Short video addiction-oriented electroencephalogram cross-subject identification framework
By employing a contrastive multi-domain EEG feature fusion algorithm and a contrastive learning mechanism, the problems of subjectivity in objective identification of short video addiction and cross-subject identification were solved, achieving high-precision cross-subject identification and auxiliary diagnosis of mental health problems.
Patent Information
- Application Number
- CN202511978828.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-03-06
AI Technical Summary
Existing technologies for the objective identification of short video addiction suffer from problems such as strong subjectivity, difficulty in cross-subject identification, and insufficient utilization of EEG signals. In particular, the individual differences in EEG signals limit the generalization ability of the model.
A contrastive multi-domain EEG feature fusion algorithm is adopted, including data preprocessing, frequency band separation, EEG encoder module and classification head. It combines multi-scale convolution and residual network, and trains the model through contrastive learning mechanism to reduce individual differences and achieve cross-subject recognition.
It achieves high-precision cross-subject short video addiction identification, has biological rationality, is applicable to the auxiliary diagnosis of various mental health problems, and has good recognition performance.
Smart Images
Figure CN121606302A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of neural network and EEG signal processing technology, and relates to deep learning-based spatiotemporal frequency feature extraction and comparative learning technology for EEG signals. Specifically, it relates to a deep learning framework for cross-subject identification of short video addiction using a comparative multi-domain EEG feature fusion algorithm on EEG signal data. Background Technology
[0002] In recent years, with the rapid popularization of mobile internet technology, short video applications have sprung up globally, and short video addiction has become increasingly serious, gradually becoming a significant issue affecting public mental health. Traditional addiction diagnosis methods mainly rely on questionnaires and scales. While convenient, these methods are highly subjective and struggle to objectively and in real-time reflect an individual's true psychological and physiological state. Electroencephalography (EEG), as an objective physiological signal, has garnered significant attention in the fields of brain science and mental health monitoring due to its advantages such as low acquisition cost, high temporal resolution, and non-invasiveness.
[0003] In the field of EEG signal processing and classification, early EEG analysis mainly relied on manually extracting features (such as power spectral density and wavelet transform coefficients) and combining them with traditional machine learning methods, such as support vector machines or random forests, for identification. However, these methods heavily depend on prior knowledge for feature engineering, require a great deal of expertise, and struggle to capture the complex nonlinear features in EEG signals.
[0004] With the rapid development of deep learning technology, end-to-end deep learning models such as Convolutional Neural Networks (CNNs) (e.g., EEGNet, DeepConvNet) are widely used in EEG decoding. These models can automatically learn feature representations from raw signals, significantly improving classification performance. While general-purpose deep learning models perform well on some tasks, they often neglect the multimodal characteristics of EEG signals in the frequency, time, and spatial domains, resulting in insufficient mining of addiction-related features. Furthermore, most frameworks are optimized only from the perspective of end-to-end classification tasks, without paying particular attention to their biological rationale, leading to poor interpretability.
[0005] Another key issue is the significant non-stationarity and substantial individual variability of EEG signals, which constitutes a major bottleneck limiting the practical application of existing technologies. The inherent differences in neural activity patterns among individuals make it difficult to directly transfer models trained on specific subject data to new subjects. Although some studies have attempted to alleviate this problem using contrastive learning or transfer learning, existing cross-subject frameworks often fail to fully utilize the spatial topological information of EEG signals, or their contrastive strategies are ineffective in bridging the distribution of different subjects in the feature space. How to effectively eliminate interference from individual differences while maintaining feature discriminative power, and construct a high-precision cross-subject classification model for short video addiction, is a core technical challenge that urgently needs to be addressed in this field. Summary of the Invention
[0006] To address the gaps in existing research on the objective identification of short video addiction, and the problems of insufficient signal utilization and poor cross-subject generalization of existing technologies when decoding EEG signals, this invention provides a cross-subject short video addiction identification system based on EEG signals, which can use EEG signals to identify short video addiction with high accuracy and cross-subject capability.
[0007] The short video addiction cross-subject identification system based on EEG signals provided by this invention is based on a comparative multi-domain EEG feature fusion algorithm, which includes: a data preprocessing module, a frequency band separation module, an EEG encoder module, and a classification head.
