Student learning ability dynamic evaluation method fused with artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a student learning ability dynamic evaluation method fused with artificial intelligence. According to the method, physiological signals, behavior postures and task interaction data are synchronously collected through a multi-source heterogeneous sensor, after time alignment and noise suppression, high-order semantic features are extracted through a multi-modal feature fusion model, and dynamic weight distribution is achieved through a cross-modal attention mechanism; and the cognitive state discrimination module outputs the concentration level, the cognitive load intensity and the knowledge mastering confidence, so that the learning content difficulty, the rhythm and the interaction form are dynamically adjusted, and a closed-loop personalized learning path is formed. According to the method, the recognition accuracy and the evaluation robustness of the real cognitive state of the student are remarkably improved, and the learning efficiency and the input degree are effectively optimized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a dynamic assessment method for student learning ability that integrates artificial intelligence. Background Technology

[0002] With the deep integration of artificial intelligence and educational technology, the refined monitoring and assessment of students' learning processes has become a core research direction in the field of smart education. Traditional learning assessment methods mainly rely on standardized test scores or teachers' subjective observations, which are difficult to reflect students' cognitive state and attention levels in real time and objectively. In recent years, learning analysis technologies based on behavioral data have gradually emerged, using cameras, keyboard logs, or wearable devices to collect single-modal signals (such as head posture and click frequency) to infer attention levels. However, such methods are easily affected by external interference or individual habitual actions, and cannot effectively distinguish "false focus"—that is, students maintain a static posture or continue operating, but have actually experienced attention drift or even inattentiveness, leading to distorted assessment results and seriously restricting the accuracy and timeliness of personalized teaching interventions.

[0003] Multimodal learning analytics, by fusing heterogeneous data from multiple sources including visual, physiological, and interactive behaviors, offers a new path for achieving robust attention state recognition. This technology aims to simultaneously capture students' explicit behaviors and implicit cognitive responses during learning tasks, leveraging cross-modal associations to uncover deep features of dynamic attentional changes. Micro-expression recognition, a key branch, can capture cognitive load and emotional fluctuations reflected in subtle movements of the brow muscles; eye tracking reveals the stability of visual attention focus through changes in pupil diameter and the distribution of fixation hotspots; and sequences of operational behaviors (such as temporal variations in keyboard typing intervals) can indirectly characterize the fluency of thought and the degree of task engagement. While each of these modalities carries complementary information, efficiently integrating their spatiotemporally asynchronous and scale-independent characteristics remains a major bottleneck for current technology implementation.

[0004] Existing technologies for multimodal fusion generally suffer from structural rigidity, weak temporal modeling capabilities, and shallow feature interactions. Most solutions employ simple weighted averaging or early splicing strategies, neglecting the differences in response delays and contribution weights of different modalities in attention drift events. While some deep learning models incorporate LSTM or CNN to process temporal data, they struggle to model long-distance dependencies and cross-modal dynamic alignment relationships. Especially in real classrooms or online learning scenarios, variations in lighting, device noise, and individual behavioral differences further exacerbate the uncertainty of modal data, resulting in insufficient sensitivity of the fusion results to distinguish complex states such as "pseudo-focus." Therefore, there is an urgent need for a dynamic evaluation mechanism that can adaptively perceive the intrinsic correlation between multimodal behavior and physiological signals and accurately capture the critical point of attention drift to support highly reliable real-time diagnosis of student learning abilities. Summary of the Invention

[0005] This invention provides a dynamic assessment method for student learning ability that integrates artificial intelligence, aiming to solve the technical problem in existing technologies that rely on single-modal data and cannot accurately distinguish students' "pseudo-focus" states. "Pseudo-focus" refers to a situation where a student outwardly maintains typical focused behaviors such as sitting still or gazing, but their cognitive attention is not truly concentrated on the learning task. This problem stems from existing assessment systems relying solely on video image analysis of facial orientation or eye gaze direction, or solely on behavioral outcome indicators such as answer accuracy, lacking the ability to simultaneously perceive and cross-validate multiple dimensions of physiological, behavioral, and cognitive states. This results in distorted assessment results that fail to accurately reflect students' learning engagement and cognitive load levels.

