Adaptive learning path recommendation system and head-mounted device

Abstract

The invention discloses a self-adaptive learning path recommendation system and a head-mounted device, and belongs to the technical field of intelligent education, the system comprises a multi-modal data acquisition module, a signal feature fusion module, a knowledge graph management module and a self-adaptive path recommendation module, electroencephalogram signals, visual behaviors and learning content data are synchronously collected through a head-mounted device, a fusion feature vector is generated by adopting a signal quality adaptive weighted fusion strategy, the knowledge mastery degree is evaluated based on a deep learning model, and a dynamic knowledge graph is updated in real time. According to the method, the recommendation priority is calculated by integrating the mastery gap, the knowledge importance, the dependence emergency degree, the forgetting risk and the fatigue state, dynamic planning and closed-loop adjustment of a personalized learning path are realized, the mastery evaluation accuracy and the learning efficiency are remarkably improved, and meanwhile, the data privacy security is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to an adaptive learning path recommendation system and head-mounted device based on multimodal physiological signal monitoring and dynamic knowledge graph, which is suitable for scenarios such as personalized learning, intelligent tutoring and cognitive ability training. Background Technology

[0002] With the rapid development of online education and intelligent tutoring technologies, personalized learning path recommendation has become a key technical means to improve learning efficiency and optimize the learning experience. However, existing learning path recommendation systems still have significant technical limitations in areas such as knowledge state perception, adaptive recommendation strategies, and the design of learning monitoring devices.

[0003] At the level of learning status perception, existing online education platforms primarily rely on statistical data such as answer accuracy and time taken to estimate learners' knowledge mastery. This approach has significant accuracy limitations. Firstly, statistical data fails to reflect learners' real-time cognitive state; learners with the same mastery score may exhibit drastically different cognitive characteristics—for example, one may understand but be careless, while another may rely on rote memorization. Secondly, mastery assessments based solely on answer results are easily influenced by interfering factors such as guessing correctly or plagiarism; research indicates that such interference can lead to a misjudgment rate of approximately 30%. Furthermore, existing systems generally lack personalized modeling of learners' forgetting patterns and fail to effectively integrate research findings in memory science to predict forgetting trends and optimal review times.

[0004] At the recommendation strategy level, existing learning path recommendation algorithms mostly employ collaborative filtering or content-based recommendation methods, which suffer from static recommendation strategies and insufficient consideration of multiple factors. Specifically, learning paths are often determined at the beginning of learning and cannot be dynamically adjusted according to changes in the learner's state during the learning process. When learners experience changes in their state, such as fatigue, inattention, or sudden comprehension, the system cannot respond in a timely manner and adjust the recommended content and difficulty.

[0005] At the level of learning monitoring devices, existing wearable EEG monitoring devices generally suffer from limited functionality and design flaws. Most devices can only collect data from a single modality, lacking the ability to simultaneously collect multimodal information such as EEG, vision, and behavior. In terms of hardware layout, existing devices often have cameras positioned on the side or lack cameras altogether, making it impossible to simultaneously capture learning materials and handwriting actions. Furthermore, existing devices generally employ a data acquisition and transmission model, with all calculations performed in the cloud, resulting in feedback latency typically exceeding 500ms, making true real-time interaction difficult.

[0006] To address the aforementioned technical issues, this invention provides an innovative system solution that aims to achieve truly personalized, real-time responsive intelligent learning path planning through real-time fusion analysis of multimodal physiological signals, precise updating of dynamic knowledge graphs, and a multi-factor closed-loop adaptive recommendation mechanism. Furthermore, through innovative head-mounted device hardware design and edge computing architecture, it provides low-latency intelligent tutoring services while ensuring data privacy. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive learning path recommendation system and a head-mounted device to solve the problems of insufficient accuracy in learning state perception, static recommendation strategies, and limited functionality of monitoring devices in the prior art, and to achieve real-time learning state perception and dynamic path optimization based on multimodal physiological signals.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An adaptive learning path recommendation system includes a multimodal data acquisition module, a signal feature fusion module, a knowledge graph management module, and an adaptive path recommendation module. The multimodal data acquisition module synchronously collects learners' EEG signal data, visual behavior data, and learning content data via a head-mounted device, preprocesses and extracts features from the raw data, and generates feature vectors for each modality. The signal feature fusion module receives the feature vectors from each modality, dynamically calculates adaptive weight coefficients based on the signal quality of each modality, and performs a weighted fusion operation on the feature vectors of each modality using the adaptive weight coefficients to output a fused feature vector. The knowledge graph management module maintains a knowledge graph database, calculates the learner's mastery score for the current knowledge point using a deep learning inference model, updates the mastery status attributes of corresponding knowledge point nodes in the knowledge graph in real time, and calculates the forgetting prediction value for each knowledge point based on a forgetting curve model. The adaptive path recommendation module calculates the recommendation priority value for each knowledge point based on the mastery score, forgetting prediction value, and dependency data, combined with real-time fatigue state indicators, generates a learning path sequence, and triggers a dynamic adjustment strategy during the learning process to adjust the content and difficulty parameters of the learning path in real time.

[0010] A head-mounted device includes a main body, an EEG sensor array, a visual acquisition device, and an edge computing unit, for use in conjunction with the aforementioned adaptive learning path recommendation system. The main body of the head-mounted device features an arc-shaped design and flexible materials, with an overall weight not exceeding 120g. The EEG sensor array is located in the frontal lobe region and includes at least three dry electrodes. The visual acquisition device is located at the center of the front of the main body of the head-mounted device, with a camera module tilted downwards at 45° and equipped with an infrared illumination module. The edge computing unit is used to run a lightweight inference model to achieve local real-time mastery assessment.

