Multi-source electroencephalogram signal comprehensive processing method, system and device and storage medium

Abstract

The invention relates to the technical field of neuroscience and signal processing, and discloses a multi-source electroencephalogram signal comprehensive processing method, system and device and a storage medium, and the method comprises the following steps: S1, collecting multi-source electroencephalogram signals including scalp EEG, electrootogram EOG and electromyogram EMG signals; s2, carrying out preprocessing on the multi-source signal; s3, extracting a time-frequency feature, a spatial component feature and a nonlinear dynamic feature of the preprocessed signal; s4, inputting the extracted features into a machine learning model containing an attention mechanism; s5, training the model based on the diversity enhanced electroencephalogram signal data set; and S6, deploying the trained model to a real-time monitoring system. According to the method, through the multi-source electroencephalogram signal fusion analysis technology, the scalp EEG and the electrooculogram EOG are combined, the recognition precision and comprehensiveness of the characteristic brain wave mode are remarkably improved, and more reliable data support is provided for mental health assessment.

Description

Technical Field

[0001] The present invention relates to the field of neuroscience and signal processing technology, and in particular to a method, system, device and storage medium for comprehensive processing of multi-source electroencephalogram (EEG) signals. Background Art

[0002] With the development of neuroscience, electroencephalogram (EEG) plays an important role in the study of sleep disorders. Existing technologies can analyze brain activity by collecting EEG signals, which is of great significance for understanding the brain state of people with anxiety, tension, and sleep disorders.

[0003] Currently, common EEG pattern recognition usually relies on a single type of EEG signal, such as power spectral density analysis based on a specific frequency bandwidth. However, analysis methods that rely solely on a single type of signal often ignore the complex interactions of EEG characteristics in different brain regions, different frequency bands, and different physiological states, resulting in limited accuracy in characteristic EEG pattern recognition and difficulty in fully reflecting the individual's neural state. Summary of the Invention

[0004] In order to make up for the above shortcomings, the present invention provides a method, system, device and storage medium for comprehensive processing of multi-source EEG signals, aiming to improve the problem that common EEG pattern recognition relies only on the analysis method of a single type of signal and is difficult to fully reflect the individual's neural state.

[0005] In a first aspect, the present invention provides the following technical solution, a method for comprehensive processing of multi-source EEG signals, comprising the following steps: S1, collect multi-source EEG signals, including scalp EEG, electro-oculogram (EOG), and electromyogram (EMG) signals; S2. Preprocessing the multi-source signals, including filtering, denoising and baseline correction; S3. Extracting time-frequency features, spatial component features, and nonlinear dynamic features of the preprocessed signal, wherein the nonlinear dynamic features include at least one of approximate entropy, sample entropy, and fractal dimension; S4. Input the extracted features into a machine learning model that includes an attention mechanism, and perform fusion analysis by dynamically assigning feature weights to different brain regions and time points; S5. Training the model based on a diversity-enhanced EEG signal dataset, where the dataset covers signal samples from different geographic regions, ages, genders, and under simulated interference conditions; S6. Deploy the trained model to the real-time monitoring system, combine dynamic time window and sliding window technology to update the feature set, and realize the analysis and pattern recognition of continuous EEG signals.

[0006] By adopting the above technical solution, multi-source EEG signals such as scalp EEG, electro-oculogram (EOG), and electromyogram (EMG) are first collected, and then preprocessed with filtering, denoising, and baseline correction. Subsequently, the time-frequency, spatial components, and nonlinear dynamics of the preprocessed signals are extracted and input into a machine learning model that includes an attention mechanism. Fusion analysis is performed by dynamically assigning feature weights. The model is trained using a dataset covering samples from different geographic regions, ages, genders, and simulated interference conditions. Finally, the trained model is deployed to a real-time monitoring system. Dynamic time window and sliding window technologies are combined to update the feature set, enabling precise analysis and pattern recognition of continuous EEG signals, significantly improving recognition accuracy and the comprehensiveness of the reflection of individual neural states.

[0007] Preferably, the extraction of nonlinear dynamic features in S3 further includes: The complexity of the signal is quantified through multi-scale entropy analysis, and the dynamic stability of the signal is evaluated in combination with the Lyapunov exponent.

[0008] Preferably, the attention mechanism in S4 is specifically a multi-head attention layer, which is used to dynamically adjust the model's attention weights on signals from different brain regions and time segments.

[0009] Preferably, the diversity-enhanced data set in S5 further includes interference samples generated by superimposing noise and muscle activity simulation signals.

[0010] Preferably, the feature selection algorithm in S3 is recursive feature elimination, which is used to select the feature subset with the highest discrimination degree for the target EEG pattern from the time-frequency features, spatial component features and nonlinear dynamic features.

