Cognitive state recognition system based on single-channel electroencephalogram signal task perception and multi-scale modeling
By combining multi-scale convolutional neural networks and Transformer architecture with a significant frequency band attention mechanism, the problem of limited information and noise interference in cognitive state recognition using single-channel EEG signals is solved, achieving high-precision and low-cost cognitive state recognition, which is suitable for tasks such as emotion recognition and workload assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-07
AI Technical Summary
Single-channel EEG signals have limitations in cognitive state recognition, including limited information, high noise interference, and limitations in time and frequency domain modeling. They are difficult to adapt to the needs of different cognitive tasks, resulting in poor recognition accuracy and task adaptability.
A multi-scale convolutional neural network combined with a Transformer architecture is used for temporal feature extraction. A time-frequency representation is generated through continuous wavelet transform and an independent encoder is designed. A significant frequency band attention mechanism is used to adaptively identify task-related frequency bands. By integrating temporal and frequency domain features, high-precision cognitive state recognition is achieved.
It achieves high-precision cognitive state recognition, reduces hardware and modeling costs, adapts to different cognitive tasks, improves noise resistance and recognition stability, provides interpretable analysis, and is suitable for fields such as portable EEG devices and telemedicine.
Smart Images
Figure CN121808334A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of brain-computer interface and neural engineering technology, and in particular to a cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling. Background Technology
[0002] With the rapid development of brain-computer interface technology and artificial intelligence, cognitive state recognition based on electroencephalography (EEG) has received increasing attention in fields such as emotion detection and workload assessment. These studies have not only deepened our understanding of brain function but also expanded the application scope of human-computer interaction systems. Due to its high temporal resolution, non-invasiveness, and relatively low cost, EEG has become a primary technical means for cognitive state assessment.
[0003] Most studies rely on multi-channel EEG covering the entire scalp to achieve high accuracy, but this method is costly in terms of data acquisition and inconvenient in practical applications. In contrast, single-channel EEG offers a user-friendly and hardware-efficient alternative, requiring only one or a few electrodes to acquire signals, significantly reducing cost and system complexity, making it more promising for applications in wearable devices, telemedicine, and daily health monitoring. However, single-channel EEG also faces significant challenges. Due to the lack of spatial information, it is difficult to utilize traditional spatial feature extraction methods; single-channel signals are more susceptible to noise and artifacts, resulting in a relatively low signal-to-noise ratio; and existing methods often employ a uniform feature extraction strategy, ignoring the differences in neural activity patterns across different cognitive tasks, leading to poor task adaptability.
[0004] Existing single-channel EEG cognitive recognition methods suffer from two major limitations. Firstly, temporal modeling is limited because most methods employ convolutional networks or recurrent neural networks with a single receptive field, failing to effectively capture the dynamic changes in neural activity across multiple time scales during different cognitive processes. Research shows that neural activity in different cognitive tasks varies significantly across time scales; for example, short-term local fluctuations may reflect rapid neural responses, while long-term trends may represent the maintenance of attention or working memory. Single-scale modeling struggles to adapt to diverse task requirements. Secondly, frequency domain modeling is limited because traditional methods typically apply a uniform encoder to all frequency bands (δ, θ, α, β, γ), ignoring the differences in representational ability and task relevance across different frequency bands. Neuroscience research has thoroughly demonstrated that different cognitive tasks rely on different dominant frequency bands: for example, emotion recognition is primarily associated with high-frequency activity, while workload assessment relies more on power variations in the β band. Against this backdrop, designing a cognitive state recognition system that fully leverages the portability of single-channel EEG while overcoming its limited information and achieving high accuracy and task adaptability has become a significant technical challenge. Summary of the Invention
[0005] Purpose of the invention: This invention provides a cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling. It can efficiently and accurately extract task-related features from single-channel EEG signals, achieve high-precision cognitive state recognition, and is suitable for various application scenarios such as emotion recognition and workload assessment.