[0011] The beneficial effects of this invention include: significantly improving the accuracy of mastery assessment through multimodal signal fusion, from 60%-70% of single-modality assessment to 85%-95%; maintaining high accuracy even in low-light or signal interference scenarios through an adaptive weighting mechanism for signal quality; increasing knowledge retention from 52% to 78% through forgetting curve modeling and timely review reminders; improving learning efficiency by 38% through recommendation priority calculation that comprehensively considers five factors; achieving real-time feedback with inference latency of less than 100ms through an edge computing architecture; and effectively protecting user privacy through a local priority processing strategy. Attached Figure Description

[0012] Figure 1 This is the overall architecture diagram of the adaptive learning path recommendation system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0014] Reference Figure 1 As shown, the adaptive learning path recommendation system provided by this invention includes a multimodal data acquisition module 1, a signal feature fusion module 2, a knowledge graph management module 3, an adaptive path recommendation module 4, an AI course generation module 5, and a privacy protection module 6. These modules form a deeply coupled closed-loop collaborative architecture. The output of one module serves as the key input to the next, and the processing results of the subsequent modules influence the working parameters of the preceding modules through feedback channels, thereby achieving a continuously optimized adaptive learning experience.

[0015] The multimodal data acquisition module 1 is responsible for the core function of simultaneously acquiring three types of data: learner's electroencephalogram (EEG) signals, visual behavior, and learning content. In a preferred embodiment of the invention, this module achieves hardware-level data acquisition through a specially designed head-mounted device, and performs preliminary signal preprocessing and feature extraction at the edge. The design concept of this module originates from the multimodal integration principle of cognitive neuroscience. By simultaneously capturing brain electrical activity, external behavioral performance, and learning material content, it can comprehensively characterize the learner's cognitive state from multiple dimensions. Compared to traditional single-modal monitoring schemes, multimodal data acquisition schemes have stronger robustness and higher evaluation accuracy; even if the data quality of one modality is compromised, other modalities can still provide effective supplementary information.

[0016] In terms of EEG signal acquisition, this invention adopts an electrode layout scheme conforming to the international 10-20 system standard. The international 10-20 system is a widely recognized electrode positioning standard in the field of EEG, and its naming rules are based on the percentage position of the electrode from anatomical landmarks in the head. Preferably, the EEG sensor array is configured with three core electrodes, respectively located at Fp1, Fp2, and Fz, corresponding to the left frontal lobe, right frontal lobe, and midfrontal region. The selection of the frontal lobe region has important neuroscientific basis because the prefrontal cortex is the core brain region for executive function, working memory, and attention control, and is closely related to the learning process. In application scenarios requiring higher spatial resolution, it can be expanded to five or eight electrodes, adding positions such as F3, F4, F7, and F8, which correspond to the left and right lateral frontal lobe regions, respectively, providing richer spatial information of the EEG. The electrodes adopt a dry electrode design, preferably made of gold-plated or silver / silver chloride material, with contact impedance controlled below 10kΩ. Compared to the cumbersome process of applying conductive gel with traditional wet electrodes, the dry electrode design significantly improves ease of use and user acceptance, making it suitable for extended use in daily learning scenarios. The sampling rate is set to 500Hz, which, according to the Nyquist sampling theorem, can effectively capture brain activity in the frequency range of 0Hz to 250Hz, fully covering the main frequency bands of cognitive-related brain activity. The resolution uses 16-bit quantization precision, corresponding to a quantization step size of approximately 0.2μV, accurately recording brain signals with amplitudes typically ranging from 10μV to 100μV.

[0017] The preprocessing of EEG signals involves multiple steps, each addressing a specific type of interference. First, the raw signal undergoes bandpass filtering, retaining effective components within the 0.5Hz to 45Hz range while filtering out DC drift and high-frequency noise. A low-frequency cutoff of 0.5Hz effectively eliminates slow baseline drift caused by electrode polarization potential drift and respiratory movements, while a high-frequency cutoff of 45Hz filters out electromyographic interference and high-frequency noise generated by electronic devices. The filter employs a fourth-order Butterworth infinite impulse response design, featuring a flat passband and a steep transition band, while phase delay can be compensated for using forward-backward filtering techniques. Subsequently, a 50Hz or 60Hz notch filter is used to eliminate power frequency interference. With a quality factor of 30, the notch filter accurately removes power frequency interference with minimal impact on adjacent frequency signals. In the artifact removal stage, this invention employs an automatic artifact detection and removal method based on independent component analysis, effectively separating and eliminating eye-tracking and electromyographic artifact components. Independent component analysis assumes that the observed signal is a linear mixture of multiple statistically independent source signals, achieving blind source separation by maximizing the non-Gaussianity of the components. Eye movement artifacts typically manifest as high-amplitude, low-frequency activity in the forehead channel, with temporal characteristics highly correlated with blinking or eye movement events; electromyographic artifacts, on the other hand, exhibit broadband, high-frequency activity. This invention employs an automatic artifact component identification algorithm, classifying components based on their temporal-frequency characteristics and spatial patterns, achieving an accuracy rate exceeding 95%. Preferably, adaptive filtering technology can be combined, using auxiliary channel signals as a reference for real-time artifact removal; this technique is particularly suitable for handling motion artifacts related to head movements.

[0018] EEG feature extraction encompasses three levels: frequency domain, time domain, and time-frequency domain, constructing a comprehensive cognitive state representation system. In the frequency domain, power spectral density is calculated using short-time Fourier transform or the Welch method, and the average power values ​​of the Delta band (0.5Hz to 4Hz), Theta band (4Hz to 8Hz), Alpha band (8Hz to 13Hz), Beta band (13Hz to 30Hz), and Gamma band (30Hz to 45Hz) are extracted to form 5-dimensional frequency domain features. The power characteristics of each frequency band have clear cognitive neuroscience significance: Delta waves mainly appear in deep sleep; their enhancement during waking learning may indicate extreme fatigue or a severe decline in attention; Theta waves are associated with working memory encoding and episodic memory retrieval, and their power usually increases when learning new knowledge; Alpha waves are considered a marker of cortical idleness, and their power is inhibited during attentional concentration and enhanced during relaxation or fatigue; Beta waves are associated with active thinking, problem-solving, and alertness, and their power increases during cognitive task performance; Gamma waves are associated with higher cognitive functions such as feature binding, perceptual integration, and conscious processing. In one embodiment of the invention, the θ / α power ratio and β / α power ratio are also calculated as indicators of cognitive load. An increased θ / α ratio generally indicates increased cognitive load, and studies have shown that this ratio is positively correlated with task difficulty and mental effort, while the β / α ratio is related to the degree of attentional concentration, and its increase indicates enhanced alertness.