[0008] The data preprocessing module performs preprocessing operations on the raw EEG signal data, including filtering, independent component analysis, and normalization, to obtain high-quality EEG data. In addition, the preprocessing module also performs data segmentation, dividing the complete data into multiple window data. Where s represents the s-th subject, and i represents the i-th segment of the subject's EEG signal. After segmentation, data windows belonging to subjects who are addicts will be labeled 1, and those belonging to other subjects will be labeled 0.
[0009] The frequency band separation module uses a Butterworth filter to perform frequency band separation on the data before the input network, separating the data into five frequency bands: Delta, Theta, Alpha, Beta, and Gamma, denoted as... Where b represents the b-th frequency band. The data after frequency band separation are input into the EEG encoder in parallel. Each branch network has an identical structure and processes the data of one frequency band.
[0010] The EEG encoder module extracts and fuses features from the data in each frequency band, and finally projects them onto the EEG feature representation. The EEG encoding module includes a spatial feature extraction module, a temporal feature extraction module, a frequency band fusion module, and a projector module. For the input frequency band data... First, the spatial feature extraction module extracts the dependency information between EEG channels to obtain spatial features. Then, through the time feature extraction module, hierarchical time feature representations are extracted to obtain spatiotemporal features. The aforementioned spatiotemporal feature extraction process is performed in parallel across five frequency bands. The frequency band fusion module performs weighted fusion of the five spatiotemporal features to obtain a preliminary EEG representation after feature fusion. The projector maps the EEG representation to a space in which contrastive loss is applied, resulting in the final EEG feature representation. This design can significantly improve the quality of the learned representations.
[0011] The spatial feature extraction module is implemented through multiple parallel two-dimensional unfilled convolutional layers. Each convolutional layer processes the same input data, with kernel sizes ranging from small to large. Convolution is performed only on the channel dimension without changing the data size in the time dimension. After obtaining each parallel feature map, they are concatenated, and then their dimensions are mapped back to the original dimensions through another convolutional layer for further propagation. This module combines the dependency information between EEG channels at different distances and extracts EEG spatial features from different receptive field sizes to obtain EEG spatial features that fuse full-scale information.
[0012] The temporal feature extraction module is designed based on ResNet, extracting hierarchical temporal feature representations of EEG data through deep residual learning. The module takes the EEG spatial features obtained in the previous step as input and first performs preliminary temporal feature extraction through a one-dimensional convolutional layer. The core of the module consists of six residual blocks, each employing a skip connection mechanism to alleviate the gradient vanishing problem in deep networks. Each residual block uses a two-layer convolutional structure, with the number of channels expanding exponentially, doubling after every two residual blocks, ultimately forming multi-scale EEG spatiotemporal features.
[0013] The classification head maps the EEG features extracted by the aforementioned modules into a 1-dimensional numerical value, which is the final output representing the probability of short video addiction among the subjects belonging to that data window.
[0014] The training process of the contrastive multi-domain EEG feature fusion algorithm includes two stages: a contrastive representation learning stage and a classification fine-tuning stage. In the contrastive representation learning stage, only the EEG encoder is trained to obtain high-quality EEG feature representations, without enabling the classification head. In the classification fine-tuning stage, both the classification head and the EEG encoder are fine-tuned simultaneously, enabling the framework to acquire recognition and classification capabilities.
[0015] The contrastive representation learning phase includes both supervised and unsupervised strategies. The supervised strategy randomly samples signals from different categories of participants (one segment per participant) and encodes them as EEG feature representations. It maximizes the representational similarity between signals of the same category and minimizes the representational similarity between signals of different categories, aiming to teach the model how to represent different categories in different encoding spaces. The loss function for the supervised strategy is Loss. sup Defined as follows, where A and N represent the sets of addicted and non-addicted subjects, respectively, and each subject is sampled from only one data window, meaning that no two data points come from the same subject; n s τ represents the batch size. s Temperature coefficient:
[0016]
[0017] An unsupervised strategy randomly samples signals from different subjects (class-independent, two segments per subject) and encodes them into EEG feature representations. It maximizes the representational similarity between signals from different subjects and minimizes the representational similarity between signals from the same subject. The goal is for the model to learn an encoding method that represents signals from both the same and different subjects into a common space, thereby reducing the impact of individual subject differences on network training. The loss function for the unsupervised strategy is Loss. uns Defined as follows, where n u τ represents the batch size. u Temperature coefficient:
[0018]
[0019] The two strategies are trained simultaneously, and the comparison represents the total loss during the learning phase. rep Defined as:
[0020] Loss rep =Loss sup +Loss uns
[0021] The classification fine-tuning phase follows the contrastive representation learning phase. After EEG encoding, a classification head is enabled, mapping the EEG feature representations to 1D numerical values representing the addiction probability of the subject associated with that data. Training is performed by calculating the cross-entropy loss with the data labels. The loss function used in the classification fine-tuning phase is Loss. cls As shown in the following formula, where This represents the output of the classification head, where y represents the true label:
[0022]
[0023] Compared with the prior art, the advantages and positive effects of the present invention are as follows:
[0024] (1) This invention addresses the gap in existing research on the objective identification of short video addiction and is the first to design a solution specifically for the task of identifying short video addiction using EEG signals. This system directly identifies and classifies EEG signals in an end-to-end manner without the need for manual feature extraction. Experimental results have verified that this system has excellent performance in the task of identifying short video addiction.