[0006] The dynamic assessment method for student learning ability integrating artificial intelligence described in this invention includes the following steps: synchronously collecting students' physiological signal data, behavioral posture data, and task interaction data during the learning process using multi-source heterogeneous sensors; performing time alignment and noise suppression processing on the collected multimodal raw data to generate a multi-channel time-series data stream under a unified time reference; constructing a multimodal feature fusion model based on a deep neural network, extracting high-order semantic features from each modality of data, and realizing dynamic weight allocation and semantic alignment between features through a cross-modal attention mechanism; inputting the fused multimodal representation into a cognitive state discrimination module, outputting the student's current cognitive focus level, cognitive load intensity, and knowledge mastery confidence level; and dynamically adjusting the difficulty gradient, presentation rhythm, and interaction form of subsequent learning content based on the cognitive state discrimination results to form a closed-loop feedback personalized learning path.

[0007] Furthermore, the multi-source heterogeneous sensor includes a non-contact infrared thermal imaging sensor, a high frame rate visible light camera, a capacitive heart rate monitoring pad, and an embedded learning terminal interactive log acquisition unit. The non-contact infrared thermal imaging sensor is deployed in front of the learning terminal to continuously acquire the temperature distribution sequence of the student's facial area, wherein the temperature change rate in the frontal lobe region is used to characterize the activation level of the prefrontal cortex; The high frame rate visible light camera captures the student's head posture angle, eyelid opening and closing frequency, pupil diameter changes, and gaze point coordinates at a sampling frequency of 60 frames per second; the capacitive heart rate monitoring pad is integrated into the surface of the learning chair and extracts the time and frequency domain features of heart rate variability by detecting chest cavity micro-movement signals; the embedded learning terminal interactive log collection unit records the student's key response latency, number of option modifications, page dwell time, and mouse trajectory entropy value during the question-answering process.

[0008] Furthermore, the time alignment and noise suppression processing includes: applying hardware-level timestamps to each sensor data stream with the start time of the learning task as global time 0; using a sliding window mid-value filter to remove impulse noise from physiological signals; applying a Kalman filter algorithm to the head pose angle sequence in the video stream to eliminate instantaneous jitter caused by sudden changes in illumination or occlusion; performing three-sigma criterion outlier removal on the response delay data in the interaction log and using linear interpolation to complete missing data segments; and finally resampling all modal data to a uniform sampling frequency to form a multi-channel time series matrix of consistent length.

[0009] Furthermore, the multimodal feature fusion model consists of three parallel encoding branches and one cross-modal interaction layer. The first encoding branch is a 1D convolutional neural network that receives a preprocessed physiological signal sequence and sequentially passes it through a 1D convolutional layer with a kernel size of 5 and a stride of 2, a batch normalization layer, and a modified linear unit activation function to extract the temporal patterns of heart rate variability and forehead temperature changes. The second encoding branch is a spatiotemporal graph convolutional network that models head pose angles, eye movement parameters, and gaze coordinates as joint graph structures that evolve over time and captures gaze-pose cooperative behavior patterns through a dynamic update mechanism of the adjacency matrix. The third encoding branch is a bidirectional gated recurrent unit network that processes the interaction log sequence. Its forward propagation captures the strategy evolution trend during the question-answering process, and its backward propagation backtracks the reflection path for error correction. The cross-modal interaction layer adopts a multi-head scaling dot product attention mechanism, using physiological signal features as query vectors, behavioral posture features and interaction features as key vectors and value vectors, respectively. It calculates the correlation weights between each modality in the cognitive attention dimension and then weights and fuses them to generate a unified multimodal context representation.

[0010] Furthermore, the cognitive state discrimination module includes three parallel output heads. The first output head is a fully connected classifier that receives multimodal contextual representations and outputs discrete levels of cognitive focus, which are divided into four categories: high focus, moderate focus, low focus, and distraction. The second output head is a regression network that outputs continuous values ​​of cognitive load intensity, ranging from 0 to 100, with higher values ​​indicating that working memory resources are closer to saturation. The third output head is a Bayesian belief network that combines the current correctness of the answer with the activation state of historical knowledge graph nodes to output a probability distribution of the degree of mastery of a specific knowledge point, with the output form being a Gaussian distribution parameter composed of the mean and variance.