[0019] At the temporal level, peak amplitude and latency characteristics of P300 and N400 event-related potentials (ERPs) were extracted for key cognitive events in learning tasks. ERPs are time-locked neural responses of the brain to specific stimuli or events, accurately reflecting the temporal progression of cognitive processing. The P300 component is associated with attention allocation and working memory updating, with amplitudes typically ranging from 5 μV to 20 μV and latency from 250 ms to 500 ms. The P300 amplitude is negatively correlated with the subjective probability of the stimulus and its task relevance; that is, the P300 amplitude is larger for rare or important stimuli. In learning scenarios, learners typically generate a larger P300 response when they successfully identify and understand key knowledge points. The N400 component is associated with semantic processing and shows enhanced performance when learners encounter semantic conflict or comprehension difficulties. The N400 is a negatively trending ERP component with a peak latency of approximately 400 ms, and its amplitude is positively correlated with the degree of semantic expectation violation. In a learning context, when a learner's understanding of a concept is inconsistent with the content presented in the material, a relatively large N400 response will be generated. This feature can be used to detect learners' comprehension difficulties.

[0020] At the time-frequency domain level, continuous wavelet transform is used to extract the time-frequency feature matrix, which can capture the transient characteristics of EEG signals. Compared with short-time Fourier transform based on a fixed window, continuous wavelet transform achieves adaptive time-frequency resolution through variable-scale basis wavelet functions, exhibiting higher frequency resolution in the low-frequency band and higher time resolution in the high-frequency band, which is more consistent with the non-stationary characteristics of EEG signals. This invention uses Morlet wavelet as the mother wavelet function, with a time-frequency product value of 6, achieving a good balance between time and frequency resolution. The squared modulus of the wavelet transform coefficients represents the instantaneous power at each time-frequency point. By statistically analyzing the power in a specific time-frequency region, features reflecting transient changes in cognitive state can be extracted. Combining the features from all the above levels, a 30-dimensional EEG feature vector is formed, comprehensively characterizing the learner's real-time cognitive state.

[0021] In terms of visual data acquisition, this invention employs an innovative 45° tilted camera layout, a key hardware innovation. Traditional head-mounted learning monitoring devices typically place the camera on the side or vertically downward. A side-mounted layout can only capture the learner's facial expressions, not the learning material; while a vertically downward layout can capture the learning material, its narrow field of view makes it difficult to simultaneously cover the hand movement area. This invention places the camera at the center of the front of the head-mounted device, tilted downward at 45°, allowing for ±15° angle adjustments. This layout design enables a single camera to simultaneously cover both the learning material and the hand movement area, achieving the synchronous monitoring concept of "what you see is what you learn." Preferably, the camera resolution is set to 1920×1080, the frame rate is 30fps, and the field of view is 120°. The high resolution ensures the clarity of the text recognized by OCR, the 30fps frame rate meets the temporal resolution requirements for hand movement tracking, and the 120° wide-angle field of view provides ample spatial coverage. At a distance of 30cm from the learner, the field of view covers approximately 35cm wide and 25cm high, sufficient to encompass a standard A4 sheet of paper and the entire handwriting area. The camera uses a 1 / 3-inch CMOS image sensor with a pixel size of 1.4μm, offering excellent low-light performance. The lens has a focal length of 2.8mm and an aperture of F / 2.0; this large aperture design allows for greater light capture in low-light conditions, and combined with a high ISO sensitivity setting, it achieves acceptable image quality even in dimly lit learning environments. The lens supports phase-detection autofocus with a focus response time of less than 200ms, enabling rapid tracking of changes in the distance to the learning material.

[0022] The visual data processing flow begins with image stabilization. Because the head-mounted device is tightly integrated with the learner's head, even minor head movements directly translate into image jitter. Without compensation, this severely impacts the stability and accuracy of OCR recognition. This invention utilizes integrated 6-axis IMU gyroscope data and employs an electronic image stabilization algorithm to compensate for image jitter caused by minor head movements, ensuring the stability of OCR recognition. The electronic image stabilization algorithm first reads angular velocity data from the IMU, calculates the head's rotational angle offset along each axis through integration, then calculates the perspective transformation matrix of the image based on the camera's intrinsic parameter matrix and the rotation angle, and finally performs an inverse transformation on the original image to compensate for the image displacement caused by rotation. The electronic image stabilization algorithm of this invention works effectively when the head movement amplitude does not exceed ±5°, and the image stability after compensation is improved by more than 80%.

[0023] Subsequently, a locally deployed lightweight OCR engine is invoked for text recognition. In one embodiment of this invention, the PaddleOCR Lite model is used. This model undergoes knowledge distillation and quantization compression, resulting in a model size of approximately 8MB. On an ARM processor, the single-frame inference time is approximately 30ms, meeting the requirements for real-time processing. The OCR engine employs a two-stage pipeline architecture: text detection and text recognition. The text detection stage uses a lightweight segmentation network to locate text regions in the image, while the text recognition stage uses a sequence-to-sequence attention network to convert the text image into a character sequence. The recognition rate can reach over 90% under normal lighting conditions, over 95% for printed text, and approximately 85% for handwritten text. The OCR recognition result includes the text content and the corresponding confidence score. The confidence score is calculated from the output probability of the recognition network, ranging from 0 to 1. This invention marks recognition results with a confidence score below 0.5 as low reliability and assigns them lower weight in subsequent fusion processing.

[0024] Hand motion tracking employs a deep learning-based keypoint detection method, which is a crucial approach for capturing learning behavior features in this invention. Preferably, the MediaPipe Hands model is used for real-time detection of 21 hand keypoints. These keypoints include four joints in the wrist, palm, and thumb, and four joints in each of the other four fingers, totaling 20 points, which can comprehensively describe hand posture and movements. The model achieves an inference frame rate of 30fps on edge computing units, meeting real-time tracking requirements. Based on keypoint sequence analysis, behavioral features such as writing speed, pause duration, number of pauses, and number of corrections are extracted. Writing speed is defined as the distance the pen tip keypoint moves per unit time, calculated as the time average of the displacement of the index finger tip keypoint between adjacent frames. A normal writing speed range is 10cm / s to 30cm / s. An excessively slow writing speed may indicate that the learner is thinking carefully or encountering difficulties, while an excessively fast writing speed may indicate that the learner is rushing or is already highly proficient. Pause detection uses a speed threshold; a pause is defined as a movement speed of less than 2 cm / s for a duration exceeding 0.5 s. Pauses are an outward manifestation of internal cognitive processing, and their frequency and duration are positively correlated with task difficulty. Correction detection is based on the overlap analysis of handwriting trajectories. Repeated overwriting marks in the same area are counted as corrections. Correction behavior reflects learners' uncertainty about their answers or their corrections after discovering errors. These behavioral characteristics are closely related to learners' cognitive states; frequent pauses and corrections usually indicate blocked thinking or difficulty in comprehension, providing important auxiliary information for mastery assessment.