[0011] In a second aspect, the present invention provides the following technical solution: a multi-source EEG signal comprehensive processing system, the system comprising: Signal acquisition module, used to obtain multi-source EEG signals; Preprocessing module, used to filter, denoise and baseline correct the signal; a feature extraction module configured to extract time-frequency features, spatial component features, and nonlinear dynamic features; Model analysis module, including machine learning models optimized with attention mechanisms for multi-source feature fusion and pattern recognition; Dataset construction module, used to generate training datasets with enhanced diversity; The real-time monitoring module, combined with dynamic time window and sliding window technology, deploys the model to achieve continuous signal analysis.

[0012] Preferably, the dynamic time window in the real-time monitoring module is an adaptive adjustable window, and its window length is dynamically adjusted according to the non-stationarity of the signal, specifically including: evaluating the non-stationarity of the current time segment by calculating the entropy or variance of the signal in real time; if the non-stationarity exceeds a preset threshold, shortening the time window length to capture instantaneous changes; if the signal tends to be stable, extending the window length to improve the efficiency of feature calculation.

[0013] A multi-source EEG signal comprehensive processing device, comprising: Wearable sensor array for synchronous acquisition of scalp EEG, EOG, and EMG signals; An embedded processor configured to execute the steps of the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 5; Real-time feedback unit for outputting EEG pattern recognition results and warning information In the third aspect, the invention provides the following technical solution: a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a method for comprehensive processing of multi-source EEG signals when executing the computer program.

[0014] In a fourth aspect, the present invention provides the following technical solution: a readable storage medium having a computer program stored thereon, wherein the computer program implements the above-mentioned multi-source EEG signal comprehensive processing method when executed by a processor.

[0015] The present invention has the following beneficial effects: 1. The present invention uses multi-source EEG signal fusion analysis technology, combined with multi-dimensional signals such as scalp EEG, electro-oculogram (EOG) and electromyogram (EMG), to significantly improve the recognition accuracy and comprehensiveness of characteristic EEG patterns. At the same time, it can capture the complex interactions between different brain regions, frequency bands and physiological states, especially in scenarios such as anxiety, tension, and sleep disorders, and can fully reflect the dynamic characteristics of neural activity.

[0016] 2. In this invention, a multi-head attention mechanism is introduced to optimize model training, dynamically assigning feature weights to different brain regions and time points, greatly enhancing the model's sensitivity to key signals. By focusing on the importance of multi-dimensional feature combinations and time segments in parallel, the model can still accurately screen key information even in the presence of noise interference or high signal complexity.

[0017] 3. By constructing a diverse training dataset encompassing signal samples from diverse geographic regions, ages, genders, and simulated interference conditions, the present invention significantly improves the model's generalization and robustness. Incorporating data on genetic differences, environmental variables, and non-ideal physiological conditions enables the model to adapt to complex and changing real-world scenarios. The system can stably extract target EEG features even in the presence of electromyographic artifacts or environmental noise, ensuring the reliability and universality of clinical analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the multi-source EEG signal comprehensive processing method proposed by the present invention; Figure 2 This is a system architecture diagram of the multi-source EEG signal comprehensive processing system proposed by the present invention; Figure 3 This is a schematic diagram of the multi-source EEG signal comprehensive processing device proposed in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1 Reference Figure 1 In a first embodiment of the present invention, the present invention provides a method for comprehensive processing of multi-source EEG signals, comprising the following steps: S1, collect multi-source EEG signals, including scalp EEG, electro-oculogram (EOG), and electromyogram (EMG) signals; S2, preprocessing of multi-source signals, including filtering, denoising and baseline correction; S3. Extracting time-frequency features, spatial component features, and nonlinear dynamic features of the preprocessed signal, wherein the nonlinear dynamic features include at least one of approximate entropy, sample entropy, and fractal dimension; S4. Input the extracted features into a machine learning model that includes an attention mechanism, and perform fusion analysis by dynamically assigning feature weights to different brain regions and time points; S5. Training models based on a diversity-enhanced EEG signal dataset, which covers signal samples from different geographic regions, ages, genders, and simulated interference conditions; S6. Deploy the trained model to the real-time monitoring system, combine dynamic time window and sliding window technology to update the feature set, and realize the analysis and pattern recognition of continuous EEG signals.

[0021] Specifically, S1. Multi-source signal acquisition: First, a wearable sensor array synchronously collects multi-source EEG signals, including scalp electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG). These signals cover different brain regions and physiological states, and can fully reflect the complexity of brain activity.

[0022] S2. Signal preprocessing: The collected multi-source signals are preprocessed to improve the signal-to-noise ratio. Preprocessing includes adaptive filtering to remove environmental noise and electromyographic interference, baseline correction to eliminate signal drift, and bandpass filtering to retain the target frequency band (such as δ, θ, α, β, and γ waves).