[0011] Furthermore, the dynamic adjustment of the difficulty gradient, presentation rhythm, and interaction form of subsequent learning content specifically includes: when the cognitive focus level is inattentive or low focus and the cognitive load intensity is below 30, the system inserts micro-interactive gamification elements to reawaken attention; when the cognitive focus level is high focus but the cognitive load intensity exceeds 80, the system automatically splits the current complex task into multiple sub-tasks and extends the presentation time of each sub-task; when the variance of the knowledge mastery confidence is greater than 15, the system pushes targeted variation exercises to strengthen concept differentiation; when the response delay exceeds a preset threshold for three consecutive questions and the eyelid closure frequency is higher than 20 times per minute, the system triggers a short rest prompt and adjusts the ambient light color temperature to a cool tone to enhance alertness.

[0012] Furthermore, the method also includes an initialization phase for constructing an individualized baseline model. Before the formal assessment, the system guides students to complete a set of standardized calibration tasks covering reading comprehension, logical reasoning, and memory retrieval activities at different cognitive load levels. During this process, multimodal data is collected simultaneously and labeled with known cognitive state tags. The labeled dataset is then used to fine-tune the multimodal feature fusion model and the cognitive state discrimination module through transfer learning, adapting the model parameters to the specific student's circadian rhythm, behavioral habits, and cognitive style, thereby improving the individual sensitivity and discrimination accuracy of subsequent dynamic assessments.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multi-dimensional cross-validation mechanism for students' cognitive states by simultaneously integrating three heterogeneous data sources: physiological signals, behavioral postures, and task interactions. This effectively eliminates the risk of misjudgment caused by external interference or individual differences in a single modality, and significantly improves the accuracy of identifying "pseudo-focused" states. 2. This invention introduces a cross-modal attention mechanism to achieve dynamic semantic alignment at the feature level, enabling the system to adaptively focus on the modal combination that best reflects the true cognitive state, thereby enhancing the robustness and generalization ability of the evaluation model. 3. Based on the joint criterion of cognitive load intensity and knowledge mastery confidence, this invention realizes real-time closed-loop control of the difficulty, pace and interaction of learning content. This not only avoids the decline in learning efficiency caused by cognitive overload, but also prevents the distraction caused by overly simple tasks, thereby maintaining continuous cognitive input while ensuring the depth of learning. 4. This invention overcomes the performance degradation problem of general models when applied across individuals by using an individualized baseline model initialization process, thus ensuring the fairness and effectiveness of the evaluation results across different student groups. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multimodal feature fusion model in this invention; Figure 3 This is a flowchart illustrating the logical flow of multi-source heterogeneous sensor data acquisition and time alignment preprocessing in this invention. Figure 4 This is a diagram showing the multi-dimensional output structure and criterion logic framework of the cognitive state discrimination module in this invention; Figure 5 This is a closed-loop control framework diagram of the learning content dynamic adjustment mechanism based on cognitive state feedback in this invention; Figure 6 This is a flowchart illustrating the personalized adaptation process of individualized baseline model initialization and migration fine-tuning in this invention. Detailed Implementation

[0015] Please refer to the attached document. Figure 1 To be continued Figure 6 This invention provides a dynamic assessment method for student learning ability that integrates artificial intelligence, aiming to solve the technical problem in existing technologies that rely on single-modal data and cannot accurately distinguish students' "pseudo-focus" states. "Pseudo-focus" refers to a situation where a student outwardly maintains typical focused behaviors such as sitting still or gazing, but their cognitive attention is not truly concentrated on the learning task. This problem stems from existing assessment systems relying solely on video image analysis of facial orientation or eye gaze direction, or solely on behavioral outcome indicators such as answer accuracy, lacking the ability to simultaneously perceive and cross-validate physiological, behavioral, and cognitive states across multiple dimensions. This results in distorted assessment results that fail to accurately reflect students' learning engagement and cognitive load levels.