[0025] To address low-light learning environments, this invention is equipped with an infrared supplementary lighting system. The supplementary lighting module uses two 940nm wavelength infrared LEDs, symmetrically distributed on both sides of the camera, with each LED having a power of no more than 0.25W. The system monitors the light intensity in real time using an ambient light sensor. When the ambient light level is detected to be below 100 lux, the infrared supplementary lighting is automatically activated, and the camera exposure compensation is adjusted by +1EV. When the ambient light level recovers to above 150 lux, the supplementary lighting is turned off, and the automatic exposure mode is restored. Experimental tests show that in a 50 lux low-light environment, the OCR recognition rate reaches 93% with supplementary lighting, while it is only 62% without supplementary lighting.

[0026] The learning content analysis unit further processes the text recognized by OCR. This invention employs a pre-trained natural language processing model for knowledge point tag extraction, preferably a lightweight distilled version of models such as BERT or RoBERTa. The model output includes the recognized knowledge point tags, a difficulty coefficient estimate, and information about the chapter to which the question belongs. The difficulty coefficient estimate comprehensively considers features such as the number of words in the question, the complexity of the formulas involved, the number of charts, and the expected number of solution steps, using a regression model for prediction, with a difficulty coefficient ranging from 0.0 to 1.0.

[0027] The signal feature fusion module 2 is responsible for intelligently fusing feature vectors from three modalities: EEG, vision, and content. Its core innovation lies in dynamically adjusting the fusion weights based on the signal quality of each modality.

[0028] The spatiotemporal alignment unit first ensures the consistency of data across modalities in the time dimension. This invention divides time windows using questions as the basic unit, with the window length adaptively adjusted according to question difficulty: 60-120 seconds for easy questions, 120-300 seconds for medium-difficulty questions, and 300-600 seconds for complex questions. Data from each sensor is synchronized using hardware timestamps, with timestamp accuracy controlled within 10ms, and periodically calibrated via the NTP network time protocol.

[0029] Signal quality assessment is the basis of adaptive weighting. For EEG signals, this invention calculates the signal-to-noise ratio (SNR) as a quality assessment metric. The SNR is calculated as the ratio of signal power to background noise power, obtained by comparing the power of the effective EEG frequency band with the power of the high-frequency noise band. The calculation formula is as follows:

[0030] ,

[0031] in, Average power in the effective EEG frequency band (0.5Hz to 45Hz), in units of ; The average power in the high-frequency noise band (50Hz to 100Hz), in units of Under normal contact conditions, the SNR is typically in the range of 15dB to 30dB. However, the SNR will decrease significantly when hair interference causes poor electrode contact.

[0032] For visual data, this invention calculates image sharpness. As a quality assessment indicator, sharpness is evaluated using the Laplace operator variance method, and the calculation formula is as follows:

[0033] ,

[0034] in, Indicates the image The result of performing a Laplacian convolution operation; This indicates the calculation of variance; The normalization constant is set to 1000, so that... The value range is normalized to the [0,1] interval. A higher sharpness value indicates a sharper image. A value below 0.5 indicates severe image blurring.

[0035] For content data, the confidence level output by the OCR recognition engine is used directly. As a quality assessment indicator, its value range is [0,1].

[0036] Based on the above quality assessment indicators, this invention uses the Sigmoid function mapping to calculate the adaptive weighting coefficients for each modality. EEG modality weighting coefficients. The calculation formula is:

[0037] ,

[0038] in, The signal-to-noise ratio threshold is set to 15dB. This is the kurtosis parameter of the Sigmoid function, set to 0.3, used to control the sensitivity of the weights to changes in signal-to-noise ratio. When the SNR is above the threshold, the weights approach 1; when it is below the threshold, the weights decrease rapidly.

[0039] Visual modal weighting coefficient The calculation formula is:

[0040] ,

[0041] in, The resolution threshold is set to 0.6; The steepness parameter is set to 5.0. In this formula... This is a sharpness assessment value for visual behavior data, ranging from 0 to 1. It is calculated by a video frame sharpness detection algorithm and reflects the sharpness of the images captured by the camera.

[0042] Content modal weighting coefficient Directly using OCR confidence level:

[0043] ,

[0044] The three modal weight coefficients need to be normalized so that their sum equals 1:

[0045] ,

[0046] in, For the normalized first Modal weighting coefficients These are the original weight coefficients before normalization. Normalization ensures the mathematical validity of the fusion operation.

[0047] This invention also incorporates a dynamic weight adjustment mechanism to address special scenarios. When the system detects that the learner is fatigued (brainwave alpha power continuously increases by more than 30%), it automatically increases the weight of the visual modality. The baseline value is 0.2 to compensate for the impact of decreased EEG signal reliability. When the system detects a low-light environment ( When the value is below 0.5, the weight of the EEG modality is automatically increased. The base value is 0.2.

[0048] Feature fusion operations employ a weighted concatenation method. Let the EEG feature vector be... The visual behavior feature vector is The feature vector of the learning content is Then the feature vectors are fused. The calculation formula is:

[0049] ,

[0050] in, This represents a vector concatenation operation; This indicates a weighted scaling of the elements of the feature vector. The dimension of the fused feature vector is... dimension.

[0051] Preferably, the present invention can also employ an attention mechanism for more refined feature fusion. Specifically, the feature vectors of the three modalities are first unified to the same dimension through a linear mapping. Then, a multi-head self-attention mechanism is used to calculate the cross-modal attention weight matrix, and finally, the fused feature vector is obtained by weighted summation. The attention mechanism can automatically learn the relative importance of features from different modalities under different cognitive states, further improving the fusion effect.