[0006] Technical Solution: The present invention discloses a cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling, comprising: a task-aware temporal feature extraction module, a frequency-domain independent feature extraction module, a task-aware frequency-domain feature fusion module, an adaptive feature fusion module, and a result output module; the task-aware temporal feature extraction module employs a multi-scale convolutional neural network combined with a Transformer architecture to capture multi-time-scale features from local to global; the frequency-domain independent feature extraction module generates time-frequency representations through continuous wavelet transform and designs independent encoders for different frequency bands to extract band-specific features; the task-aware frequency-domain feature fusion module adaptively identifies task-related frequency bands and guides cross-band feature fusion through a salient frequency band attention mechanism; the adaptive feature fusion module integrates temporal and frequency domain features using an attention mechanism; and the result output module completes cognitive state classification and provides interpretable analysis results.
[0007] Furthermore, the task-aware temporal feature extraction module includes a multi-scale local feature encoder and a global feature encoder, given a single-channel EEG input signal. The multi-scale local feature encoder first processes the time series through three parallel one-dimensional convolutional layers with kernel sizes of 16, 8, and 4, respectively. Then, batch normalization and ReLU activation are performed to connect the multi-scale features along the channel dimension, and these features are merged through depthwise convolution. Finally, a Transformer-based global feature encoder is used to capture long-range temporal dependencies. The final output is a temporal representation of the EEG signal, denoted as . .
[0008] Furthermore, in the frequency domain independent feature extraction module, given the input signal... First, continuous wavelet transform is applied to map it into a two-dimensional time-frequency diagram. Where F represents the maximum frequency, set to 50Hz; M is divided into five frequency bands: ,set up For frequency band k, the frequency index set is [set], and the data segment for frequency band k is [data segment]. .
[0009] Furthermore, for each frequency band k, an independent spectrum encoder is trained. ,in It is the set of trainable parameters in the encoder to extract frequency band-specific features. : .
[0010] Furthermore, each Employing a separable time-frequency convolution design, it is used along the time axis. Convolution models short-term dynamics, using along the frequency axis. Convolution captures cross-frequency dependencies, followed by ReLU activation and pooling, resulting in... Frequency band specific characteristics as downstream components.
[0011] Furthermore, the task-aware frequency domain feature fusion module uses a salient frequency band attention fusion mechanism to achieve task-adaptive feature integration, given the features of each frequency band. The contribution score for each frequency band is calculated using a lightweight MLP: The significance weights are obtained through Softmax normalization: ,in This reflects the significance of frequency band k for the current task; the most significant frequency band is selected: Based on the characteristics of the most prominent frequency band To guide the process, features from all frequency bands are fused using a multi-head cross-attention mechanism. This mechanism focuses on complementary information in non-salient frequency bands, suppresses redundancy and noise, and generates task-aware frequency domain feature representations, ultimately yielding the frequency domain features. .
[0012] Furthermore, the adaptive feature fusion module is responsible for integrating temporal and frequency domain features and completing the final cognitive state classification; after obtaining temporal features... and frequency domain features Then, attention-based time-frequency fusion is performed, the concatenated features are linearly projected into queries, keys and values, and self-attention processing is used to model cross-feature correlations, finally obtaining the fused feature representation Z.
[0013] Furthermore, the fused feature Z is passed through a fully connected layer and a Softmax activation function to output the probability distribution of each cognitive state category. Cross-entropy loss is used as the objective function, and the specific formula is as follows:
[0014]
[0015] Where K represents the number of categories, It's a real label.
[0016] Furthermore, the results output module outputs cognitive state category labels, generates interpretability analysis information based on the significant frequency band attention mechanism, calculates the contribution score of each frequency band through a lightweight MLP, and obtains the corresponding significance weight through Softmax. The higher the weight of the frequency band, the greater its contribution to the cognitive state discrimination of the current sample.
[0017] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention achieves a key breakthrough in cognitive state recognition using single-channel EEG through the organic combination of task-aware temporal modeling, frequency-domain independent modeling, and task-aware frequency-domain fusion; the task-aware temporal feature extraction module utilizes multi-scale convolution and Transformer to capture neural dynamic patterns from short to long; the frequency-domain independent feature extraction module generates time-frequency representations through continuous wavelet transform and constructs independent encoders for each frequency band, effectively extracting specific features of each band; the task-aware frequency-domain feature fusion module adaptively locates task-related frequency bands through a significant frequency band attention mechanism, thereby reducing redundancy and effectively modeling local and cross-frequency band dependencies; the adaptive feature fusion module integrates temporal and frequency domain features through self-attention, significantly improving noise resistance and recognition stability; the result output module provides accurate cognitive state classification results and interpretable analysis information; the system only requires a single-channel EEG to achieve high-precision recognition, reducing hardware and modeling costs, and can automatically adapt to different cognitive tasks, providing strong support for the application of portable EEG devices in fields such as affective computing, human-computer interaction, and telemedicine. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the system structure of the present invention.