[0016] The dynamic assessment method for student learning ability integrating artificial intelligence as described in this invention includes the following steps: S1 synchronously collects students' physiological signal data, behavioral posture data, and task interaction data during the learning process through multi-source heterogeneous sensors; S2 performs time alignment and noise suppression processing on the collected multimodal raw data to generate a multi-channel time-series data stream under a unified time reference; S3 is a multimodal feature fusion model built on deep neural networks. It extracts high-order semantic features from each modality of data and achieves dynamic weight allocation and semantic alignment between features through a cross-modal attention mechanism. S4 inputs the fused multimodal representation into the cognitive state discrimination module and outputs the student's current cognitive attention level, cognitive load intensity, and knowledge mastery confidence level. S5. Based on the cognitive state judgment result, dynamically adjust the difficulty gradient, presentation rhythm and interaction form of subsequent learning content to form a closed-loop feedback personalized learning path.

[0017] In step S1, physiological signal data, behavioral posture data, and task interaction data of students during the learning process are simultaneously collected using multi-source heterogeneous sensors. These multi-source heterogeneous sensors include a non-contact infrared thermal imaging sensor, a high-frame-rate visible light camera, a capacitive heart rate monitoring pad, and an embedded learning terminal interaction log acquisition unit.

[0018] The non-contact infrared thermal imaging sensor is deployed in front of the learning terminal to continuously acquire temperature distribution sequences of the student's facial area. The rate of temperature change in the frontal lobe region is used to characterize the activation level of the prefrontal cortex. The prefrontal cortex is a core brain region for working memory and executive control; increased metabolic activity in this area is usually accompanied by increased local blood flow, leading to a slight increase in skin surface temperature. The system records facial thermal images at a sampling frequency of 10 frames per second and uses a preset facial landmark localization algorithm to locate the area between the eyebrows and the central forehead region. The average grayscale value sequence of pixels in this region is extracted and then converted into a Celsius temperature sequence through radiometric calibration.

[0019] The high frame rate visible light camera captures students' head posture angles, eyelid opening and closing frequency, pupil diameter changes, and gaze point coordinates at a sampling frequency of 60 frames per second. The camera incorporates an automatic exposure and white balance module to ensure stable image quality under different lighting conditions. The image data stream is input to a real-time face detection and key point tracking module. This module employs a lightweight convolutional neural network structure to locate 68 facial key points in each frame and calculates the Euler angles of the head around the X-axis (pitch), Y-axis (yaw), and Z-axis (roll). Simultaneously, it estimates the pupil diameter using an iris-sclera boundary fitting algorithm and calculates the eyelid opening and closing ratio based on the distance between the upper and lower eyelid key points. When this ratio remains below 0.3 for more than 1.5 seconds, it is considered a blink event. The gaze point coordinates are back-projected onto the screen coordinate system using a 3D eyeball model, with an accuracy error controlled within 2% of the screen diagonal length.

[0020] The capacitive heart rate monitoring pad is integrated into the surface of the learning chair. It extracts the time and frequency domain features of heart rate variability by detecting micro-motion signals in the chest cavity. The pad contains a high-sensitivity capacitive sensing array that can sense millimeter-level displacements in the chest cavity caused by the heartbeat. After the signal is filtered by a bandpass filter to remove respiratory baseline drift, the R-wave interval sequence is extracted using a peak detection algorithm. Then, heart rate variability indicators such as standard deviation SDNN, root mean square of the difference between adjacent RR intervals RMSSD, low-frequency power LF, and high-frequency power HF are calculated.

[0021] The embedded learning terminal's interaction log collection unit records the student's key response latency, number of option modifications, page dwell time, and mouse trajectory entropy during the question-answering process. The interaction log records each user operation event with a millisecond-level timestamp, including mouse movement, clicks, keyboard input, and page switching. The system calculates the time interval from question presentation to the first answer as the response latency; it counts the number of option switches for the same question as a strategy hesitation index; page dwell time is defined as the time span from entering the page to submitting an answer or jumping to the next page; the mouse trajectory entropy is calculated by segmenting the trajectory point sequence into a sliding window, calculating the standard deviation of the direction angle of each trajectory segment, and then calculating the global average, to measure the randomness of the exploration behavior.

[0022] In step S2, the acquired multimodal raw data undergoes time alignment and noise suppression processing to generate a multi-channel time-series data stream with a unified time reference. Taking the start time of the learning task as global time 0, all sensors receive the same trigger signal at the hardware level to initiate data acquisition and embed a timestamp with nanosecond-level precision.