[0025] (2) The present invention introduces a contrastive multi-domain EEG feature fusion algorithm, which has a complete and comprehensive temporal, spatial and frequency feature extraction and fusion mechanism, and can efficiently mine the implicit information of EEG signals. In addition, the framework is trained through a unique contrastive learning mechanism, and the encoding module of the framework is fine-tuned through supervised and unsupervised strategies to enable it to obtain the encoding ability to fully represent EEG signals. Combining the two mechanisms, the framework can greatly reduce the individual differences of EEG signals and achieve cross-subject recognition.
[0026] (3) The system of the present invention can be used as a general solution for the identification of mental health problems with the aid of EEG signals. It is not only applicable to short video addiction, but can also be extended to other mental health scenarios, such as pornography addiction. Through experimental verification, the system of the present invention has shown excellent recognition performance in different scenarios and has a certain biological rationality, which helps to provide auxiliary support for the identification and intervention of related mental health problems. Attached Figure Description
[0027] Figure 1 This is a diagram illustrating the main framework implementation of the EEG cross-subject recognition system according to an embodiment of the present invention.
[0028] Figure 2 This is a comparison chart of the recognition system of this invention and other deep learning EEG classification models in terms of short video addiction recognition performance;
[0029] Figure 3 This is a comparison chart of the pornography addiction recognition performance of the recognition system of this invention with other models;
[0030] Figure 4 This is a frequency band weight heatmap of the recognition system of the present invention during training on different data. Detailed Implementation
[0031] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0032] This invention addresses the significant individual differences in EEG signals and the lack of objective diagnostic criteria in cross-subject identification tasks for short video addiction. It constructs a multi-domain feature extraction encoder based on a deep learning classification model, decouples spatiotemporal features using multi-scale convolution and residual networks, introduces a learnable weighted frequency band fusion mechanism, and designs a dual-strategy contrastive representation learning module. This module combines a supervised strategy aimed at differentiating addiction categories with an unsupervised strategy aimed at bridging the differences in representations among different subjects for joint optimization, effectively reducing inter-subject distributional differences and accurately identifying addiction states. Furthermore, this invention conducts frequency band weight analysis to verify the consistency between the system's operation and neurophysiology. This invention not only overcomes the subjectivity limitations of traditional questionnaire-based diagnoses but also provides a highly robust and physiologically interpretable general solution for the auxiliary diagnosis of mental disorders based on EEG signals.
[0033] This invention uses a short video addiction recognition task as an example to illustrate the implementation of an EEG cross-subject recognition framework for short video addiction. The application environment used is: Ubuntu system, Python 3.8, and PyTorch deep learning architecture. The core of the EEG cross-subject recognition framework for short video addiction is a contrastive multi-domain EEG feature fusion algorithm, including: a data preprocessing module, a frequency band separation module, an EEG encoder module, and a classification head. The main improvement of this invention lies in the training process of the EEG encoder module and the framework, such as... Figure 1 As shown.
[0034] The data preprocessing module preprocesses the raw EEG signal data to obtain high-quality EEG data. The entire preprocessing process is completed using built-in methods of python-mne 1.6.1. The preprocessing module first performs a 50Hz mains frequency filter to remove mains interference, followed by a 1-50Hz bandpass filter. After filtering, automatic independent component analysis is performed to remove artifacts. Finally, the min-max normalization method is used to linearly map all data to the range [0,1], eliminating the influence of differences in the absolute values of EEG signals among different subjects. After data preprocessing, a sliding window method is used to segment the data, selecting a 3-second window size with 66% overlap (2 seconds) between adjacent windows to eliminate phase shift effects. This expands the data while enhancing the model's ability to classify data using only a few seconds of data. Since the data was collected from subjects with or without addiction, the data preprocessing module assigns the same label to all window data for each subject: 0 for non-addictive and 1 for addicted.
