Intelligent interactive training system and terminal based on electroencephalogram signal and attention mechanism

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent interactive training system and terminal based on an electroencephalogram signal and an attention mechanism, and the system comprises a full-process technical chain including non-invasive electroencephalogram acquisition, multi-dimensional data processing, personalized model adaptation, stepped training feedback and full-role cooperative interaction. The problems that traditional training lacks quantitative monitoring, is insufficient in individuation and is single in feedback are solved, scientific quantification, accurate intervention and long-acting migration of attention training are achieved, and meanwhile non-intrusive type and no side effect are guaranteed in the whole process; and meanwhile, the problem that personalized and continuous training services cannot be provided according to the individual characteristics of the training object due to the fact that most traditional attention training schemes are universal contents is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to an intelligent interactive training system and terminal based on electroencephalogram (EEG) signals and attention mechanisms. Background Technology

[0002] Children and adolescents (6-12 years old) are in a stage of rapid attention development, exhibiting significant age-stratification characteristics: children aged 6-8 primarily rely on involuntary attention, with short attention spans (only 10-15 minutes), and are easily distracted by external stimuli; children aged 9-12 gradually develop voluntary attention, with improved attention stability and allocation abilities, but still exhibiting significant individual differences and weak resistance to distractions. This stage of attention development directly impacts cognitive abilities, learning efficiency, and the formation of behavioral habits, making it a golden period for intervention and training.

[0003] Currently, attention training, especially for children and adolescents, faces numerous challenges. First, existing attention training processes lack scientifically quantifiable real-time monitoring methods, making it difficult to promptly detect inattentiveness and accurately intervene during training. Second, traditional attention training programs often employ generic content, failing to provide personalized and continuous training services based on individual characteristics (such as age, occupation, and basic attention level), and cannot dynamically adjust training difficulty and frequency to match the trainee's progress. Third, existing training systems employ simplistic feedback intervention methods and lack a tiered intervention logic, failing to provide efficient and humane guidance to refocus attention when trainees exhibit signs of distraction.

[0004] Therefore, the present invention provides an intelligent interactive training system and terminal based on EEG signals and attention mechanisms to solve the above problems. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides an intelligent interactive training system and terminal based on EEG signals and attention mechanisms, so as to solve the problem that the traditional attention training programs are mostly generalized and cannot provide personalized and continuous training services according to the individual characteristics of the trainees.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an intelligent interactive training system based on EEG signals and attention mechanisms, comprising: a data acquisition module for acquiring multi-dimensional monitoring data; a data processing module for preprocessing and extracting multi-dimensional features from the multi-dimensional monitoring data to obtain core feature vectors and individual baseline data; a model analysis module for analyzing the multi-dimensional monitoring data, core feature vectors, and individual baseline data using a neurophysiological habit model to obtain individual core features and individual label features; a fusion analysis of the individual core features and individual label features uploaded from various terminals using an EEG big data intelligent recommendation model to output a personalized recommendation strategy; and a tiered feedback module for executing the personalized recommendation strategy according to preset tiered training feedback rules.

[0007] Preferably, the preprocessing and multi-dimensional feature extraction of the multi-dimensional monitoring data to obtain the core feature vector and individual baseline data includes: filtering the electroencephalogram (EEG) data using an IIR notch filter to obtain initial EEG data; filtering the signals of a preset band in the initial EEG data using a 4th-order Butterworth bandpass filter to obtain second EEG data; performing artifact separation on the multi-dimensional monitoring data and second EEG data using the ICA algorithm to obtain clean EEG data; and extracting features from the clean EEG data to obtain the core feature vector and individual baseline data.

[0008] Preferably, the step of using the ICA algorithm to separate artifacts from the multi-dimensional monitoring data and the second EEG data to obtain clean EEG data includes: using the ICA algorithm to perform blind source separation on the multi-dimensional monitoring data and the second EEG data, outputting independent component data; performing artifact recognition on the independent component data based on a preset artifact feature library, outputting effective component data; reconstructing the effective component data into EEG signals based on the inverse ICA transform, obtaining initial clean EEG data; and performing baseline drift correction on the initial clean EEG data to obtain clean EEG data.

[0009] Preferably, the step of extracting features from clean EEG data to obtain core feature vectors and individual baseline data includes: calculating using Fast Fourier Transform. Wave, Wave, Wave, The average power of the wave; the key ratio is calculated based on each average power to obtain the frequency domain characteristics; the F3 / F4 locus in the prefrontal cortex is calculated. The ratio of the product of the covariance and standard deviation of wave power, the sample entropy, and the standard deviation of power in each wave band are used to obtain time-domain features; the EEG data corresponding to the training actions are superimposed, averaged, and components are extracted to obtain event-related potential features; the frequency domain features, time domain features, and event-related potential features are integrated to obtain the core feature vector; and individual baseline data are calculated based on resting-state EEG data.

[0010] Preferably, the step of using a neurophysiological habit model to analyze multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain individual core features and individual label features includes: performing feature analysis on multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain multi-dimensional input features; processing EEG features in the multi-dimensional input features using a combination of attention mechanisms and wavelet packet transform to obtain reinforced EEG features; using principal component analysis to reduce the dimensionality of the multi-dimensional input features and reinforced EEG features to obtain individual core features; using an incremental FTRL algorithm to analyze the individual core features to optimize the parameters of the neurophysiological habit model and output the optimal parameter vector, EEG baseline parameters, and association rule library; and using a multi-dimensional decision tree to classify the optimal parameter vector, EEG baseline parameters, and association rule library to obtain individual label features.