[0023] The system first removes impulse noise from physiological signals: a sliding window median filter is applied to the heart rate variability sequence, with a window length of 5 seconds. If a certain RR interval deviates from the median of the window by more than 30%, it is considered an outlier and replaced with a linear interpolation result of the nearest valid value; wavelet thresholding is used for forehead temperature sequences. The wavelet basis function is used for three-level decomposition, and the signal is reconstructed after applying soft thresholding to the high-frequency coefficients.

[0024] Kalman filtering is applied to the head pose angle sequence in the video stream to eliminate transient jitter caused by sudden changes in lighting or occlusion. The state vector of the Kalman filter contains the current pose angle and its first derivative. The observation model is directly taken from the output of the visual tracking module, and the process noise covariance matrix is ​​dynamically adjusted based on the historical motion smoothness. Three-sigma criterion outlier removal is performed on the response latency data in the interaction log: the mean of all response latency is calculated. with standard deviation Remove those larger than +3 The data points were collected, and missing data segments were filled using linear interpolation, with the interpolation range not exceeding three consecutive points. Finally, all modal data were resampled to a uniform sampling frequency of 10 Hz to form a multi-channel time series matrix of consistent length. Each row of this matrix corresponds to a time step, and each column corresponds to a feature channel, with a total of 27 dimensions: 8 dimensions for physiological signals, 12 dimensions for behavioral postures, and 7 dimensions for interaction logs.

[0025] In step S3, a multimodal feature fusion model is constructed based on a deep neural network to extract high-order semantic features from each modality of data, and dynamic weight allocation and semantic alignment between features are achieved through a cross-modal attention mechanism. The multimodal feature fusion model consists of three parallel encoding branches and one cross-modal interaction layer.

[0026] The first encoding branch is a 1D convolutional neural network that receives a preprocessed physiological signal sequence and sequentially passes it through a 1D convolutional layer with a kernel size of 5 and a stride of 2, a batch normalization layer, and a modified linear unit activation function, stacked in four layers. Each layer outputs 16, 32, 64, and 128 channels respectively, ultimately outputting a 128-dimensional physiological feature vector. This branch effectively captures the temporal patterns of heart rate variability and forehead temperature changes; for example, a deceleration of the heart rate accompanied by an increase in forehead temperature often indicates concentrated cognitive resource allocation.

[0027] The second encoding branch is a spatiotemporal graph convolutional network, which models head pose angles, eye movement parameters, and gaze coordinates as an articulated graph structure that evolves over time. Graph nodes include four key locations: head center, eye center, nose tip, and gaze point. Edge weights are dynamically calculated based on the Euclidean distance between nodes. Graph convolution operations alternate between time and space dimensions. The temporal convolution kernel size is 3, and the spatial adjacency matrix is ​​updated through a learnable attention mechanism to enhance the representation of gaze-pose cooperative behavior patterns, such as prolonged gazing combined with slight head tilt to indicate depth understanding.

[0028] The third encoding branch is a bidirectional gated recurrent unit network (GRU) that processes the interaction log sequence, with its hidden unit dimension set to 128. Forward propagation captures the strategy evolution trend during the answering process; for example, a gradually shortening response latency indicates increased proficiency. Backward propagation backtracks the reflection path of error correction; for example, choosing the original answer after multiple modifications may reflect self-doubt. The feature vectors output from the three branches are linearly projected to the same 256-dimensional plane and then input to the cross-modal interaction layer. The cross-modal interaction layer employs a multi-head scaling dot product attention mechanism with 8 heads. Physiological signal features are used as the query vector. Behavioral posture features are key vectors Interactive features are value vectors Calculate attention weights: ; in The dimension is the key vector. This mechanism dynamically evaluates the correlation strength of information from different modalities under the current cognitive state. For example, when a student is under high load, the weight of heart rate variability features in physiological signals increases, while the weight of mouse trajectory entropy values ​​in interaction logs decreases. After weighted fusion, a unified multimodal context representation is generated with a dimension of 256.

[0029] In step S4, the fused multimodal representation is input to the cognitive state discrimination module, which outputs the student's current cognitive attention level, cognitive load intensity, and knowledge mastery confidence level. The cognitive state discrimination module contains three parallel output heads.

