the process noise covariance of the time-varying Kalman smoother is adaptively adjusted with the comprehensive deviation.

[0098] Based on the acceleration signal of sudden changes in body movement, the amplitude of the skin conductance phase peak caused by sudden changes in body movement in the skin conductance signal of multimodal physiological signals is identified and compressed to obtain the artifact-suppressed skin conductance signal.

[0099] The successive heartbeat interval sequence was extracted from the photoplethysmography pulse wave signal of multimodal physiological signals, and ectopic beats in the successive heartbeat interval sequence were detected and removed based on the dynamic threshold of median absolute deviation, thus obtaining artifact-suppressed heart rate variability.

[0100] By aligning artifact-suppressed EEG signals, artifact-suppressed EEG signals, and artifact-suppressed heart rate variability according to a unified timestamp, a multimodal physiological signal time series was obtained. Multidimensional physiological features were then extracted frame by frame from the multimodal physiological signal time series to obtain a multi-channel physiological feature sequence.

[0101] In one embodiment, the short-term cognitive state module 202 is further configured to:

[0102] Calculate the difference between the feature vector of the current frame and the feature vector of the previous frame in the multi-channel physiological feature sequence, and maintain the moving average of the difference over multiple consecutive frames;

[0103] By comparing the moving average with the baseline fluctuation threshold in the individual physiological baseline, the deviation difference is obtained, and based on the deviation difference, the adaptive noise scale of the current frame is calculated by integral updating.

[0104] By utilizing the state transition probability matrix of the hierarchical probabilistic state space model, the posterior probability of the short-term cognitive state of the previous frame, the feature vector of the current frame, and the adaptive noise scale of the current frame, the posterior probability distribution of each short-term cognitive state in the current frame is calculated using the Bayesian forward recursive algorithm, thus obtaining the short-term cognitive state and its posterior probability.

[0105] In one embodiment, the stage-based cognitive state module 203 is further configured to:

[0106] For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state is greater than the preset upper threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the upper threshold of the posterior probability is added to the cumulative sum statistic corresponding to the short-term cognitive state.

[0107] For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state is less than the preset threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the threshold of the posterior probability is subtracted from the cumulative sum statistic corresponding to the short-term cognitive state.

[0108] When the cumulative sum of any short-term cognitive state is greater than or equal to the decision boundary for the first time, the short-term cognitive state is determined to be a stage-established state, and the cumulative sum of the short-term cognitive state is reset to zero.

[0109] If the established phase state conflicts with the cognitive load level of the currently executed learning sequence during the lock-in period of the dynamic refractory period constraint that is activated after the previous major learning action, the established phase state is suppressed.

[0110] In one embodiment, the IoT-based intelligent education management system 200 is further used for:

[0111] When the decision oscillation index exceeds the preset second threshold, a protective degradation mechanism is triggered until the decision oscillation index falls back to the safe range and remains within the safe range for the target duration. The protective degradation mechanism is used to switch from adaptive learning path adjustment to a predefined fixed lesson plan mode and send a prompt message to the teacher.

[0112] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0113] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0114] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0115] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A smart education management method based on the Internet of Things, characterized in that, The method includes: The system acquires multimodal physiological signals synchronously collected by sensors worn by students, and performs adaptive artifact suppression on the multimodal physiological signals based on individual physiological baselines to obtain multichannel physiological feature sequences. The multi-channel physiological feature sequence is input into a hierarchical probabilistic state space model to obtain short-term cognitive state and short-term cognitive state posterior probability; the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequence during the inference of short-term cognitive state posterior probability. The stability statistical test based on continuous time window is performed on the posterior probability of the short-term cognitive state. When the posterior probability of the short-term cognitive state meets the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a staged establishment state. The historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence are input into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a future preset time period. Based on the established phased state and the probabilistic evolution trajectory, a cost function is constructed that includes the cost of state deviation and the penalty for the magnitude of action change. Then, rolling time-domain optimization is performed based on minimizing the cost function and the dynamic refractory period constraint to obtain the optimal learning action for the current decision cycle. The currently executed learning action sequence is updated by executing the optimal learning action, and the decision oscillation index is calculated based on the updated currently executed learning action sequence. When the decision oscillation index exceeds a preset first oscillation threshold, the weight of the penalty for the change amplitude of the action is increased by setting a step size and the locking time of the dynamic refractory period constraint is extended.