[0052] The Knowledge Graph Management Module 3 is responsible for maintaining the dynamic knowledge graph database, enabling real-time assessment of mastery and accurate prediction of forgetting curves. This module serves as a crucial bridge connecting feature fusion and path recommendation. Its core concept is to expand the traditional static knowledge graph into a dynamic learner knowledge state graph. Through continuous evaluation and updates, it accurately depicts each learner's mastery of each knowledge point and its temporal changes.

[0053] The knowledge graph database uses the graph database Neo4j for storage, but a relational database combined with a graph index can also be used. The choice of graph database is based on the inherent characteristics of knowledge graph data. There are complex dependency networks between knowledge points, and graph databases can efficiently support graph traversal queries and path planning algorithms. The data structure of the knowledge graph contains two types of elements: nodes and edges. The attributes of knowledge point nodes include: unique identifier ID, name (e.g., quadratic function - vertex form), subject, applicable grade, difficulty coefficient (0.0 to 1.0), curriculum standard weight (0 to 100), and exam frequency statistics for the past three years. The difficulty coefficient is labeled by education experts based on the cognitive complexity of the knowledge point and the average mastery rate of students. The curriculum standard weight reflects the importance of the knowledge point in the curriculum standard, and the exam frequency statistics are based on the analysis of knowledge points in the past three years' middle school or college entrance examination questions. The attributes of dependency edges include: edge type (precursor dependency or subsequent extension), dependency strength (strong dependency is set to 1.0, weak dependency is set to 0.1 to 0.9), and text description. Prerequisite edges represent the foundational knowledge required to learn a particular knowledge point, such as mastering completing the square before learning vertex expressions. Subsequent extension edges represent advanced knowledge that a knowledge point can be extended to, such as extending vertex expressions to optimization problems. Dependency strength distinguishes between necessary and beneficial prerequisites; strong dependencies indicate that understanding subsequent knowledge is impossible without prior knowledge, while weak dependencies indicate that prior knowledge is helpful but not essential. Furthermore, each knowledge point node is associated with learner-specific mastery attributes, including the current mastery score. (0 to 100) Assess confidence level (0 to 1), last update timestamp and historical mastery record sequence. The mastery attribute is learner-specific; different learners will have different mastery scores for the same knowledge point.

[0054] The mastery assessment employs a deep learning inference model, which is the core step in transforming multimodal features into a quantifiable mastery score. In a preferred embodiment of this invention, the inference model uses a lightweight Transformer architecture, with a 2-layer encoder, 4 attention heads, and 128 hidden layer dimensions. The Transformer architecture is chosen based on its superior performance in sequence modeling tasks, as the self-attention mechanism effectively captures the complex interactions between the dimensions of the feature vector. The model input is a 55-dimensional fused feature vector. After processing through two layers of Transformer encoders, a mastery score is output through a classification head. and confidence level The classification head employs a two-layer fully connected network, using ReLU activation and Dropout regularization in between. The model has approximately 2M parameters, a 98% reduction in size compared to the 110M parameters of a complete Transformer-Based model, enabling real-time inference on edge devices. After conversion to ONNX format, the inference latency on edge computing units is less than 80ms, meeting the requirements for real-time feedback. Confidence level... The confidence level reflects the reliability of the evaluation results. It is calculated based on the entropy value of the softmax probability distribution of the model's output layer. A lower entropy value indicates higher certainty in the model's predictions, resulting in higher confidence. Conversely, a higher entropy value indicates difficulty in distinguishing between multiple possible levels of mastery, leading to lower confidence. The confidence level decreases when the input data quality is poor or learner behavior is abnormal, guiding the system to adopt a more cautious approach to processing the evaluation results.

[0055] The mastery level is categorized into four levels: below 40 points is considered "no mastery" (marked in red); 40 to 70 points is considered "partial mastery" (marked in yellow); 70 to 90 points is considered "proficient mastery" (marked in green); and above 90 points is considered "expert mastery" (marked in dark green). The grading results are used for user interface visualization and differentiated recommendation strategies.

[0056] The knowledge graph is updated in real time using an exponential smoothing strategy. After each question is completed, the system updates the knowledge graph based on the user's level of mastery during the assessment. Update the mastery attribute of the corresponding knowledge point in the knowledge graph using the following formula:

[0057] ,

[0058] in, The updated mastery score is in points. The score is based on the level of mastery before the update, and the unit is points. The mastery level in this assessment is scored in points; This is a smoothing coefficient, ranging from 0.7 to 0.9. Larger values... A value that makes historical knowledge have a greater impact on update results is suitable for scenarios with low confidence levels; a smaller value... This value allows the new evaluation result to have a greater impact on the update, making it suitable for scenarios with high evaluation confidence. In one embodiment of the present invention, Value based on assessment confidence level Dynamic adjustment: .

[0059] This invention also implements a prerequisite dependency propagation mechanism. When the mastery level of a certain knowledge point is below 40 points after an update, the system automatically traces all prerequisite dependency nodes of that knowledge point and marks these nodes as needing review. This mechanism ensures that learners consolidate their prerequisite knowledge base before relearning a certain knowledge point, which conforms to the learning sequence principle of cognitive science.

[0060] Forgetting curve modeling is another core function of the knowledge graph management module. This invention models the forgetting curve based on the classic form of the Ebbinghaus forgetting curve and introduces a personalized parameter optimization mechanism. The basic forgetting curve formula is:

[0061] ,

[0062] in, for The accuracy of predicting the timing of events is scored in points. Initial time The mastery level is scored, in points; For the current moment, The time difference is the initial assessment time. The unit is h; The forgetting time constant, expressed in hours, represents the decay of mastery to its initial value. (Approximately 36.8%) of the time required.

[0063] This invention relates to the forgetting time constant Personalized optimization is performed, taking into account the impact of sleep quality and study habits. The optimized time constant is calculated using the following formula:

[0064] ,

[0065] in, The basal forgetting time constant is set to 72h, representing the typical forgetting rate without any consolidation measures. The percentage of slow-wave sleep is [0, 0.5], which can be obtained by linking with sleep monitoring equipment. Slow-wave sleep is a key stage for memory consolidation. The number of effective review sessions is a non-negative integer. The sleep impact coefficient is set to 3.0, meaning that for every 0.1% increase in the proportion of slow-wave sleep, the forgetting time constant increases. h; The review impact coefficient is set to 0.5, indicating that each effective review increases the forgetting time constant. h.