[0035] In this embodiment of the invention, the data preprocessing module processes the segmented data segments as follows: Where s represents the s-th subject, i represents the i-th segment of the subject's EEG signal, C represents the number of channels in the EEG data, and T represents the number of sampling points for this segment of data. For example... Figure 1 As shown, before signal encoding, the frequency band separation module first separates signals of different frequencies in the EEG signal. It uses a 4th-order Butterworth filter for bandpass filtering, separating the data into five frequency bands: Delta (1-3Hz), Theta (4-7Hz), Alpha (8-13Hz), Beta (14-30Hz), and Gamma (31-50Hz). The frequency band data is denoted as... Where b represents the b-th frequency band.
[0036] The EEG encoder module is the core of the entire system. It extracts and fuses features from data in each frequency band and finally projects them onto the EEG feature representation. The EEG encoder module of this invention possesses a complete and comprehensive temporal, spatial, and frequency feature extraction and fusion mechanism, enabling efficient mining of implicit information from EEG signals. Its spatial feature extraction module employs full-scale convolution to extract and integrate all distance-dependent EEG channel dimensional information. The temporal feature extraction module uses a residual network structure with progressively increasing convolutional kernel size to extract multi-level time-series features. The learnable frequency band fusion module has biological rationale. Compared with feature extraction methods in other deep learning frameworks, the EEG encoder module of this invention extracts and represents more comprehensive EEG features, providing a new perspective for future related research.
[0037] like Figure 1 As shown, the implementation structure of the EEG encoder module includes a spatial feature extraction module, a temporal feature extraction module, a frequency band fusion module, and a projector module.
[0038] The spatial feature extraction module is implemented based on multiple convolutional neural networks. Frequency band data. First, the data is copied C times and implemented through C parallel two-dimensional unfilled convolutional layers. The kernel size of each convolutional layer ranges from (1,1) to (C,1), and convolution is performed only on the channel dimension without changing the data size in the time dimension. By designing convolutional kernels of varying sizes to process the data separately, the spatial dimension features of the EEG data can be extracted from different receptive field sizes by combining the dependency information between channels at different distances.
[0039] Conv of the kth convolutional layer k The convolution kernel size is (k,1), and spatial features are obtained after convolution. As shown in the following equation, where Conv(·) represents the convolution operation. Since the convolutional layer has no padding, The number of channels and They are not the same.
[0040]
[0041] After obtaining the outputs of C convolutional layers, they are concatenated into a long vector. The spatial features at the full scale are represented by the following formula, where [·, ·] represent the stitching operation:
[0042]
[0043] Full-scale spatial features The size is too large, so as shown in the following formula, it is then passed through a convolutional layer Conv. con Will Mapped to As the final spatial feature delivery time feature extraction, D1 is the intermediate dimension, which is set to 16 in this embodiment of the invention.
[0044]
[0045] The temporal feature extraction module is implemented based on a residual network, extracting hierarchical temporal feature representations of EEG data through deep residual learning. The module takes as input the EEG spatial features extracted in the previous step. First, a preliminary temporal feature extraction is performed on the input features using a one-dimensional convolutional layer. The core of the temporal feature extraction module consists of six residual blocks, each employing a skip connection mechanism to alleviate the gradient vanishing problem in deep networks. Each residual block uses a two-layer convolutional structure, with the number of channels expanding exponentially, doubling after every two residual blocks, ultimately forming a multi-scale EEG temporal feature representation.
[0046] Each convolution in the module is followed by batch normalization, and then a non-linear transformation is performed using the ReLU activation function. The module integrates a Dropout mechanism with a probability p set to 0.2, effectively mitigating overfitting. In the final stage, global average pooling is used to aggregate temporal features, and a fully connected layer outputs a D2-dimensional (64-dimensional in this embodiment) feature vector, providing a compact and expressive EEG feature representation. The processing procedure of the time feature extraction module can be expressed as the following formula, where This represents the process of extracting the time features of the b-th frequency band:
[0047]
[0048] like Figure 1 As shown, after the spatiotemporal feature extraction is completed, the next step is the frequency band fusion module. This module performs a weighted summation of the five frequency band feature vectors extracted in parallel to obtain the fused representation.