[0023] S3. Multi-dimensional feature extraction: Three key features are extracted from the preprocessed signal: time-frequency features, spatial component features, and nonlinear dynamic features. The time-frequency features obtain the time-frequency distribution characteristics of the signal through continuous wavelet transform; the spatial component features use independent component analysis (ICA) to separate independent EEG signal components; nonlinear dynamic features include approximate entropy, sample entropy, fractal dimension and other indicators to quantify the complexity and dynamic stability of the signal. Combined with multi-scale entropy analysis and Lyapunov index evaluation, the dynamic behavior of the signal is further revealed. The feature subset with the highest discrimination for the target EEG pattern is screened out through the recursive feature elimination algorithm to reduce redundant information.

[0024] S4. Feature fusion driven by attention mechanism: The extracted multi-source features are input into a machine learning model containing a multi-head attention mechanism. By dynamically adjusting the weight distribution of signals in different brain regions and time segments, the model's sensitivity to key features is enhanced. The multi-head attention layer pays attention to different feature combinations in parallel, calculates the attention weight according to the importance of the signal at the time point, optimizes the input of the fully connected layer, and can still effectively screen key information in a complex signal environment, significantly improving the recognition accuracy of characteristic EEG patterns such as anxiety, tension, and sleep disorders. The machine learning model of the attention mechanism is a model that integrates multi-head attention layers. By dynamically allocating feature weights, it enhances sensitivity to key information and achieves efficient fusion of multi-source features and accurate EEG pattern recognition.

[0025] S5. Diversity-enhanced model training: The model is trained on a diversity-enhanced EEG dataset. This dataset covers samples from diverse geographic regions, age groups, and genders, and simulates signals under interference conditions such as noise and muscle activity. By introducing data on genetic differences, environmental variables, and non-ideal physiological conditions, the model's generalization and robustness are enhanced. Cross-validation and parameter optimization are combined during training to ensure that the model consistently outputs high-precision analysis results across diverse individuals and environments.

[0026] S6. Real-time monitoring and dynamic analysis: The trained model is deployed to the real-time monitoring system, and continuous signal analysis is achieved by combining dynamic time window and sliding window technology. The dynamic time window adaptively adjusts the window length according to the non-stationarity of the signal: the signal stability is evaluated by real-time calculation of entropy or variance. When the non-stationarity is high, the window is shortened to capture instantaneous changes. When the stability is high, the window is extended to improve computing efficiency. The sliding window technology continuously updates the feature set to ensure that the system can provide instant feedback on EEG pattern recognition results, assisting doctors in neurological status assessment and intervention.

[0027] The extraction of nonlinear dynamic features in S3 further includes: The complexity of the signal is quantified through multi-scale entropy analysis, and the dynamic stability of the signal is evaluated in combination with the Lyapunov exponent.

[0028] Specifically, in the nonlinear dynamic feature extraction of step S3, the complexity and dynamic stability of the EEG signal are comprehensively quantified through the collaborative analysis of multi-scale entropy and Lyapunov exponent. Multi-scale entropy is based on the multi-time scale characteristics of sample entropy, and the pre-processed signal is coarse-grained to generate sequences of different time scales and calculate the sample entropy value at each scale, thereby revealing the hierarchical changes in signal complexity. In the anxious state, the EEG signal may show a decrease in entropy value at a specific scale, indicating that the neural activity tends to be regular, while the normal state is characterized by a high entropy value, reflecting the increase in complexity.

[0029] The Lyapunov exponent uses phase space reconstruction technology to map one-dimensional time series to high-dimensional space, calculate the maximum Lyapunov exponent, and evaluate the chaotic characteristics of the signal: a positive exponent indicates low dynamic stability, while a negative exponent represents high stability. The combination of the two can analyze the nonlinear characteristics of EEG signals in multiple dimensions - multi-scale entropy characterizes complexity differences from the time dimension, while the Lyapunov exponent quantifies stability changes from the perspective of dynamic systems. In sleep monitoring, the slow-wave sleep stage may exhibit low complexity (low multi-scale entropy) and high stability, while the rapid eye movement sleep stage exhibits high complexity (high multi-scale entropy) and low stability (the exponent approaches zero or a positive value).

[0030] During the implementation process, by optimizing the scale range (such as 1 to 20 times the sampling interval) and adaptive window length selection, computational efficiency and accuracy were ensured. At the same time, preprocessing retained the 0.5-50Hz frequency band to avoid nonlinear information loss, which significantly improved the sensitivity and accuracy of EEG pattern recognition, providing more comprehensive dynamic feature support for mental health assessment, sleep disorder diagnosis and neurological disease warning.

[0031] The attention mechanism in S4 is specifically a multi-head attention layer, which is used to dynamically adjust the model's attention weights on signals from different brain regions and time segments.