[0066] Let's illustrate the prediction and application of the forgetting curve with a specific case: Learners learn about the method of matching on Day 1, and their mastery is assessed. Points. Assuming slow-wave sleep accounts for a certain percentage. Number of times to review The optimized time constant h. If the learner does not review at all on Day 3 (after 48 hours), the predicted mastery level is [missing value]. The score has dropped from proficient to partially mastered. Based on this, the system automatically sends a review reminder on Day 2, suggesting that learners consolidate their knowledge before significant forgetting occurs.

[0067] The adaptive path recommendation module 4 is the core decision engine of this invention, responsible for calculating recommendation priorities, planning learning paths, and performing dynamic adjustments based on the knowledge graph status and the learner's real-time status.

[0068] Recommendation priority calculation comprehensively considers five dimensions of factors. For each knowledge point in the knowledge graph... Its recommendation priority The calculation formula is:

[0069] ,

[0070] in, For the first The recommended priority value for knowledge points is a dimensionless value, ranging from [0, 100]. To determine the degree gap, the calculation formula is as follows: , The current level of mastery is scored, dimensionless; The formula for scoring the importance of knowledge is as follows: ,in The weight of the curriculum standard (0 to 100). The normalized value of the frequency (from 0 to 1). The coefficient is for the grade level (1.0 for junior high school and 1.2 for senior high school). The calculation formula is based on urgency. ,in For knowledge points The set of subsequent nodes, The edge weight is the dependency strength; The forgetting risk value is calculated using the following formula: ,in This is the last study time; The fatigue penalty value is calculated using the following formula: ,in The brain fatigue index is 0 to 1. This represents the difficulty level of the current question (0 to 1). Weighting coefficient. to All are dimensionless coefficients, ensuring that the weighted terms can be directly added together.

[0071] The weighting coefficients for the five dimensions have been validated through extensive experiments and are optimally set as follows: , , , , The mastery gap and the importance of knowledge are the dominant factors. Urgency is relied upon to ensure the rationality of the learning sequence, the risk of forgetting triggers timely review, and fatigue punishment avoids excessive consumption.

[0072] The learning path planning unit generates specific learning sequences based on recommendation priorities. This invention provides three planning strategies to choose from. The greedy strategy selects the knowledge point with the highest priority at each decision point, suitable for short-term sprint scenarios. The graph traversal strategy uses depth-first or breadth-first methods to traverse the knowledge graph, ensuring complete coverage of prior knowledge, suitable for system learning scenarios. The reinforcement learning strategy uses Q-learning or Actor-Critic frameworks to learn the optimal recommendation strategy based on historical learning data, better balancing exploration and utilization.

[0073] Several constraints must be followed during path planning. The difficulty gradient constraint requires that the difficulty difference between adjacent recommended questions not exceed 0.2, avoiding cognitive shock caused by sudden changes in difficulty. The time limit constraint requires that a single study session not exceed 90 minutes, avoiding fatigue accumulation caused by prolonged study. The knowledge span constraint requires avoiding frequent switching between subjects; content from the same subject should be arranged consecutively to maintain cognitive continuity.

[0074] A dynamic adjustment mechanism is key to achieving closed-loop self-adaptation. The system assesses the learning status every 5 minutes and executes corresponding adjustment strategies based on preset trigger conditions. When the system detects that the mastery score for three consecutive questions is below 40, it determines that the learner is encountering difficulties, automatically reducing the difficulty coefficient of subsequent recommended questions by 0.1 to 0.2, and prioritizing basic reinforcement content. When the system detects that the mastery score for five consecutive questions is above 85, it determines that the current content is too easy, automatically increasing the difficulty coefficient of subsequent recommended questions by 0.1 to 0.2, and recommending extension and improvement content. When the system detects that the EEG fatigue index (calculated based on Alpha wave power changes) exceeds 0.8, it determines that the learner is fatigued, provides tactile reminders via a vibration motor, pushes rest suggestions via pop-up windows, and switches to relaxing content such as fun science videos. When the system detects that the learner has not reached the mastery standard after spending more than 30 minutes on a certain knowledge point, it triggers a human tutor intervention reminder, suggesting that the learner seek help from a teacher.

[0075] The dynamically adjusted feedback signals are also transmitted to the front-end acquisition module, forming a complete closed loop. For example, when the system determines that the learner has entered a state of fatigue, it can automatically adjust the EEG sampling strategy and increase the monitoring frequency of fatigue-related frequency bands; when the system determines that the learner has difficulty understanding the current knowledge point, it can improve the acquisition accuracy of visual modalities and analyze hand behavior characteristics in more detail to identify the specific type of difficulty.

[0076] AI Course Generation Module 5 is responsible for providing learners with personalized learning content. This module maintains a standard course content library and integrates a real-time course generation engine based on a large language model.

[0077] The standard course content library covers approximately 70% of common learning scenarios. The library is indexed by knowledge point ID and difficulty level, storing pre-built, tiered course resources, including video lectures, text-based tutorials, and interactive exercises. Each knowledge point typically comes with three difficulty levels: basic, intermediate, and advanced.

[0078] The real-time language model generation engine handles approximately 30% of personalized scenarios, including those with rare knowledge point combinations, unique error patterns, or requiring targeted explanations. When triggering conditions are met, the system constructs personalized course generation prompts based on mastery data provided by the knowledge graph management module and behavioral feature data provided by the multimodal data acquisition module. These prompts include learner profile information, such as the current knowledge point, mastery score, cognitive load index, behavioral characteristics, and specific error patterns. This invention preferably uses Claude Sonnet 4, GPT-4, or domestic large-scale models such as Wenxin Yiyan and Tongyi Qianwen as the generation engine.

[0079] AI-generated content undergoes a quality control process. The automated review stage checks for sensitive words or incorrect knowledge points. A manual review stage randomly selects 5% of the generated content for review by the teaching team. The feedback stage tracks changes in learning progress; courses with an improvement of less than 10 points in learning progress are marked as low-quality and regenerated.

[0080] Privacy protection module 6 ensures user data security by adopting a local-first processing strategy to minimize the transmission of sensitive data.