[0011] Preferably, the objective function of the neurophysiological habit model is: , where the objective function This represents the model parameter vector. Perform a minimization solution. For the number of training parameters, The loss function is Logistic Loss. For the first The 64-dimensional core features of each training sample For the first The true attention state labels of each sample The L1 regularization coefficient is... is the L2 regularization coefficient.

[0012] Preferably, the step of using an EEG big data intelligent recommendation model to fuse and analyze the individual core features and individual tag features uploaded by each terminal and output a personalized recommendation strategy includes: obtaining randomly generated initial cluster center features; concatenating the initial cluster center features with the individual core features to obtain an initial feature vector; mapping the initial feature vector to a fused feature vector using a multilayer perceptron; clustering the fused feature vector using the K-means++ algorithm to generate a group profile; performing temporal pattern analysis on the fused feature vector using a temporal LSTM model to obtain an initial personalized dynamic recommendation strategy; selecting the top-ranked strategy from the training strategies of multiple similar users based on the cosine similarity between the individual core features and the group profile as a collaborative filtering strategy; and fusing the initial personalized dynamic recommendation strategy and the collaborative filtering strategy to obtain a personalized recommendation strategy.

[0013] Preferably, the step of using the K-means++ algorithm to cluster the fused feature vectors to generate a group profile includes: initializing cluster centers; calculating the distance between each feature vector using cosine similarity; assigning the fused feature vector of each individual to the nearest cluster center; calculating the feature mean of each class of samples as the new cluster center; stopping the iteration when the cluster center offset is less than a preset offset value or the number of iterations reaches a preset number, extracting features from the clustering results, and generating a group profile.

[0014] Preferably, the step of executing a personalized recommendation strategy according to a preset tiered training feedback rule includes: executing training modes according to a preset training order and obtaining feedback data for each mode; and matching a training strategy suitable for the current mode to the feedback data based on the personalized recommendation strategy.

[0015] Preferably, the terminal includes: an EEG acquisition terminal for acquiring EEG physiological data; a training interaction terminal for training and acquiring interaction data, and completing feature extraction, label generation, and personalized recommendation strategy generation processes based on an integrated neurophysiological habit model and an EEG big data intelligent recommendation model; and a multimodal feedback terminal for generating and executing intervention actions based on the personalized recommendation strategy.

[0016] The beneficial effects of this invention are as follows: 1. This invention solves the problems of traditional training, such as lack of quantitative monitoring, insufficient personalization, and single feedback, through a complete technical chain of non-invasive EEG acquisition, multi-dimensional data processing, personalized model adaptation, tiered training feedback, and full-role collaborative interaction. It achieves scientific quantification, precise intervention, and long-term transfer of attention training, while ensuring non-invasiveness and no side effects throughout the process. It also solves the problem that traditional attention training programs are mostly generalized and cannot provide personalized and continuous training services based on the individual characteristics of the trainees.

[0017] 2. The individual label features and core features obtained from the neurophysiological habit model analysis can comprehensively cover core features such as EEG baseline, feedback preference, and physiological tolerance, achieving a precise description of each individual's unique features; and the built-in incremental FTRL algorithm supports incremental parameter updates after each training session, which can adapt to the pace of children's attention improvement.

[0018] 3. This invention sets up a progressive approach from preparation, basic, core and consolidation, which is adapted to the cognitive process of children’s attention awakening, strengthening and consolidation. In addition, the training time and difficulty are dynamically adjusted according to the child’s attention state, which can avoid excessive fatigue or inefficient training. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of an intelligent interactive training system based on electroencephalogram (EEG) signals and attention mechanisms according to the present invention. Detailed Implementation

[0020] The following will refer to the attached reference. Figure 1 The various embodiments of the present invention will be described in detail below. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0021] An intelligent interactive training system based on EEG signals and attention mechanisms, as shown in the attached figure. Figure 1 As shown, it includes: a data acquisition module, used to collect multi-dimensional monitoring data; the multi-dimensional monitoring data includes electroencephalographic data, behavioral interaction data, peripheral feedback data, and environmental status data, providing support for multi-dimensional analysis.

[0022] Regarding the selection of sampling sites, based on the developmental characteristics of the prefrontal cortex in adolescents, the F3 / F4 core sites (key brain regions for attention regulation) were selected to avoid invasive intracranial sampling, and silver / silver chloride disc electrodes were used. To reduce skin irritation, the sampling rate is set to 256Hz (meeting the Nyquist sampling requirements of the 4-40Hz target band), with a timestamp accuracy of 1ms to ensure no signal aliasing and accurate synchronization across devices; Bluetooth 5.0 BLE Low Energy transmission is used to reduce radiation and device power consumption.

[0023] The data processing module preprocesses and extracts multi-dimensional features from the multi-dimensional monitoring data to obtain core feature vectors and individual baseline data. Addressing the susceptibility of children's EEG signals to eye movement and electromyography (EMG) interference, a comprehensive preprocessing process is employed, including a 50Hz notch filter to remove power frequency interference, a Butterworth bandpass filter to retain the 4-40Hz target band, ICA artifact removal to separate eye movement / EMG interference, and baseline calibration, ensuring signal purity. Focusing on attention-related indicators in adolescents and children, the module extracts... Core features such as ratios, cross-brain region synchronization coefficients, and P3 component amplitudes are used to establish an individual EEG baseline, avoiding errors caused by age and individual differences.