[0019] Figure 2 This is a schematic diagram of the frequency domain independent feature extraction module of the present invention.
[0020] Figure 3 This is a schematic diagram of the task-aware frequency domain feature fusion module of the present invention.
[0021] Figure 4 This is a schematic diagram of the task-aware temporal feature extraction module of the present invention.
[0022] Figure 5 This is a schematic diagram of the adaptive feature fusion module structure of the present invention.
[0023] Figure 6 This is a graph showing the proportion of different frequency bands selected as significant frequency bands in the two tasks of this invention.
[0024] Figure 7 This is a comparison chart showing the recognition accuracy of different electrode positions in two tasks according to the present invention. Detailed Implementation
[0025] like Figure 1 As shown, a cognitive state recognition system based on task perception and multi-scale modeling using single-channel EEG signals includes: a task-aware temporal feature extraction module, a frequency-domain independent feature extraction module, a task-aware frequency-domain feature fusion module, an adaptive feature fusion module, and a result output module. The task-aware temporal feature extraction module uses a multi-scale convolutional neural network combined with a Transformer architecture to capture multi-time-scale features from local to global. The frequency-domain independent feature extraction module generates time-frequency representations through continuous wavelet transform and designs independent encoders for different frequency bands to extract band-specific features. The task-aware frequency-domain feature fusion module adaptively identifies task-related frequency bands and guides cross-band feature fusion through a salient frequency band attention mechanism. The adaptive feature fusion module integrates temporal and frequency domain features using an attention mechanism. The result output module completes cognitive state classification and provides interpretable analysis results.
[0026] To fully capture the temporal representation in EEG signals, a two-stage temporal feature extractor composed of a multi-scale local feature encoder and a global feature encoder was designed. For example... Figure 4 As shown, the multi-scale local feature encoder models signals across different receptive fields to capture dynamic patterns of brain signals at multiple time scales. The global feature encoder focuses on modeling long-range dependencies, extracting contextual information from the entire EEG sequence.
[0027] Given a single-channel EEG input signal The multi-scale local feature encoder first processes the time series through three parallel one-dimensional convolutional layers with kernel sizes of 16, 8, and 4, followed by batch normalization and ReLU activation. This design enables the extraction of temporal patterns at multiple scales, thereby capturing both fine-grained and coarse-grained dynamic characteristics of EEG signals. The multi-scale features are then concatenated along the channel dimension and merged through depthwise convolution. Subsequently, a Transformer-based global feature encoder is employed to capture long-range temporal dependencies. Specifically, the Transformer layer uses self-attention to capture global dependencies, where queries and values are normalized using layer normalization before entering the attention layer to ensure training stability. The final output is a temporal representation of the EEG signal, denoted as . .
[0028] Given input signal First, continuous wavelet transform is applied to map it into a two-dimensional time-frequency diagram. Where F represents the maximum frequency, set to 50Hz. Based on prior knowledge, M is divided into five frequency bands: .set up This is the frequency index set for frequency band k. The data segment for frequency band k is... .
[0029] For each frequency band k, train an independent spectrum encoder. ,in It is the set of trainable parameters in the encoder to extract frequency band-specific features. : .
[0030] Each Employing a separable time-frequency convolution design, such as... Figure 2 As shown, use along the time axis Convolution models short-term dynamics, using along the frequency axis. Convolution captures cross-frequency dependencies, followed by ReLU activation and pooling. Results Frequency band specific characteristics as downstream components.