[0081] The video stream processing employs a local OCR recognition scheme. Video frames captured by the camera are processed locally on the edge computing unit. After the OCR engine extracts the text content, the original image frames are immediately discarded without being saved or uploaded. Only the OCR-recognized text content and confidence scores are uploaded to the cloud in encrypted JSON format. This scheme reduces the data volume by over 99.999%, from approximately 1.3 GB / min of the original video to approximately 100 bytes / question of OCR text.

[0082] The EEG data processing employed a feature vectorization scheme. Raw EEG waveforms were extracted and deleted locally, with only a 10-minute cache retained for anomaly backtracking. Uploaded data consisted of 30-dimensional feature vectors, reducing the data volume from approximately 180KB / item for the raw waveform to approximately 120 bytes / item for the feature vector, a reduction of 99.93%.

[0083] The system allows users to set privacy masking areas. Users can select sensitive areas in the video frame, such as the location of names and student IDs on exam papers, in the settings interface. The system saves the masking coordinates, and the area will be automatically skipped or blurred during subsequent OCR processing.

[0084] Data transmission employs AES-256 encryption and HTTPS protocol. The device generates a public-private key pair, while the cloud stores only the public key, ensuring data security during transmission. The system complies with domestic regulations such as the Personal Information Protection Law, the Data Security Law, and the Regulations on the Protection of Children's Personal Information Online, as well as international standards such as GDPR.

[0085] This invention also provides an adaptive learning path recommendation method based on the above system, comprising the following steps:

[0086] Step S1, Data Acquisition and Feature Extraction: The learner's EEG signals, visual behavior, and learning content are simultaneously acquired using a head-mounted device. The raw data is preprocessed, and feature vectors are extracted. For a detailed description of the multimodal data acquisition module 1 in the aforementioned system embodiment, please refer to the specific processing procedure.

[0087] Step S2, Adaptive Feature Fusion: Evaluate the quality index of each modal signal, calculate the adaptive weight coefficient of each modality based on the quality index, and perform weighted fusion of the feature vectors of each modality to generate a fused feature vector. For a detailed description of the signal feature fusion module 2 in the aforementioned system embodiment, please refer to the above description.

[0088] Step S3, Mastery Assessment and Knowledge Graph Update: The fused feature vector is input into the deep learning inference model to calculate the learner's mastery score and confidence level for the current knowledge point. Based on the assessment results, the mastery status attributes of the corresponding knowledge point nodes in the knowledge graph are updated, and the forgetting prediction value of each knowledge point is calculated based on the forgetting curve model. For a detailed description of the process, please refer to the knowledge graph management module 3 in the aforementioned system embodiment.

[0089] Step S4, Priority Calculation and Path Planning: Based on the mastery score, forgetting prediction value, and knowledge point dependencies, combined with the learner's real-time fatigue status index, the priority value of each knowledge point is calculated, and a learning path sequence is generated based on the priority values. For a detailed description of the process, please refer to the adaptive path recommendation module 4 in the aforementioned system embodiment.

[0090] Step S5, Dynamic Adjustment and Closed-Loop Feedback: During the learning process, the learning status is continuously monitored. When a preset adjustment condition is triggered, the corresponding dynamic adjustment strategy is executed to correct the content and difficulty parameters of the learning path in real time. The adjustment signal is then fed back to the data acquisition stage to achieve closed-loop optimization. The specific processing procedure can be found in the description of the dynamic adjustment mechanism of the adaptive path recommendation module 4 in the aforementioned system embodiment.

[0091] The steps in the above method form a continuously looping process. Steps S1 to S3 are triggered once after each question is completed. Step S4 is triggered to update when the knowledge graph state changes significantly. Step S5 continuously performs state monitoring and adjustment judgment with a 5-minute cycle.

[0092] This invention underwent a 4-week controlled experiment for verification. Thirty second-year junior high school students were recruited and randomly divided into an experimental group and a control group, with 15 students in each group. The experimental group used the adaptive learning path recommendation system of this invention, while the control group used a traditional online learning platform. The learning content was a unit on quadratic functions, including eight core knowledge points such as completing the square, vertex form, maximum / minimum problems, and practical applications.

[0093] In terms of mastery improvement, the experimental group learners improved their mastery of weak knowledge points by an average of 27.6 points, while the control group only improved by 11.2 points. The difference was statistically significant (p<0.01).

[0094] In terms of learning efficiency, the average learning time required to achieve the same level of mastery improvement was 5.2 hours for the experimental group and 7.8 hours for the control group, representing a 38% increase in efficiency.

[0095] Regarding knowledge retention, a retest was conducted one month after the learning ended. The knowledge retention rate of the experimental group was 78%, while that of the control group was 52%, an increase of 26 percentage points.

[0096] In terms of user satisfaction, the experimental group students scored 4.3 out of 5, while the control group scored 3.1 out of 5; 92% of the parents in the experimental group were satisfied with the learning outcomes.

[0097] The experimental results above demonstrate that the present invention, through the synergistic effect of multimodal physiological signal fusion, dynamic knowledge graph management, and closed-loop adaptive recommendation mechanism, can significantly improve learning effectiveness and efficiency, and has important application value.

[0098] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. An adaptive learning path recommendation system, characterized in that, It includes a multimodal data acquisition module, a signal feature fusion module, a knowledge graph management module, and an adaptive path recommendation module; The multimodal data acquisition module is used to simultaneously acquire learners' EEG signal data, visual behavior data, and learning content data through a head-mounted device, and to preprocess and extract features from the acquired raw data to generate EEG feature vectors, visual behavior feature vectors, and learning content feature vectors. The signal feature fusion module is used to receive the feature vectors of each modality output by the multimodal data acquisition module, calculate the signal quality assessment value of the EEG signal data, the clarity assessment value of the visual behavior data, and the recognition confidence assessment value of the learning content data, respectively, dynamically calculate the adaptive weight coefficient of the corresponding modality based on the signal quality assessment value, clarity assessment value, and recognition confidence assessment value, and use the adaptive weight coefficient to perform a weighted fusion operation on the feature vectors of each modality to output the fused feature vector; The knowledge graph management module is used to maintain a knowledge graph database containing knowledge point node data and dependency relationship edge data. It receives the fused feature vector output by the signal feature fusion module, calculates the learner's mastery score and evaluation confidence value for the current knowledge point through a deep learning inference model, dynamically adjusts the smoothing coefficient in the knowledge graph update strategy when the evaluation confidence value is lower than a preset confidence threshold, updates the mastery status attribute of the corresponding knowledge point node in the knowledge graph database in real time according to the mastery score, and calculates the forgetting curve prediction value of each knowledge point based on the Ebbinghaus forgetting curve model. The adaptive path recommendation module is used to calculate the recommendation priority value of each knowledge point based on the mastery score, forgetting curve prediction value and dependency edge data provided by the knowledge graph management module, combined with the fatigue state index collected in real time. The module generates a learning path sequence based on the recommendation priority value, and triggers a dynamic adjustment strategy based on continuous answer performance and cognitive load during the learning process to modify the content and difficulty parameters of the learning path sequence in real time.