[0024] Set valid data judgment criteria, such as signal-to-noise ratio (SNR) ≥ 10dB and artifact rate < 5%. If the data is below the criteria, automatically prompt to adjust the electrode wearing position to ensure data availability.

[0025] The model analysis module uses a neurophysiological habit model to analyze multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain individual core features and individual label features. The EEG big data intelligent recommendation model is used to fuse and analyze the individual core features and individual label features uploaded from various terminals to output personalized recommendation strategies.

[0026] The tiered feedback module executes personalized recommendation strategies based on preset tiered training feedback rules.

[0027] This invention addresses the shortcomings of traditional training methods, such as lack of quantitative monitoring, insufficient personalization, and limited feedback, through a complete technological chain encompassing non-invasive EEG acquisition, multi-dimensional data processing, personalized model adaptation, tiered training feedback, and multi-role collaborative interaction. It achieves scientific quantification, precise intervention, and long-term transfer of attention training while ensuring non-invasiveness and no side effects throughout the process. Furthermore, it solves the problem that traditional attention training programs often offer generic content, failing to provide personalized and continuous training services based on the individual characteristics of the trainees.

[0028] In one embodiment of the present invention, the preprocessing and multi-dimensional feature extraction of multi-dimensional monitoring data to obtain core feature vectors and individual baseline data includes: filtering the electroencephalogram (EEG) data using an IIR notch filter to obtain initial EEG data; filtering signals in a preset band from the initial EEG data using a 4th-order Butterworth bandpass filter to obtain second EEG data; performing artifact separation on the multi-dimensional monitoring data and second EEG data using the ICA algorithm to obtain clean EEG data; and extracting features from the clean EEG data to obtain core feature vectors and individual baseline data.

[0029] Specifically, an IIR notch filter is used to filter the EEG physiological data. The parameters are specifically adapted to children's EEG signals, with a center frequency of 50Hz and a Q value of 30. This ensures that only the target frequency is filtered without affecting adjacent effective bands, removing 50Hz power frequency interference and harmonics generated by the power grid, and avoiding interference. , The core waveband signals are used to output initial EEG data.

[0030] A fourth-order Butterworth bandpass filter was used to filter signals in preset bands from the initial EEG data. Specifically, the 4-40Hz band, which is directly related to attention, was retained, while low-frequency drift (<4Hz) and high-frequency noise (>40Hz) were filtered out. This reduced the interference of irrelevant band noise on subsequent feature extraction, improved feature recognition, and output the second EEG data.

[0031] The ICA algorithm is used to separate artifacts from multi-dimensional monitoring data and second EEG data. Common eye movement artifacts (blinking, eye movement), electromyography artifacts (facial muscle contraction), and electrocardiogram artifacts in children's training are separated and removed. The pure EEG signal is reconstructed and baseline drift correction is performed to output pure EEG data.

[0032] Feature extraction was performed on pure EEG data to extract three core features that are strongly correlated with the attention of adolescents and children, thus obtaining core feature vectors while avoiding redundant information.

[0033] In this way, the present invention can provide a high-quality data foundation for subsequent analysis steps, which helps to improve the accuracy of identification.

[0034] In one embodiment of the present invention, the method of using the ICA algorithm to separate artifacts from multi-dimensional monitoring data and second EEG data to obtain pure EEG data includes: using the ICA algorithm to perform blind source separation on multi-dimensional monitoring data and second EEG data, and outputting independent component data; performing artifact recognition on the independent component data based on a preset artifact feature library, and outputting effective component data; reconstructing the effective component data into EEG signals based on inverse ICA transform to obtain initial pure EEG data; and performing baseline drift correction on the initial pure EEG data to obtain pure EEG data.

[0035] Specifically, the ICA algorithm is used to perform blind source separation on multi-dimensional monitoring data and second EEG data. When setting the ICA algorithm, the number of independent components = number of acquisition channels + 2; the negative entropy maximization is used as the objective function; the upper limit of the number of iterations is 100; after the ICA algorithm analyzes and processes the multi-dimensional monitoring data and second EEG data, the independent component data is output.

[0036] Based on a pre-set forgery feature library, forgery identification is performed on independent component data. A three-dimensional recognition system is constructed based on time-domain features, frequency-domain features, and waveform features formed from a large amount of historical data. Combined with the typical characteristics of common forgeries in children, a forgery feature library is constructed to automatically identify and label forgery components.

[0037] Automatic recognition mainly includes: eye-tracking artifact recognition, meeting the requirements of " Eye movement artifacts are identified by any two of the following three conditions: "low-frequency energy percentage > 80%" and "peak waveform". Considering the high blinking frequency of children (approximately 15-20 blinks / minute for 6-12 year olds), the maximum allowable percentage of artifact components is set at 30% to avoid excessive rejection of effective signals. For electromyography (EMG) artifact identification, "high-frequency energy percentage > 70%", "irregular waveform", and "sudden amplitude increase" are considered preliminary conditions. Combining these with "energy > 3 times the average energy of effective beta waves" confirms it as an EMG artifact. Considering the incomplete development of facial muscles in children and the tendency for involuntary contractions, the EMG artifact identification threshold is relaxed. (That is, to include in the judgment) to avoid omission.