[0049] The frequency band feature fusion module takes D2-dimensional features from B frequency bands as input data. The features of each frequency band are mapped to 1D using a linear layer, then the values are concatenated, and finally converted into B frequency band weights using a softmax function. During training, the network automatically adjusts the weights of different frequency bands, highlighting the bands that have a greater impact on classification. The five frequency bands are then weighted and fused to output the fused D2-dimensional features. As shown in the following formula, where ω b Weights for the corresponding frequency bands:
[0050]
[0051] like Figure 1 As shown, the feature vector after band fusion is fed into the projector, which is implemented with two linear layers, using the ReLU function for non-linear mapping between the two layers. The projector maps the current EEG feature vector to the space where contrastive loss is applied. This design significantly improves the quality of the learned representation. The projector extracts the features... Mapped to D dimensions (set to 32 in this embodiment) as the final EEG feature representation. As shown in the following formula, where Proj(·) represents the projection operation:
[0052]
[0053] After projection, the system obtains the final EEG feature representation, which combines full-scale spatial features with multi-level temporal features. It undergoes multi-band processing and weighted fusion, and finally, after projection, the feature representation is comprehensive and the information is highly condensed. However, when using this system for short video addiction identification, a classification head structure is still needed to map the EEG feature representation to the classification result.
[0054] The classification head maps EEG features to 1-dimensional numerical values, which are then used as model outputs to represent the addiction probability of the subject associated with that data, as shown in the following formula:
[0055]
[0056] The above describes the framework structure of the system of the present invention. While designing an efficient feature extraction network, the system of the present invention also designs an innovative training method adapted to the framework. It combines two stages: optimized contrastive representation learning and classification fine-tuning. It can gradually enable the EEG encoder to learn a more robust encoding method through supervised and unsupervised training strategies, thereby reducing the impact of individual differences in EEG signals on the system's training and inference, and ultimately achieving cross-subject recognition.
[0057] like Figure 1As shown, in the contrastive representation learning phase, the EEG feature representations obtained by the EEG encoder are used for training two contrastive strategies, supervised and unsupervised, without enabling the classification head.
[0058] A supervised strategy randomly samples signals from different categories of participants (one segment per participant) for encoding representation, maximizing the representation similarity between signals of the same category and minimizing the representation similarity between signals of different categories. The goal is for the encoder to learn how to represent different categories in different encoding spaces. The loss function of this strategy is defined as follows, where A and N represent the sets of addicted and non-addicted participants (where each participant is sampled only one segment, i.e., no two segments come from the same participant), and n... s τ is the batch size (set to 4 in this embodiment). s The temperature coefficient (set to 0.1 in this embodiment of the invention):
[0059]
[0060] An unsupervised strategy randomly samples signals from different subjects (class-independent, two segments per subject) for encoding representation, maximizing the representation similarity between signals from different subjects and minimizing the representation similarity between signals from the same subject. The aim is for the model to learn an encoding method that represents signals from both the same and different subjects into a common space, thereby reducing the impact of individual subject differences on network training. The loss function for this strategy is defined as follows, where n... u τ is the batch size (set to 8 in this embodiment of the invention). u The temperature coefficient (set to 0.07 in this embodiment of the invention):
[0061]
[0062] The two strategies are trained simultaneously, and the total loss during the learning phase is defined as follows:
[0063] Loss rep =Loss sup +Loss uns
[0064] The classification fine-tuning phase follows the contrastive representation learning phase. After EEG encoding, a classification head is enabled, mapping the EEG feature representations to 1D numerical values representing the addiction probability of the subject associated with that data point. Training is performed by calculating the cross-entropy loss with the data labels. The loss function used in the classification fine-tuning phase is Loss. cls As shown in the following formula, where This represents the output of the classification head, where y represents the true label:
[0065]
[0066] The embodiments of this invention demonstrate the superiority of the system through experiments. To measure the cross-subject classification ability of the system, the training and test sets are divided by subject to ensure that a subject's signal does not appear in both the training and test sets simultaneously. The training and test sets are divided in a 9:1 ratio. To reflect the average performance of the system across different groups, stratified ten-fold cross-validation was used. Twenty subjects were divided into 10 groups, each containing one addicted and one non-addicted subject. In each experiment, one group was used as the test set, and the other nine groups were used as the training set. The model's performance was evaluated after ten experiments using the average performance metrics. Accuracy (Acc), Precision (Pre), Recall (Rec), and F1 score were selected as performance evaluation metrics for the binary classification task.