2. The method according to claim 1, characterized in that, The adaptive artifact suppression of the multimodal physiological signals based on individual physiological baselines yields a multi-channel physiological feature sequence, including: Based on a sliding window, the temporal amplitude kurtosis and specific frequency band power spectral density of the EEG signals of the multimodal physiological signals are calculated, and the comprehensive deviation of the temporal amplitude kurtosis and the specific frequency band power spectral density relative to the individual physiological baseline is calculated. When the comprehensive deviation exceeds a set deviation threshold, instantaneous artifact subtraction is performed on the EEG signals based on a time-varying Kalman smoother within the sliding window to obtain artifact-suppressed EEG signals; the process noise covariance of the time-varying Kalman smoother is adaptively adjusted according to the comprehensive deviation. Based on the acceleration signal of sudden changes in body movement, the amplitude of the skin conductance phase peak caused by sudden changes in body movement in the skin conductance signal of the multimodal physiological signal is identified and compressed to obtain the artifact-suppressed skin conductance signal. The successive heartbeat interval sequence is extracted from the photoplethysmography pulse wave signal of the multimodal physiological signal, and ectopic beats in the successive heartbeat interval sequence are detected and removed based on the dynamic threshold of median absolute deviation, so as to obtain artifact-suppressed heart rate variability. The artifact-suppressed EEG signal, the artifact-suppressed EEG signal, and the artifact-suppressed heart rate variability are aligned with a unified timestamp to obtain a multimodal physiological signal time series. Multidimensional physiological features are then extracted frame by frame from the multimodal physiological signal time series to obtain the multichannel physiological feature sequence.

3. The method according to claim 2, characterized in that, The step of inputting the multi-channel physiological feature sequence into a hierarchical probabilistic state-space model to obtain short-term cognitive states and short-term cognitive state posterior probabilities includes: Calculate the difference between the feature vector of the current frame and the feature vector of the previous frame in the multi-channel physiological feature sequence, and maintain the moving average of the difference over multiple consecutive frames; The deviation difference is obtained by comparing the moving average value with the baseline fluctuation threshold in the individual physiological baseline, and the adaptive noise scale of the current frame is calculated by integral updating based on the deviation difference. Using the state transition probability matrix of the hierarchical probabilistic state space model, the posterior probability of the short-term cognitive state of the previous frame, the feature vector of the current frame, and the adaptive noise scale of the current frame, the posterior probability distribution of each of the short-term cognitive states in the current frame is calculated by the Bayesian forward recursive algorithm, thereby obtaining the short-term cognitive states and their posterior probabilities.

4. The method according to claim 1, characterized in that, The step of performing a stability statistical test on the posterior probability of the short-term cognitive state based on a continuous time window, and marking the short-term cognitive state as a staged established state when the posterior probability of the short-term cognitive state satisfies the stability condition within a continuous time window of a preset number of frames, includes: For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state corresponding to the short-term cognitive state is greater than a preset upper threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the upper threshold of the posterior probability is added to the cumulative sum statistic corresponding to the short-term cognitive state. For each short-term cognitive state in each frame, if the posterior probability of the short-term cognitive state corresponding to the short-term cognitive state is less than a preset lower threshold of the posterior probability, then the difference between the posterior probability of the short-term cognitive state and the lower threshold of the posterior probability is subtracted from the cumulative sum statistic corresponding to the short-term cognitive state. When the cumulative sum statistic of any short-term cognitive state is greater than or equal to the decision boundary for the first time, the short-term cognitive state is determined to be the stage-established state, and the cumulative sum statistic of the short-term cognitive state is reset to zero. If, during the locking time of the dynamic refractory period constraint activated after the previous major learning action, the phased establishment state conflicts with the cognitive load level of the currently executed learning action sequence, then the phased establishment state is suppressed.