[0038] ECG artifact detection identifies artifacts with "periodic waveforms and periods within the 0.6-0.86s range," confirmed by synchronization with heart rate monitoring data (if applicable). Artifacts are characterized by a low percentage (typically <5%) and are directly removed after identification without further retention. Automatically identified artifact components are sampled for verification (1 artifact is sampled from every 10 data segments). If the artifact detection accuracy is <90%, the detection threshold is adjusted (e.g., the eye-tracking artifact amplitude threshold is adjusted from...). Down to In addition, rules for determining the effective ingredients are set (“ If the identified artifacts are characterized by an energy distribution between 4-40Hz and no obvious periodicity / peaks, they are automatically marked as "suspected effective components" and will not be included in subsequent removal. After completing the above process, the effective component data is output.

[0039] Based on the inverse ICA transform, the effective component data is reconstructed into EEG signals. First, all labeled artifact components are removed from the independent component set, retaining the effective components. The retained effective components are then subjected to the inverse ICA transform to restore the EEG signals with the same number of acquisition channels as the original. The transformation process strictly follows the reversibility principle of the ICA algorithm to ensure that the signal amplitude and phase relationship are not distorted. The reconstructed signal is then smoothed by a 5-point moving average to eliminate high-frequency noise that may be introduced by the inverse transform. The smoothing window is ≤5 points to avoid blurring of the effective signal, and the initial clean EEG data is output.

[0040] Baseline drift correction is performed on the initial pure EEG data. This involves dynamic baseline correction of the initial pure EEG data based on children's resting state data. The difference between the current signal and the sliding window baseline is calculated in real time to eliminate baseline drift caused by unstable electrode contact and slow physiological changes. This ensures that the signal amplitude reflects the real EEG activity and outputs pure EEG data.

[0041] The pure EEG signal output in this embodiment is free from major artifacts such as eye movement and electromyography, and baseline drift correction avoids baseline deviation caused by artifacts. At the same time, the pure signal ensures the accuracy of attention-related features such as the θ / β ratio and P3 component extracted subsequently, reduces the interference of artifacts on feature quantification, and helps to improve the accuracy of personalized strategies.

[0042] In one embodiment of the present invention, the step of extracting features from clean EEG data to obtain core feature vectors and individual baseline data includes: calculating using Fast Fourier Transform. Wave, Wave, Wave, The average power of the wave; the key ratio is calculated based on each average power to obtain the frequency domain characteristics; the F3 / F4 locus in the prefrontal cortex is calculated. The ratio of the product of the covariance and standard deviation of wave power, the sample entropy, and the standard deviation of power in each wave band are used to obtain time-domain features; the EEG data corresponding to the training actions are superimposed, averaged, and components are extracted to obtain event-related potential features; the frequency domain features, time domain features, and event-related potential features are integrated to obtain the core feature vector; and individual baseline data are calculated based on resting-state EEG data.

[0043] In this embodiment, feature extraction is performed on the clean EEG data. When extracting frequency domain features, a Fast Fourier Transform (FFT) is used with a window length of 256 points and an overlap rate of 50%. Wave (4-8Hz), Wave (8-13Hz), Wave (12-15Hz), The average power of the wave (13-30Hz); based on these average powers, key ratios are calculated, for example, Ratio (TBR) The ratio reflects the emotional stability and attentional state of adolescents and children.

[0044] When extracting temporal features, the ratio of the product of the covariance and standard deviation of the β wave power at the F3 / F4 sites in the prefrontal cortex is calculated to obtain the cross-brain region synchronization coefficient. A cross-brain region synchronization coefficient ≥0.5 indicates high synchronization, reflecting the coordination ability of attention regulation. The lower the sample entropy (SE) value, the more stable the EEG signal and the more focused the attention. The standard deviation of the power of each band is also calculated to reflect the fluctuation of band activity. The smaller the standard deviation, the more stable the attention.

[0045] When extracting event-related potential features, the EEG signals corresponding to the "target stimulus" (such as correct clicking of the Schulte grid, auditory training target number) in training were averaged 30 times to cancel out random noise; the amplitude and latency of the P3 component (250-500ms) were extracted, and the P3 amplitude was ≥ This indicates accurate attention and response; the shorter the latency period, the faster the reaction speed.

[0046] Furthermore, before training, children were guided to remain calm and relaxed, and 5 minutes of resting-state EEG data were collected. After preprocessing, this data served as the individual baseline. Core indicators such as TBR, power of each band, and synchronization coefficient were calculated during the baseline period as benchmarks for subsequent training evaluation, resulting in individual baseline data. Among these, TBR was... The ratio is used to compare the real-time features extracted during training with the individual baseline, calculate the relative rate of change, and output standardized features. This avoids using a uniform standard to evaluate different children (e.g., the baseline TBR of 6-year-old and 12-year-old children is significantly different) and ensures the individual relevance of the features.

[0047] The core feature vector is the core input to the model analysis module. Individual baseline data provides a benchmark for individual model construction, while real-time feature vectors provide a basis for the model to dynamically evaluate attention status and generate personalized training strategies. The standardization of features ensures that the model can be accurately adapted across ages and individuals, avoiding model bias caused by differences in feature scale.