[0030] The first output head is a fully connected classifier, consisting of two fully connected layers. The middle layer has a dimension of 128, and the output layer has a dimension of 4, corresponding to four categories: high focus, moderate focus, low focus, and distraction. The classifier is trained using the cross-entropy loss function, and the output is activated by softmax to obtain the probability of each category.

[0031] The second output head is a regression network with a single fully connected layer structure. It outputs continuous numerical values ​​of cognitive load intensity, ranging from 0 to 100. This network uses the mean squared error loss function and performs min-max normalization on the labels during training.

[0032] The third output head is a Bayesian belief network that receives multimodal representations and knowledge point identifiers corresponding to the current question, and outputs a probability distribution of the degree of mastery of a specific knowledge point. This network consists of two fully connected layers and outputs the mean. With log variance The loss function is the negative log-likelihood: ; in, The values ​​represent the correctness of the answers (1 for correct, 0 for incorrect). High variance indicates a high degree of uncertainty regarding the system's mastery of the knowledge point.

[0033] In step S5, based on the cognitive state assessment result, the difficulty gradient, presentation rhythm, and interaction format of subsequent learning content are dynamically adjusted to form a closed-loop feedback-based personalized learning path. When the cognitive focus level is inattentive or low focus and the cognitive load intensity is below 30, the system inserts micro-interactive gamification elements to reawaken attention, such as popping up a quick reaction mini-game within 3 seconds, requiring students to click on a flashing icon; when the cognitive focus level is high focus but the cognitive load intensity exceeds 80, the system automatically splits the current complex task into multiple sub-tasks and extends the presentation time of each sub-task, for example, breaking down a multi-step math problem into 3 independent prompt boxes, increasing the dwell time of each box by 50%. When the variance of the confidence level of knowledge mastery is greater than 15, the system pushes targeted variation exercises to strengthen concept differentiation. For example, when algebraic concepts are confused, it provides comparative questions with symbol substitution but the same structure. When the response delay exceeds the preset threshold for three consecutive times and the eyelid closure frequency is higher than 20 times per minute, the system triggers a short rest prompt and adjusts the ambient light color temperature to a cool tone to enhance alertness. The rest prompt lasts for 30 seconds, during which the screen displays a deep breathing guided animation.

[0034] The method also includes an initialization phase for building an individualized baseline model. Before the formal assessment, the system guides students to complete a set of standardized calibration tasks covering reading comprehension, logical reasoning, and memory retrieval activities at different cognitive load levels. These tasks comprise 12 sub-tasks, each lasting 3 minutes. During this process, multimodal data is simultaneously collected and labeled with known cognitive state tags: reading comprehension is labeled as moderate load, logical reasoning as high load, and memory retrieval as low load; attention level tags are determined independently by two trained observers and then agreed upon.

[0035] This labeled dataset was used to fine-tune the multimodal feature fusion model and the cognitive state discrimination module through transfer learning. The fine-tuning employed a two-stage strategy: in the first stage, the cross-modal interaction layer was frozen, and only three encoding branches and the output head were trained with a learning rate of 0.001; in the second stage, all parameters were unfrozen, the learning rate was reduced to 0.0001, and the number of training epochs did not exceed 10. After fine-tuning, the model parameters were adapted to the specific students' circadian rhythms, behavioral habits, and cognitive styles. For example, some students habitually blink frequently while thinking, which a general model might misjudge as inattentiveness. However, the individualized model learns the positive correlation between this behavior and high cognitive load through calibration data, thereby improving discrimination accuracy.

[0036] The system implementation of the above method relies on an integrated learning and evaluation platform. This platform consists of a front-end perception layer, a data processing layer, an intelligent inference layer, and a feedback execution layer. The front-end perception layer includes the aforementioned four types of sensors, connected to the edge computing host via Gigabit Ethernet or USB 3.0 interfaces. The data processing layer runs on a Linux operating system, employing ROS middleware to manage multi-source data streams, achieving timestamp synchronization, buffer queue management, and preprocessing algorithm scheduling. The intelligent inference layer is deployed on a GPU-accelerated server, loading a trained deep neural network model, receiving the preprocessed temporal matrix, and performing feature extraction, fusion, and state determination, with inference latency controlled within 200 milliseconds.