5. The method according to claim 1, characterized in that, The process involves constructing a cost function based on the established phased state and the probabilistic evolution trajectory, which includes state deviation costs and penalties for action change magnitudes. Then, based on minimizing the cost function and the dynamic refractory period constraint, rolling time-domain optimization is performed to obtain the optimal learning action for the current decision cycle. This includes: The cost function is obtained using the following formula: in, Pre-set time periods for the future, For the probabilistic evolution trajectory, the first... The short-term cognitive state probability vector of a frame. For the probability distribution of the ideal cognitive state, Let KL divergence be the KL divergence. The sequence of learning actions to be optimized is... For the previously executed sequence of learned actions, The penalty coefficient for the range of motion change. The cost coefficient for state deviation. For learner stable trait vectors, This is a function of the degree of matching between the learning action sequence to be optimized and the stable trait vector; The L2 norm of the change in the learned action sequence is used to activate a refractory period timer when the L2 norm of the change in the learned action sequence exceeds a preset threshold for significant actions. The base duration of the refractory period timer is determined by the product of the magnitude of the action change and the learner's individual information processing speed coefficient, and is superimposed with an extension that is positively correlated with the current decision oscillation index. The refractory period timer is used to mark learning actions whose L2 norm of the change in the learned action sequence exceeds the maximum allowable change threshold as infeasible during the lockout time period.

6. The method according to claim 1, characterized in that, When the decision oscillation index exceeds a preset first oscillation threshold, the weight of the penalty for the change in the action is increased by setting a step size, and the lock-up time of the dynamic refractory period constraint is extended, further including: When the decision oscillation index exceeds a preset second threshold, a protective degradation mechanism is triggered until the decision oscillation index falls back to a safe range and remains within the safe range for the target duration. The protective degradation mechanism is used to switch from adaptive learning path adjustment to a predefined fixed lesson plan mode and send a prompt message to the teacher.

7. An intelligent education management system based on the Internet of Things, characterized in that, The system includes: The physiological data module is used to acquire multimodal physiological signals synchronously collected by sensors worn by students, and to perform adaptive artifact suppression on the multimodal physiological signals based on individual physiological baselines to obtain multichannel physiological feature sequences. The short-term cognitive state module is used to input the multi-channel physiological feature sequence into the hierarchical probabilistic state space model to obtain the short-term cognitive state and the short-term cognitive state posterior probability; the hierarchical probabilistic state space model dynamically adjusts the observation noise scale according to the change amplitude of physiological features in the multi-channel physiological feature sequence during the inference of the short-term cognitive state posterior probability. The stage cognitive state module is used to perform a stability statistical test on the posterior probability of the short-term cognitive state based on a continuous time window. When the posterior probability of the short-term cognitive state meets the stability condition within a continuous time window of a preset number of frames, the short-term cognitive state is marked as a stage-established state. The evolution prediction module is used to input the historical multi-channel physiological feature sequence, the short-term cognitive state posterior probability sequence corresponding to the historical multi-channel physiological feature sequence, and the currently executed learning action sequence into the individualized recurrent neural network prediction model to generate the probability evolution trajectory of each short-term cognitive state within a preset time period in the future. The learning action solution module is used to construct a cost function containing state deviation cost and action change magnitude penalty based on the staged established state and the probability evolution trajectory, and to perform rolling time-domain optimization based on minimizing the cost function and dynamic refractory period constraint to solve for the optimal learning action in the current decision cycle. The oscillation suppression module is used to update the currently executed learning action sequence by executing the optimal learning action, and calculate the decision oscillation index based on the updated currently executed learning action sequence. When the decision oscillation index exceeds a preset first oscillation threshold, the weight of the penalty for the change amplitude of the action is increased by setting a step size and the locking time of the dynamic refractory period constraint is extended.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images