[0048] In one embodiment of the present invention, the method of using a neurophysiological habit model to analyze multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain individual core features and individual label features includes: performing feature analysis on multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain multi-dimensional input features; processing EEG features in the multi-dimensional input features using a combination of attention mechanism and wavelet packet transform to obtain enhanced EEG features; using principal component analysis to reduce the dimensionality of the multi-dimensional input features and enhanced EEG features to obtain individual core features; using the incremental FTRL algorithm to analyze the individual core features to optimize the parameters of the neurophysiological habit model and output the optimal parameter vector, EEG baseline parameters, and association rule library; and using a multi-dimensional decision tree to classify the optimal parameter vector, EEG baseline parameters, and association rule library to obtain individual label features.

[0049] Specifically, feature analysis is performed on multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain core EEG features, behavioral interaction features, feedback interaction features, environmental and state features, and derived features, forming multi-dimensional input features based on a neurophysiological habit model. Among these, core EEG features include time-domain features, frequency-domain features, and event-related potential features; behavioral interaction features include the mean / standard deviation of training reaction time, accuracy rate, frequency of attention wandering, and module selection preferences; feedback interaction features include the number of triggers, effective times, tolerance intensity levels, and selection preference percentages for each peripheral device; and environmental and state features include training period efficiency (achievement rate at different times), environmental noise impact coefficient, frequency of fatigue, and emotional stability. Derivative features include the correlation between EEG and behavior (such as the correlation coefficient between TBR and reaction time), the lag time between feedback and EEG improvement, and the rate of decay of training effect.

[0050] In terms of feature enhancement, a combination of "attention mechanism and wavelet packet transform (WPT)" is adopted to assign an attention weight of ≥0.8 to core bands such as θ and β, thereby strengthening key features and obtaining enhanced EEG features. Then, principal component analysis (PCA) is used to retain 64-dimensional core features with 95% variance, eliminate redundant information, and reduce the computational complexity of the neurophysiological habit model.

[0051] Meanwhile, the incremental FTRL algorithm is adopted to adapt to the dynamic updating characteristics of children's data, so as to realize the construction and iteration of baseline calibration and neurophysiological habit model.

[0052] Regarding the parameter configuration of the neurophysiological habituation model, the regularization parameter is: L1 regularization parameters, sparsified feature weights, The L2 regularization parameter controls model complexity; the forgetting coefficient is 0.05, which slowly decays the weights of features that haven't appeared in a long time, avoiding interference from outdated data. The objective function of the neurophysiological habituation model is: , Wherein, objective function This represents the model parameter vector. Perform a minimization solution. For the number of training parameters, The loss function is Logistic Loss. For the first The 64-dimensional core features of each training sample For the first The true attention state labels of each sample The L1 regularization coefficient is... is the L2 regularization coefficient. The objective function aims to smooth the model parameters (weight decay), preventing excessive weights for individual features (such as abnormal EEG features in a particular training session) that could lead to overfitting of the model to the local training data of children; it also allows the model to focus more on the overall feature patterns, improving its generalization ability to new training data and adapting to the characteristics of fluctuating attention states and the presence of some noise in the data.

[0053] The individual label features and core features obtained from the neurophysiological habit model analysis can comprehensively cover core features such as EEG baseline, feedback preference, and physiological tolerance, achieving a precise description of each individual's unique characteristics; and the built-in incremental FTRL algorithm supports incremental parameter updates after each training session, which can adapt to the pace of children's attention improvement.

[0054] In one embodiment of the present invention, the method of using an EEG big data intelligent recommendation model to fuse and analyze individual core features and individual tag features uploaded by each terminal and output a personalized recommendation strategy includes: obtaining randomly generated initial cluster center features; concatenating the initial cluster center features with individual core features to obtain an initial feature vector; mapping the initial feature vector to a fused feature vector using a multilayer perceptron; clustering the fused feature vector using the K-means++ algorithm to generate a group profile; performing temporal pattern analysis on the fused feature vector using a temporal LSTM model to obtain an initial personalized dynamic recommendation strategy; selecting the top-ranked strategy from the training strategies of multiple similar users based on the cosine similarity between individual core features and the group profile as a collaborative filtering strategy; and fusing the initial personalized dynamic recommendation strategy and the collaborative filtering strategy to obtain a personalized recommendation strategy.

[0055] Specifically, the process involves obtaining 8 randomly generated 16-dimensional initial cluster center features; concatenating these initial cluster center features with individual core features to obtain an initial feature vector; using a multilayer perceptron to map the initial feature vector into a fused feature vector; in the multilayer perceptron, the first hidden layer uses the ReLU activation function for fusion and dimensionality reduction, and normalizes it using BatchNorm to process the 80-dimensional initial feature vector into a 128-dimensional feature vector; the second hidden layer uses the ReLU activation function with Dropout=0.2 to process the 128-dimensional feature vector into a 64-dimensional feature vector; and then, in the output layer, the 64-dimensional feature vector is processed into 128 dimensions and normalized using BatchNorm to output the fused feature vector.

[0056] The K-means++ algorithm is used to cluster the fused feature vectors to generate a group profile. A temporal LSTM model is then used to analyze the temporal patterns of the fused feature vectors, capturing the temporal patterns in the children's training process (such as the recent pace of attention improvement and difficulty adaptation) to obtain an initial personalized dynamic recommendation strategy. The temporal LSTM model's network structure includes an input layer, hidden layer 1, hidden layer 2, and an output layer. Hidden layer 1 is a 64-dimensional LSTM unit with dropout=0.2 (to prevent overfitting) to capture long-term temporal dependencies. Hidden layer 2 is a 32-dimensional LSTM unit that further extracts key temporal features. The output layer outputs a 16-dimensional recommendation vector, including the proportion of training modules (4-dimensional), initial difficulty (2-dimensional), peripheral combination priority (3-dimensional), training duration suggestion (2-dimensional), dynamic adjustment threshold (3-dimensional), and training period recommendation (2-dimensional).