[0031] The task-aware frequency domain feature fusion module is one of the core innovations of this invention. It aims to adaptively fuse features from various frequency bands according to different cognitive tasks, constructing a task-aware frequency domain feature representation. For example... Figure 3 As shown, this module employs a significant frequency band attention fusion mechanism to achieve task-adaptive feature integration. Given features in each frequency band... The contribution score for each frequency band is calculated using a lightweight MLP: The significance weights are obtained through Softmax normalization: ,in This reflects the significance of frequency band k for the current task. The most significant frequency band is selected: Features of the most prominent frequency band To guide this process, features from all frequency bands are fused using a multi-head cross-attention mechanism. This cross-attention mechanism focuses on complementary information in non-salient frequency bands, suppresses redundancy and noise, and generates task-aware frequency domain feature representations, ultimately yielding the frequency domain features. .
[0032] The adaptive feature fusion module is responsible for integrating time-domain and frequency-domain features and completing the final cognitive state classification, such as... Figure 5 As shown.
[0033] Time-frequency feature fusion: obtaining time-domain features and frequency domain features Then, attention-based time-frequency fusion is performed. The concatenated features are linearly projected into queries, keys, and values, and self-attention processing is used to model cross-feature correlations, ultimately yielding the fused feature representation Z.
[0034] Loss function: The fused feature Z is passed through a fully connected layer and a softmax activation function, outputting the probability distribution of each cognitive state category. Cross-entropy loss is used as the objective function, with the specific formula as follows:
[0035]
[0036] Where K represents the number of categories, It's a real label.
[0037] The core function of the output module is to process and present the recognition results to ensure their accuracy and interpretability, facilitating subsequent analysis and application. This module outputs cognitive state category labels and provides significant frequency band recognition results and time-frequency feature importance analysis, offering reliable technical support and decision-making basis for clinical diagnosis, human-computer interaction systems, and neuroscience research.
[0038] A comprehensive evaluation was conducted on two publicly available datasets: the SEED dataset for emotion recognition tasks, which uses 15 Chinese film clips to induce three emotional states (positive, negative, and neutral), containing 3,600 samples; and the PWED dataset for workload assessment tasks, collected through a flight simulation paradigm. Specifically, workload levels (low, medium, and high) were defined based on the difficulty setting of the flight simulation task and the density of triggering events per unit time. Low workload corresponds to simple cruise tasks, while high workload corresponds to complex tasks involving handling sudden malfunctions or high-density operations. This invention uses only one channel of signal as input per iteration. Evaluation metrics included accuracy and F1 score, assessed using 5-fold cross-validation, with the final result being the average of five repeated experiments. Experimental results are shown in Table 1. Table 1 shows that the system of this invention achieved state-of-the-art performance on both datasets with a smaller standard deviation, indicating better stability and robustness. Notably, most existing methods exhibit significant performance variations across different tasks, while the system of this invention maintains excellent and stable performance on both tasks, fully demonstrating its task adaptability.
[0039] Table 1. Comparison experiments between the system and existing methods on two datasets.
[0040]
[0041] Ablation experiments were conducted to verify the effectiveness of the system and to validate each component. The experimental results are shown in Table 2. Table 2 shows that even the basic time-frequency fusion framework achieves good performance, demonstrating the importance of joint time-frequency modeling. Adding MS-TE and TA-SE both improve performance, proving the effectiveness of the two modules. The complete system achieves optimal performance, validating the synergistic effect between modules. In the emotion recognition task, MS-TE plays a more significant role; in the workload assessment task, TA-SE contributes more, consistent with the neural mechanism characteristics of different tasks.
[0042] Table 2 Validation of each module of the system
[0043]
[0044] To verify the reliability of the task-aware frequency domain encoder, the proportion of different frequency bands selected as significant bands in the test set was statistically analyzed. For example... Figure 6 As shown, experimental results indicate that in emotion recognition tasks, the γ band (30-50Hz) was most frequently selected as the salient band, while in workload assessment tasks, the β band (13-30Hz) was most frequently selected as the salient band. These findings are highly consistent with prior knowledge in the field of neuroscience, fully validating the effectiveness and interpretability of the task perception module of this invention.
[0045] The accuracy of identifying 15 candidate electrode locations on two tasks was analyzed, and the results are as follows: Figure 7 As shown in the figure. Experiments revealed that in the emotion recognition task, the AF3 electrode (frontal lobe region) achieved the highest accuracy, with most dominant electrodes concentrated in the frontal lobe, consistent with the key role of the frontal lobe in emotion regulation, decision-making, and impulse control. In the workload assessment task, the PO8 electrode (occipital lobe region) achieved the highest accuracy, with dominant electrodes mainly located in the occipital lobe, consistent with the nature of the visual stimulus task. These results further validate that the system of the present invention can effectively capture task-related neural activity patterns.