[0067] The experimental setup for this embodiment of the invention is as follows: In the contrast representation learning phase, the batch size for each training round is 20, and each batch contains 4 subjects sampled under supervised policy (4 signal segments in total) and 4 subjects sampled under unsupervised policy (8 signal segments in total). The initial learning rate for the contrast representation learning phase is set to 0.0003, and a total of 50 training rounds are conducted. In the classification fine-tuning phase, the batch size for each training round is 128, and the initial learning rate is set to 0.0005, for a total of 100 training rounds. The AdamW optimizer and gradient constraints are used in both phases to prevent excessive gradient updates, and learning rate decay is implemented, multiplying the learning rate by 1 / e every 10 rounds. The experiment is conducted on two RTX 4080S graphics cards, with a single-fold training time of approximately 3 hours during the training phase.
[0068] like Figure 2 As shown, the short video addiction identification system implemented in this embodiment of the invention is compared with several widely used EEG deep learning frameworks, including LightCNN, EEGNet, DeepConvNet, and ShallowNet. From Figure 2 As can be seen, the recognition system of this invention outperforms other frameworks in all indicators, verifying the effectiveness of the model.
[0069] Depend on Figure 3 As can be seen, in the task of identifying psychological health issues similar to short video addiction—specifically pornography addiction—the system of this invention achieved the best performance in terms of accuracy and precision compared to other widely used EEG deep learning frameworks, and also demonstrated competitive results in other metrics. Therefore, the system of this invention has strong versatility and significant advantages in identifying psychological health issues based on EEG signals, providing a new solution for the auxiliary identification of psychological health problems.
[0070] like Figure 4As shown, in the embodiment of the present invention, the frequency band weights of the frequency band fusion module in the trained model were also observed during the experiment. The horizontal axis represents 5 frequency bands, and the vertical axis represents the 10 different folds in the 10-fold cross-validation process.
[0071] From the perspective of cognitive neuroscience, different frequency bands of EEG signals carry different physiological meanings: the Delta band reflects deep sleep or unconscious state, the Theta band is related to relaxation, meditation, and attention, the Alpha band reflects a state of wakefulness but relaxation, the Beta band is related to focused attention or active state, and the Gamma band reflects higher cognitive functions such as memory and consciousness.
[0072] Figure 4 (1) and (2) represent the frequency band weights of the resting-state data. Compared to other task-state data, the model tends to amplify low-frequency signals when processing resting-state data. Figure 4 (1) Data for eyes open, Figure 4 (2) The data is for closed eyes. The frequency band weight of the open eyes data is higher in the Alpha band, while the closed eyes data is more biased towards low frequency signals, indicating that the brain activity is more relaxed compared to the open eyes state. This corresponds exactly to the physiological significance of EEG frequency bands.
[0073] Figure 4 (3)(4)(5) represent the frequency band weights of the video viewing data, which are more biased towards higher frequencies compared to the resting state data, symbolizing more active brain activity. Figure 4 (3) Data representing the viewing of short videos of interest shows that the frequency band weights are concentrated in the Beta and Gamma bands, indicating that the brain is in an active state. In contrast, Figure 4 (4) For data on watching short videos that are not of interest, the frequency band weight is higher at Beta. Figure 4 (5) Data for watching long videos shows higher band weights at the Gamma level. When faced with short videos that are not of interest, even if the brain is active, attention is not as focused as when watching content that is of interest, so the band weights are biased towards the lower-frequency Beta band. Watching long videos requires more focused attention, and unlike short videos that end quickly, the brain also needs to remember the content of the previous videos as it is watched, so the band weights are biased towards the higher-frequency Gamma band. This corresponds precisely to the physiological significance of EEG bands.
[0074] Based on the above experiments, the embodiments of the present invention found that the system will automatically amplify the frequency bands related to different data during the training process. This not only demonstrates the effectiveness of the system's frequency band fusion module, but also confirms that the system of the present invention has biological rationality and exhibits a certain degree of alignment with real physiological processes.
[0075] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.