[0037] The feedback execution layer is integrated into the learning content management system. It receives the cognitive state assessment results, calls the content adjustment strategy engine, and modifies the presentation logic of the course script in real time. The system supports offline mode, locally caching the individualized model parameters for the past 7 days in environments without network access to ensure assessment continuity. All data is encrypted using AES 256-bit during transmission and storage, complying with educational data privacy protection standards.

[0038] This embodiment constructs a multi-dimensional cross-validation mechanism for students' cognitive states by simultaneously fusing three heterogeneous data sources: physiological signals, behavioral postures, and task interactions. This effectively eliminates the risk of misjudgment caused by external interference or individual differences in a single modality, significantly improving the accuracy of identifying "pseudo-focused" states. A cross-modal attention mechanism is introduced to achieve dynamic semantic alignment at the feature level, enabling the system to adaptively focus on the modal combination that best reflects the true cognitive state, enhancing the robustness and generalization ability of the evaluation model. Based on a joint criterion of cognitive load intensity and knowledge mastery confidence, real-time closed-loop control of learning content difficulty, pace, and interaction form is achieved. This not only avoids a decline in learning efficiency due to cognitive overload but also prevents inattention caused by overly simple tasks, thus maintaining continuous cognitive engagement while ensuring learning depth. Through the initialization process of an individualized baseline model, the performance degradation problem of general models in cross-individual applications is overcome, ensuring the fairness and effectiveness of evaluation results across different student groups.

Claims

1. A dynamic assessment method for student learning ability integrating artificial intelligence, characterized in that, include: The physiological signal data, behavioral posture data, and task interaction data of students during the learning process are collected synchronously through multi-source heterogeneous sensors. The collected multimodal raw data is processed for time alignment and noise suppression to generate a multi-channel time-series data stream under a unified time reference. A multimodal feature fusion model is constructed based on deep neural networks, which extracts high-order semantic features from each modality of data, and realizes dynamic weight allocation and semantic alignment among features through a cross-modal attention mechanism; The fused multimodal representation is input into the cognitive state discrimination module, which outputs the student's current cognitive attention level, cognitive load intensity, and knowledge mastery confidence level. Based on the cognitive state assessment results, the difficulty gradient, presentation pace, and interaction format of subsequent learning content are dynamically adjusted.

2. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 1, characterized in that, The multi-source heterogeneous sensor includes a non-contact infrared thermal imaging sensor, a high frame rate visible light camera, a capacitive heart rate monitoring pad, and an embedded learning terminal interactive log acquisition unit. The non-contact infrared thermal imaging sensor is used to continuously acquire the temperature distribution sequence of the student's facial area, wherein the temperature change rate in the frontal lobe region is used to characterize the activation level of the prefrontal cortex. The high frame rate visible light camera captures the student's head posture angle, eyelid opening and closing frequency, pupil diameter changes, and coordinates of the gaze point at a sampling frequency of 60 frames per second. The capacitive heart rate monitoring pad is integrated into the surface of the learning chair and extracts the time and frequency domain features of heart rate variability by detecting micro-movement signals in the chest cavity. The embedded learning terminal interaction log collection unit records the key response latency, number of option modifications, page dwell time, and mouse trajectory entropy value of students during the question-answering process.

3. The dynamic assessment method for student learning ability integrating artificial intelligence according to claim 2, characterized in that, The collected multimodal raw data undergoes time alignment and noise suppression processing to generate a multi-channel time-series data stream with a unified time reference, including: With the start time of the learning task as global time 0, hardware-level timestamps are applied to the data streams of each sensor. A sliding window midpoint filter is used to remove impulse noise from physiological signals; Kalman filtering is applied to the head pose angle sequence in the video stream to eliminate instantaneous jitter caused by sudden changes in lighting or occlusion. Three Sigma criteria were applied to remove outliers from the response latency data in the interaction logs, and missing data segments were filled using linear interpolation. All modal data are resampled to a uniform sampling frequency to form a multi-channel time series matrix of consistent length.

4. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 3, characterized in that, A multimodal feature fusion model is constructed based on deep neural networks. High-order semantic features are extracted from data in each modality, and a cross-modal attention mechanism is used to achieve dynamic weight allocation and semantic alignment among features, including: The preprocessed physiological signal sequence is received by a 1D convolutional neural network, and then passed through a 1D convolutional layer with a kernel size of 5 and a stride of 2, a batch normalization layer, and a modified linear unit activation function to extract the temporal patterns of heart rate variability and forehead temperature changes. The head pose angle, eye movement parameters and gaze coordinates are modeled as joint graph structures that evolve over time by using a spatiotemporal graph convolutional network, and gaze-attention cooperative behavior patterns are captured through a dynamic update mechanism of the adjacency matrix. The interactive log sequence is processed by a bidirectional gated recurrent unit network. Its forward propagation captures the strategy evolution trend during the question-answering process, and its backward propagation backtracks the reflection path for error correction. The feature vectors output from the three encoding branches are linearly projected to the same dimension and then input into the cross-modal interaction layer. The cross-modal interaction layer adopts a multi-head scaling dot product attention mechanism, using physiological signal features as query vectors, behavioral posture features and interaction features as key vectors and value vectors, respectively. The correlation weights between each modality in the cognitive attention dimension are calculated and weighted and fused to generate a unified multimodal context representation.

5. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 4, characterized in that, The fused multimodal representation is input into the cognitive state discrimination module, which outputs the student's current cognitive attention level, cognitive load intensity, and knowledge mastery confidence level, including: The multimodal contextual representation is received through a fully connected classifier, and the discrete levels of cognitive attention are output. The levels are divided into four categories: high attention, moderate attention, low attention, and inattentiveness. The regression network outputs a continuous numerical value of cognitive load intensity, ranging from 0 to 100. By combining the current correctness of the answer with the activation status of nodes in the historical knowledge graph through a Bayesian belief network, the probability distribution of the mastery of a specific knowledge point is output. The output is in the form of a Gaussian distribution parameter composed of the mean and variance.

6. The dynamic assessment method for student learning ability integrating artificial intelligence according to claim 5, characterized in that, Based on the cognitive state assessment results, the difficulty gradient, presentation pace, and interaction format of subsequent learning content are dynamically adjusted, including: When the cognitive attention level is inattentive or low focus and the cognitive load intensity is below 30, the system inserts micro-interactive gamification elements to re-engage attention. When the cognitive focus level is high but the cognitive load intensity exceeds 80, the system automatically splits the current complex task into multiple sub-tasks and extends the presentation time of each sub-task. When the variance of the confidence level of knowledge mastery is greater than 15, the system pushes targeted variation exercises to strengthen concept differentiation; When the response time exceeds the preset threshold for three consecutive questions and the eyelid closure frequency is higher than 20 times per minute, the system triggers a short rest prompt and adjusts the ambient light color temperature to a cool tone to enhance alertness.

7. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 6, characterized in that, The method also includes an initialization phase for constructing a personalized baseline model, including: Prior to the formal assessment, students are guided to complete a set of standardized calibration tasks, which cover reading comprehension, logical reasoning, and memory retrieval activities at different cognitive load levels. During the calibration task, multimodal data is collected synchronously and labeled with known cognitive state tags; The multimodal feature fusion model and cognitive state discrimination module were fine-tuned through transfer learning using this labeled dataset, so that the model parameters could be adapted to the physiological rhythms, behavioral habits and cognitive styles of specific students.

8. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 7, characterized in that, The standardized calibration task consists of 12 sub-tasks, each lasting 3 minutes; The cognitive state labels include: reading comprehension tasks are labeled as moderate load, logical reasoning as high load, and memory retrieval as low load; the focus label is determined by consensus after independent evaluation by two trained observers.

9. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 8, characterized in that, The transfer learning fine-tuning employs a two-stage strategy: In the first stage, the cross-modal interaction layer is frozen, and only 3 encoding branches and output heads are trained with a learning rate of 0.

001. In the second stage, all parameters are unfrozen, the learning rate is reduced to 0.0001, and the number of training rounds does not exceed 10.

10. The method for dynamic evaluation of student learning ability integrating artificial intelligence according to claim 4, characterized in that, The 1D convolutional neural network consists of four stacked layers, with each layer having 16, 32, 64, and 128 output channels, respectively, ultimately outputting a physiological feature vector with a dimension of 128.

Images