[0057] Regarding the training parameter settings, the optimizer is the Adam optimizer, with a learning rate of 0.001 and a decay rate of 0.99; the loss function is the mean squared error (MSE), which measures the deviation between the recommended vector and the true optimal policy vector; and the number of training iterations is 50 rounds.

[0058] Based on the cosine similarity between individual core features and group profiles, the top-ranking strategies are selected from the training strategies of multiple similar users as collaborative filtering strategies; the initial personalized dynamic recommendation strategy and the collaborative filtering strategy are fused to obtain a personalized recommendation strategy.

[0059] Based on the cosine similarity between individual fusion features and 8 types of profiles, the top ten similar users (the 10 individuals with the highest similarity within the same profile) are selected; the optimal solution with "achievement rate ≥ 80% and feedback effectiveness ≥ 90%" is extracted from the training strategies of similar users as candidate recommendation strategies; the strategies are sorted according to their matching degree with the current individual's label, and the top three are retained as candidate strategies.

[0060] The initial personalized dynamic recommendation strategy output by the temporal LSTM model has a weight of 0.7, focusing on individual dynamic trends, while the collaborative filtering strategy output by the collaborative filtering model has a weight of 0.3, focusing on group optimal experience. The corresponding dimensions of the 16-dimensional recommendation vector are weighted, summed, and integrated to obtain the personalized recommendation strategy.

[0061] Furthermore, the method of using the K-means++ algorithm to cluster the fused feature vectors and generate a group profile includes: initializing cluster centers; calculating the distance between each feature vector using cosine similarity; assigning the fused feature vector of each individual to the nearest cluster center; calculating the feature mean of each class of samples as the new cluster center; stopping the iteration when the cluster center offset is less than a preset offset value or the number of iterations reaches a preset number, extracting features from the clustering results, and generating a group profile.

[0062] In the initialization of cluster centers, after multiple experiments, the K value was set to 8; cosine similarity was used to calculate the distance between feature vectors to adapt to high-dimensional feature distance calculation and reduce the impact of the curse of dimensionality; the preset offset value was 0.001, and the preset number of iterations was 100; features were extracted from the 8 clustering results, and the core features of each category (EEG, behavior, feedback preference, age fit) were summarized to form 8 profiles of the morning attention groups of teenagers and children. For example, profile ID 1 has the core features of high theta wave, effective sound feedback, and high efficiency in the morning, and the corresponding training strategy is Schulte training (40%), broadcasting doll, and 7:30-8:00 training; profile ID 2 has the core features of insufficient SMR wave, sensitive vibration feedback, and short-term high frequency, and the corresponding training strategy is control-type games, low-intensity vibration (level 1-2), 15 minutes / time × 2 times / day.

[0063] In one embodiment of the present invention, the step of executing a personalized recommendation strategy according to a preset tiered training feedback rule includes: executing training modes according to a preset training order and obtaining feedback data for each mode; and matching a training strategy suitable for the current mode to the feedback data based on the personalized recommendation strategy.

[0064] Specifically, the main process of executing the training mode according to the preset training order includes four modes, which are arranged in the preset training order as preparation mode, basic mode, core mode and consolidation mode; The preparation mode, designed for children who are easily distracted and prone to mood swings during the initial stages of training, activates their attention through relaxation techniques. Waves (8-13Hz) lower cortisol levels, stabilize children's emotions, calibrate the EEG baseline, and lay the foundation for focused state in subsequent training; select corresponding audio according to age (6-8 years old: cartoon character-guided meditation, paired with natural white noise; 9-12 years old: simple language guidance, incorporating mild mindfulness training); The audio is 3-5 minutes long with a soothing rhythm. The interface displays dynamic breathing waveforms to guide children's synchronized breathing; real-time monitoring is also available. Wave power and Ratio, when Wave power increased by ≥20% compared to baseline and When the value is ≥1.8, it is considered ready and automatically enters the basic mode; if the value is not met, the process will be extended by 1-2 minutes or the guidance will be repeated.

[0065] The basic mode, diaphragmatic breathing training, is suitable for children to establish basic attention control skills, while improving cross-brain region synchronization between the prefrontal and parietal lobes, strengthening autonomic nervous system regulation, and establishing a linkage mechanism between breathing, attention, and EEG, activating β waves (13-30Hz) and SMR waves (12-15Hz). The interface displays a balloon inflation / deflation animation, accompanied by voice prompts "Inhale for 4 seconds, hold your breath for 2 seconds, exhale for 6 seconds," and a synchronous vibration feedback device triggers level 1 vibration according to the breathing rhythm; the initial breathing rate is 6 breaths / minute, which is subsequently adjusted to 8 breaths / minute based on the child's adaptation (breathing rhythm stability ≥65%); monitoring β wave power and cross-brain region synchronization coefficient, when β wave power increases by ≥18% from baseline and the synchronization coefficient is ≥0.55, the core mode is entered.