Claims
1. A cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling, characterized in that, include: The system comprises a task-aware temporal feature extraction module, a frequency-domain independent feature extraction module, a task-aware frequency-domain feature fusion module, an adaptive feature fusion module, and a result output module. The task-aware temporal feature extraction module employs a multi-scale convolutional neural network combined with a Transformer architecture to capture multi-time-scale features from local to global perspectives. The frequency-domain independent feature extraction module generates time-frequency representations through continuous wavelet transforms and designs independent encoders for different frequency bands to extract band-specific features. The task-aware frequency-domain feature fusion module adaptively identifies task-related frequency bands and guides cross-band feature fusion through a salient frequency band attention mechanism. The adaptive feature fusion module integrates time-domain and frequency-domain features using an attention mechanism; The results output module completes the cognitive state classification and provides interpretable analysis results.
2. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 1, characterized in that, The task-aware temporal feature extraction module includes a multi-scale local feature encoder and a global feature encoder, given a single-channel EEG input signal. The multi-scale local feature encoder first processes the time series through three parallel one-dimensional convolutional layers with kernel sizes of 16, 8, and 4, respectively. Then, batch normalization and ReLU activation are performed to connect the multi-scale features along the channel dimension, and these features are merged through depthwise convolution. Finally, a Transformer-based global feature encoder is used to capture long-range temporal dependencies. The final output is a temporal representation of the EEG signal, denoted as . .
3. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 1, characterized in that, In the frequency domain independent feature extraction module, given the input signal First, continuous wavelet transform is applied to map it into a two-dimensional time-frequency diagram. Where F represents the maximum frequency, set to 50Hz; M is divided into five frequency bands: ,set up For frequency band k, the frequency index set is [set], and the data segment for frequency band k is [data segment]. .
4. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 3, characterized in that, For each frequency band k, train an independent spectrum encoder. ,in It is the set of trainable parameters in the encoder to extract frequency band-specific features. : .
5. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 4, characterized in that, Each Employing a separable time-frequency convolution design, it is used along the time axis. Convolution models short-term dynamics, using along the frequency axis. Convolution captures cross-frequency dependencies, followed by ReLU activation and pooling, resulting in... Frequency band specific characteristics as downstream components.
6. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 1, characterized in that, The task-aware frequency domain feature fusion module uses a salient frequency band attention fusion mechanism to achieve task-adaptive feature integration, given features of each frequency band. The contribution score for each frequency band is calculated using a lightweight MLP: The significance weights are obtained through Softmax normalization: ,in This reflects the significance of frequency band k for the current task; the most significant frequency band is selected: Features of the most prominent frequency band To guide the process, features from all frequency bands are fused using a multi-head cross-attention mechanism; The cross-attention mechanism can focus on complementary information in non-salient frequency bands, suppress redundancy and noise information, generate task-aware frequency domain feature representations, and finally obtain frequency domain features. .
7. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 1, characterized in that, The adaptive feature fusion module is responsible for integrating temporal and frequency domain features and completing the final cognitive state classification; after obtaining temporal features... and frequency domain features Then, attention-based time-frequency fusion is performed, the concatenated features are linearly projected into queries, keys and values, and self-attention processing is used to model cross-feature correlations, finally obtaining the fused feature representation Z.
8. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 7, characterized in that, The fused feature Z is passed through a fully connected layer and a softmax activation function to output the probability distribution of each cognitive state category. The cross-entropy loss is used as the objective function, and the specific formula is as follows: Where K represents the number of categories, It's a real label.
9. The cognitive state recognition system based on single-channel EEG signal task perception and multi-scale modeling as described in claim 1, characterized in that, The output module outputs cognitive state category labels, generates interpretability analysis information based on the salient frequency band attention mechanism, calculates the contribution score of each frequency band through a lightweight MLP, and obtains the corresponding saliency weight through Softmax. The higher the weight of the frequency band, the greater its contribution to the cognitive state discrimination of the current sample.