Claims
1. An electroencephalogram cross-subject identification framework for short video addiction, characterized in that, The method comprises the following steps: a data preprocessing module is used to receive raw EEG signal data, preprocess and cut the data, and obtain EEG data windows with addiction labels; a frequency band separation module is used to connect the data preprocessing module, separate the EEG data windows into multiple frequency band data of different frequencies; an EEG encoder module is used to connect the frequency band separation module, receive each frequency band data, and perform feature extraction and fusion in parallel, and finally output an EEG feature representation; a classification head is used to connect the EEG encoder module, map the EEG feature representation to a probability value of short video addiction; wherein the EEG encoder module is configured to be pre-trained through a contrastive representation learning stage, and the contrastive representation learning stage is combined with a supervised strategy and an unsupervised strategy for joint optimization to obtain cross-subject EEG feature representation.
2. The EEG cross-subject identification framework for short video addiction according to claim 1, wherein, The EEG encoder module comprises: a spatial feature extraction module for extracting EEG inter-channel dependency information for each input frequency band data to obtain spatial features; a temporal feature extraction module connected to the spatial feature extraction module for extracting hierarchical temporal feature representation of the spatial features to obtain spatio-temporal features; a frequency band fusion module connected to the temporal feature extraction module for weighting and fusing the spatio-temporal features corresponding to each frequency band to obtain a preliminary EEG representation; and a projector module connected to the frequency band fusion module for mapping the preliminary EEG representation to a space applying a contrastive loss to obtain a final EEG feature representation.
3. The EEG cross-subject identification framework for short video addiction according to claim 2, characterized in that, The spatial feature extraction module comprises: a plurality of parallel two-dimensional non-padding convolution layers, each convolution layer is used to process the same frequency band data, and the convolution kernel size of each convolution layer is different, and only the channel dimension is convolved; a splicing unit for splicing the feature maps output by the plurality of parallel two-dimensional non-padding convolution layers to form full-scale spatial features; and a mapping convolution layer for mapping the dimension of the full-scale spatial features to a preset dimension to output the spatial features.
4. The EEG cross-subject identification framework for short video addiction according to claim 3, characterized in that, The number of the plurality of parallel two-dimensional non-padding convolution layers is equal to the number of EEG data channels C, and the convolution kernel size of the kth convolution layer is (k, 1), wherein k is valued from 1 to C.
5. The EEG cross-subject identification framework for short video addiction according to claim 2, wherein, The temporal feature extraction module is constructed based on a residual network and comprises: a one-dimensional convolution layer for performing preliminary temporal feature extraction on the input spatial features; a plurality of cascaded residual blocks, each residual block comprising a skip connection mechanism and a two-layer convolution structure, and the number of channels is expanded in an exponential growth mode according to the cascade depth; a global average pooling layer for aggregating temporal features; and a full connection layer for outputting the spatio-temporal features.
6. The EEG cross-subject identification framework for short video addiction according to claim 2, wherein, The frequency band fusion module comprises: a linear mapping unit for mapping and converting the spatio-temporal features of each frequency band to corresponding frequency band weights; and a weighted summation unit for weighting and fusing the spatio-temporal features of all frequency bands according to the frequency band weights.
7. The EEG cross-subject identification framework for short video addiction according to claim 1, wherein, The frequency band separation module is configured with a Butterworth filter for separating data according to five frequency bands of Delta, Theta, Alpha, Beta and Gamma; the EEG encoder module comprises five parallel branches with identical structures, which are respectively used for processing data of the five frequency bands.
8. The EEG cross-subject identification framework for short video addiction according to claim 1, wherein, The system is configured to perform the contrastive representation learning stage by: using the supervised strategy, randomly sampling signals from subjects of different categories, maximizing the representation similarity between signals of the same category, and minimizing the representation similarity between signals of different categories, with a loss function as follows, where A, N represent the subjects of addiction and non-addiction, n s is the batch size, τ s is the temperature coefficient: With the unsupervised strategy, signals from different subjects are randomly sampled, maximizing the representation similarity between signals from different subjects and minimizing the representation similarity between signals from the same subject, with the loss function as follows, where n u is the batch size, τ u is the temperature coefficient: The total loss of the contrast representation learning stage is composed of the loss of the supervised strategy and the loss of the unsupervised strategy, and is defined as follows: Loss rep = Loss sup + Loss uns .
9. The EEG cross-subject identification framework for short video addiction according to claim 1, wherein, The system is also configured to perform a classification fine-tuning stage after the contrast representation learning stage; in the classification fine-tuning stage, the system enables the classification head, calculates the cross-entropy loss with data with addiction labels, and fine-tunes the classification head and the EEG encoder module.