[0066] The core model, combined with children's gamification preferences, integrates attention training into fun scenarios to improve training compliance. Simultaneously, it enhances neuroplasticity through high-intensity training, strengthening core abilities such as sustained attention, visual search, and working memory. It also targets and regulates the theta / β ratio. For Schulte training, it starts with a 3×3 grid, upgrading to 4×4 with a success rate ≥85%, and reaching a maximum of 6×6. For EEG games, control-based games (such as airplane battles) initially have a slower speed; speed and obstacle density are increased when SMR wave power ≥20%. TBR and attentional wandering frequency are assessed every 30 seconds, triggering corresponding feedback interventions.

[0067] The consolidation mode plays visual clips of attentional states during training (e.g., green markings for "focused periods" and yellow markings for "distracted periods"), accompanied by concise voice feedback ("You focused for 8 minutes during the Schulte training, great job!"). It also plays soothing music and uses an aromatherapy device (lavender scent, concentration 0.2mg / m³) to guide children in deep breathing and reduce gamma wave power (<8μV² is considered fatigue relief). A training summary for each stage (achievement rate, focus duration) is generated and simultaneously pushed to parents for future reference. This visual feedback allows children to intuitively perceive training results while simultaneously promoting relaxation and preventing over-fatigue, thus increasing their motivation for long-term training.

[0068] Regarding the hierarchical multimodal feedback mechanism, the feedback priority is: core feedback (tactile, auditory) > auxiliary feedback (visual) > extended feedback (olfactory), ensuring that key feedback takes effect quickly, auxiliary feedback enhances the effect, and extended feedback optimizes the experience. Peripheral combinations include, but are not limited to, a talking doll (auditory), a vibration feedback device (tactile), a visual terminal (visual), and an aromatherapy device (olfactory, optional). All peripherals adopt a child-friendly design (rounded corners, soft materials). Feedback latency control: end-to-end latency ≤150ms (EEG state determination → peripheral feedback triggering), timing deviation ≤30ms, avoiding cognitive interference caused by asynchronous multimodal feedback.

[0069] Attention states include focused, mildly inattentive, moderately inattentive, severely inattentive, and fatigued; each state has corresponding judgment criteria and feedback combination strategies, which are used to intervene in training and improve the attention training effect of adolescents and children.

[0070] This invention sets up a progressive approach from preparation, basic, core, and consolidation, which is adapted to the cognitive process of children's attention awakening, strengthening, and consolidation. Furthermore, the training duration and difficulty are dynamically adjusted according to the child's attention state, which can avoid excessive fatigue or inefficient training.

[0071] In one embodiment of the present invention, the terminal used in the intelligent interactive training system based on EEG signals and attention mechanisms includes: an EEG acquisition terminal for acquiring EEG physiological data.

[0072] The training interactive terminal is used to train and acquire interactive data, and completes the process of feature extraction, label generation, and personalized recommendation strategy generation based on the integrated neurophysiological habit model and EEG big data intelligent recommendation model. The multimodal feedback terminal generates and executes intervention actions based on the personalized recommendation strategy.

[0073] Specifically, the EEG acquisition terminal is a wearable EEG acquisition device for the prefrontal cortex (headband / hairband form), a non-invasive core acquisition device that integrates silver / silver chloride disc electrodes, a Bluetooth transmission module, and a low-power processing chip for acquiring EEG physiological data.

[0074] The training interactive terminal serves as the core for training execution and data processing. Following a four-stage process of "preparation, foundation, core, and consolidation," it pushes content such as music meditation, diaphragmatic breathing, Schulte training, and EEG games, supporting dynamic combination of training modules and adaptive adjustment of difficulty. It also integrates a data preprocessing algorithm chain (notch filtering, ICA artifact separation, etc.) and intelligent models, completing feature extraction, label generation, and training strategy inference locally. For child-friendly interaction, it adopts a large icon and high color contrast design, supporting voice-guided operation.

[0075] The multimodal feedback terminal generates and executes intervention actions based on personalized recommendation strategies. Intervention actions include, but are not limited to, voice guidance, tiered feedback, environmental adaptation, haptic feedback, security control, reward animations, and emotion regulation.

[0076] The data management terminal is used for storage, training monitoring, and effect evaluation. For training monitoring, it allows real-time viewing of key indicators such as children's training progress, attention span, and achievement rate.

[0077] The cloud server stores the global training data pool (after anonymization), global model parameters, and training resource packages in a distributed manner.

[0078] The individual label features and core features obtained from the neurophysiological habit model analysis can comprehensively cover core features such as EEG baseline, feedback preference, and physiological tolerance, achieving a precise description of each individual's unique characteristics; and the built-in incremental FTRL algorithm supports incremental parameter updates after each training session, which can adapt to the pace of children's attention improvement.

[0079] This invention sets up a progressive approach from preparation, basic, core, and consolidation, which is adapted to the cognitive process of children's attention awakening, strengthening, and consolidation. Furthermore, the training duration and difficulty are dynamically adjusted according to the child's attention state, which can avoid excessive fatigue or inefficient training.

[0080] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0081] It should be noted that in the description of this invention, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0082] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0083] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0084] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0085] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0086] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0087] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An intelligent interactive training system based on electroencephalogram (EEG) signals and attention mechanisms, characterized in that, include: The data acquisition module is used to collect multi-dimensional monitoring data; The data processing module preprocesses and extracts multi-dimensional features from the multi-dimensional monitoring data to obtain core feature vectors and individual baseline data. The model analysis module uses a neurophysiological habit model to analyze multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain individual core features and individual label features; it uses an EEG big data intelligent recommendation model to fuse and analyze the individual core features and individual label features uploaded from various terminals and output personalized recommendation strategies. The tiered feedback module executes personalized recommendation strategies based on preset tiered training feedback rules.