2. The adaptive learning path recommendation system according to claim 1, characterized in that, The method for calculating adaptive weight coefficients by the signal feature fusion module includes: calculating EEG modal weight coefficients by mapping the difference between the signal quality assessment value of the EEG signal data and a preset signal-to-noise ratio threshold using the Sigmoid function; calculating visual modal weight coefficients by mapping the difference between the clarity assessment value of the video frame and a preset clarity threshold using the Sigmoid function; directly using the recognition confidence assessment value of the learning content data as the content modal weight coefficient; and normalizing the three modal weight coefficients so that their sum equals 1.

3. The adaptive learning path recommendation system according to claim 1, characterized in that, When the knowledge graph management module calculates the predicted value of the forgetting curve, the forgetting time constant is optimized and adjusted based on the learner's slow-wave sleep ratio data and the number of historical reviews. The basic forgetting time constant is set to 72 hours. For every 0.1% increase in the slow-wave sleep ratio, the forgetting time constant is extended by 30 hours. Each effective review extends the forgetting time constant by 36 hours.

4. The adaptive learning path recommendation system according to claim 1, characterized in that, The adaptive path recommendation module considers five dimensions when calculating the recommendation priority value: mastery gap, knowledge importance, urgency of dependence, risk of forgetting, and fatigue penalty. The weight coefficients for each dimension are set to 0.35, 0.30, 0.20, 0.10, and 0.05, respectively.

5. The adaptive learning path recommendation system according to claim 1, characterized in that, The multimodal data acquisition module extracts features from EEG signal data including: bandpass filtering from 0.5Hz to 45Hz and notch filtering from 50Hz or 60Hz on the raw EEG signal; removing eye movement and electromyography artifacts using independent component analysis; extracting power spectral density features of five frequency bands: Delta, Theta, Alpha, Beta, and Gamma; extracting amplitude features of P300 and N400 event-related potential components; and calculating the power spectral density ratios of the Theta and Alpha bands and the Beta and Alpha bands as indicators of cognitive load.

6. The adaptive learning path recommendation system according to claim 1, characterized in that, The knowledge point node data in the knowledge graph management module includes the unique identifier, name, subject, grade, difficulty coefficient, curriculum standard weight, and examination frequency in the past three years; the dependency relationship edge data includes the type of preceding dependency and the type of subsequent extension, and the edge weight distinguishes between strong dependency and weak dependency, with the weight of strong dependency set to 1.0 and the weight of weak dependency set to 0.1 to 0.

9.

7. The adaptive learning path recommendation system according to claim 1, characterized in that, The dynamic adjustment strategy of the adaptive path recommendation module includes: automatically reducing the difficulty coefficient of subsequent recommended questions when the mastery score of three consecutive questions is below 40; automatically increasing the difficulty coefficient of subsequent recommended questions when the mastery score of five consecutive questions is above 85; pushing rest suggestions and switching to lighter content when the brain fatigue index exceeds 0.8; and triggering human tutor intervention and reminders when a knowledge point is not mastered after more than 30 minutes.

8. The adaptive learning path recommendation system according to claim 1, characterized in that, It also includes an AI course generation module, which maintains a standard course content library and provides a real-time course generation engine based on a large language model. When learners make mistakes that are not covered in the standard course content library or answer the same knowledge point incorrectly three times in a row, the module constructs personalized course generation prompts based on the mastery score provided by the knowledge graph management module and the EEG feature vector, visual behavior feature vector, and learning content feature vector provided by the multimodal data acquisition module, and calls the large language model to generate targeted micro-lesson content.

9. The adaptive learning path recommendation system according to claim 1, characterized in that, It also includes a privacy protection module, which is used to upload only text data after performing local OCR recognition processing on the video stream at the edge, and to delete the original data and retain only the feature vector after extracting features from the EEG signal data. It supports users to set privacy masking areas and uses AES-256 encryption algorithm and HTTPS protocol for data transmission.

10. A head-mounted device, characterized in that, The device includes a head-mounted device body, an EEG sensor array, a visual acquisition device, and an edge computing unit, and is used in conjunction with the adaptive learning path recommendation system according to any one of claims 1 to 9. The main body of the head-mounted device features an arc-shaped design with a contact layer made of flexible silicone or memory foam on the inside. The overall weight does not exceed 120g, and it is equipped with a rear adjustment mechanism to accommodate head circumferences ranging from 50cm to 62cm. The EEG sensor array is located in the frontal lobe region of the main body of the head-mounted device, including at least three dry electrodes, which correspond to the Fp1, Fp2 and Fz positions of the international 10-20 system, respectively. The electrodes are made of gold-plated or silver / silver chloride material, with a contact impedance of less than 10kΩ and a sampling rate of not less than 250Hz. The visual acquisition device is located at the center of the front of the head-mounted device and includes a camera module and an infrared fill light module. The camera module is tilted downward at a 45° angle with an adjustable range of ±15°, has a resolution of not less than 1280×720, a frame rate of not less than 30fps, and a field of view of not less than 90°. The infrared fill light module uses a 940nm wavelength LED and automatically starts when the ambient light intensity is less than 100 lux. The edge computing unit is integrated inside the main body of the head-mounted device, including an ARM processor with a main frequency of not less than 1.2 GHz, a running memory of not less than 2 GB and a storage of not less than 16 GB, for running a lightweight inference model to achieve local real-time mastery assessment, with an inference latency of not more than 100 ms.

Images