2. The intelligent interactive training system according to claim 1, characterized in that, The aforementioned preprocessing and multi-dimensional feature extraction of multi-dimensional monitoring data to obtain core feature vectors and individual baseline data includes: filtering the electroencephalogram (EEG) data using an IIR notch filter to obtain initial EEG data; filtering signals in a preset band from the initial EEG data using a 4th-order Butterworth bandpass filter to obtain second EEG data; separating artifacts from the multi-dimensional monitoring data and second EEG data using the ICA algorithm to obtain clean EEG data; and extracting features from the clean EEG data to obtain core feature vectors and individual baseline data.

3. The intelligent interactive training system according to claim 2, characterized in that, The method of using the ICA algorithm to separate artifacts from multi-dimensional monitoring data and second EEG data to obtain pure EEG data includes: using the ICA algorithm to perform blind source separation on multi-dimensional monitoring data and second EEG data, outputting independent component data; performing artifact recognition on the independent component data based on a preset artifact feature library, outputting effective component data; reconstructing the effective component data into EEG signals based on inverse ICA transform, obtaining initial pure EEG data; and performing baseline drift correction on the initial pure EEG data to obtain pure EEG data.

4. The intelligent interactive training system according to claim 2, characterized in that, The aforementioned feature extraction from clean EEG data to obtain core feature vectors and individual baseline data includes: calculating the average power of theta waves, alpha waves, SMR waves, and beta waves using Fast Fourier Transform; calculating key ratios based on each average power to obtain frequency domain features; calculating the ratio of the product of covariance and standard deviation of beta wave power at the F3 / F4 sites in the prefrontal cortex, sample entropy, and standard deviation of power in each band to obtain time domain features; performing superposition averaging and component extraction on EEG data corresponding to training actions to obtain event-related potential features; integrating frequency domain features, time domain features, and event-related potential features to obtain the core feature vector; and calculating individual baseline data based on resting-state EEG data.

5. The intelligent interactive training system according to claim 1, characterized in that, The method described above uses a neurophysiological habit model to analyze multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain individual core features and individual label features. This includes: performing feature analysis on multi-dimensional monitoring data, core feature vectors, and individual baseline data to obtain multi-dimensional input features; processing EEG features from the multi-dimensional input features using a combination of attention mechanisms and wavelet packet transform to obtain reinforced EEG features; using principal component analysis to reduce the dimensionality of the multi-dimensional input features and reinforced EEG features to obtain individual core features; using an incremental FTRL algorithm to analyze the individual core features to optimize the parameters of the neurophysiological habit model, outputting the optimal parameter vector, EEG baseline parameters, and association rule database; and using a multi-dimensional decision tree to classify the optimal parameter vector, EEG baseline parameters, and association rule database to obtain individual label features.

6. The intelligent interactive training system according to claim 5, characterized in that, The objective function of the neurophysiological habit model is: , Wherein, objective function This represents the model parameter vector. Perform a minimization solution. For the number of training parameters, The loss function is Logistic Loss. For the first The 64-dimensional core features of each training sample For the first The true attention state labels of each sample The L1 regularization coefficient is... is the L2 regularization coefficient.

7. The intelligent interactive training system according to claim 1, characterized in that, The method described above employs an EEG big data intelligent recommendation model to fuse and analyze individual core features and individual tag features uploaded from various terminals, and outputs a personalized recommendation strategy. This includes: obtaining randomly generated initial cluster center features; concatenating the initial cluster center features with individual core features to obtain an initial feature vector; mapping the initial feature vector to a fused feature vector using a multilayer perceptron; clustering the fused feature vector using the K-means++ algorithm to generate a group profile; performing temporal pattern analysis on the fused feature vector using a temporal LSTM model to obtain an initial personalized dynamic recommendation strategy; selecting the top-ranked strategy from the training strategies of multiple similar users based on the cosine similarity between individual core features and the group profile as a collaborative filtering strategy; and fusing the initial personalized dynamic recommendation strategy and the collaborative filtering strategy to obtain a personalized recommendation strategy.

8. The intelligent interactive training system according to claim 7, characterized in that, The method of using the K-means++ algorithm to cluster fused feature vectors and generate a group profile includes: initializing cluster centers; calculating the distance between each feature vector using cosine similarity; assigning the fused feature vector of each individual to the nearest cluster center; calculating the feature mean of each class of samples as the new cluster center; stopping the iteration when the cluster center offset is less than a preset offset value or the number of iterations reaches a preset number, extracting features from the clustering results, and generating a group profile.

9. The intelligent interactive training system according to claim 1, characterized in that, The method of executing a personalized recommendation strategy based on a preset tiered training feedback rule includes: executing training modes according to a preset training order and obtaining feedback data for each mode; and matching a training strategy suitable for the current mode to the feedback data based on the personalized recommendation strategy.

10. The intelligent interactive training system according to claim 1, characterized in that, The terminal includes: an EEG acquisition terminal, used to acquire electroencephalographic data; The training interactive terminal is used to train and acquire interactive data, and complete the process of feature extraction, label generation and personalized recommendation strategy generation based on the integrated neurophysiological habit model and EEG big data intelligent recommendation model. The multimodal feedback terminal generates and executes intervention actions based on personalized recommendation strategies.

Images