[0032] Specifically, by introducing a multi-head attention layer, the model dynamically optimizes the attention weights for signals from different brain regions and time segments, thereby improving the recognition accuracy of characteristic EEG patterns. This layer consists of multiple attention heads working in parallel, each of which independently learns feature interactions in different dimensions. During model training, a dynamic weight allocation mechanism adaptively adjusts the weights based on the characteristics of the input signal: after mapping multi-source signals to query, key, and value spaces, similarity scores are calculated using dot-product attention and normalized to a probability distribution using the Softmax function. At the same time, the refined focus on time segments segments the continuous signal using a sliding window technique, with each attention head independently evaluating the importance of each segment.

[0033] The outputs of the attention heads are concatenated or weighted together before being fed into a downstream fully connected layer. The cross-entropy loss function and gradient descent algorithm are used to optimize parameters, gradually strengthening the weights of key features (such as high-frequency oscillations) and weakening interfering signals. Compared to traditional methods, the multi-head attention layer significantly improves the model's robustness to complex spatial and temporal heterogeneity through multi-dimensional feature interaction and dynamic weight allocation.

[0034] The diversity-enhanced dataset in S5 also includes interference samples generated by superimposing noise and muscle activity simulation signals.

[0035] Specifically, EEG signal samples from different geographical regions, age and gender groups are first included in the data collection phase to capture the impact of genetic differences, living habits and physiological characteristics on EEG activity. At the same time, artificially synthesized interference conditions are used to simulate non-ideal signals in real scenarios: for example, high-frequency environmental noise and low-frequency physiological noise are superimposed, and muscle activity simulation signals are introduced to generate mixed signal samples containing multiple interference types. In addition, a dynamic interference intensity gradient is used to gradually increase the noise level from mild to significant to train the model's adaptability under different levels of interference. To ensure data balance, the sample distribution is evenly divided by age, gender and interference type to avoid the model being biased towards specific groups or conditions. Through such diversity enhancement strategies, the model can learn to distinguish between real EEG features and interference signals during training.

[0036] The feature selection algorithm in S3 is recursive feature elimination, which is used to select the feature subset with the highest discrimination for the target EEG pattern from time-frequency features, spatial component features and nonlinear dynamic features.

[0037] Specifically, three types of features are first extracted from preprocessed multi-source EEG signals: time-frequency features are generated using a continuous wavelet transform to generate a time-frequency distribution matrix; spatial component features are separated using independent component analysis (ICA) to separate independent EEG signal components; and nonlinear dynamic features, including metrics such as approximate entropy, sample entropy, and fractal dimension, quantify the signal's complexity and dynamic behavior. All features are then normalized to eliminate the impact of dimensional differences on model evaluation.

[0038] Next, a classification model is initialized, an initial classifier is trained based on the full feature set, and an importance score is calculated for each feature. For example, the power spectral density of high-frequency beta waves in the time-frequency features may be assigned a higher weight, while certain nonlinear dynamic features may contribute significantly to distinguishing anxiety states. The algorithm gradually removes the lowest-ranking features based on their scores, retraining the model and evaluating its performance after each iteration until the predefined number of features is reached or the model performance stabilizes.

[0039] During this process, a recursive mechanism ensures the optimality of feature subsets. For example, in a sleep staging task, the initial stage may contain hundreds of features. After multiple rounds of iteration, key features are ultimately selected, such as the time-frequency energy of delta waves in the temporal lobe, the spatial component entropy of alpha waves in the occipital lobe, and the multi-scale entropy of prefrontal signals. These features can significantly distinguish slow-wave sleep from rapid eye movement (REM) sleep, while eliminating interfering features associated with electromyographic artifacts or environmental noise.

[0040] Example 2 Reference Figure 2 In a second embodiment of the present invention, the present invention provides a multi-source EEG signal comprehensive processing system, the system comprising: Signal acquisition module, used to obtain multi-source EEG signals; Preprocessing module, used to filter, denoise and baseline correct the signal; a feature extraction module configured to extract time-frequency features, spatial component features, and nonlinear dynamic features; Model analysis module, including machine learning models optimized with attention mechanisms for multi-source feature fusion and pattern recognition; Dataset construction module, used to generate training datasets with enhanced diversity; The real-time monitoring module combines dynamic time window and sliding window technology to deploy models for continuous signal analysis.

[0041] Specifically, the signal acquisition module: synchronously collects multi-source EEG signals through a wearable sensor array, including scalp electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG). The sensor array uses high-sensitivity electrodes and low-noise circuit design to ensure the synchronization and stability of signal acquisition. For example, EEG electrodes cover key brain areas such as the frontal lobe, parietal lobe, temporal lobe and occipital lobe, and EOG electrodes are placed around the eye sockets to capture eye movements.

[0042] Preprocessing module: The preprocessing module is responsible for filtering, denoising, and baseline correction of the raw signal. Specifically, it includes: Adaptive filtering: Using an adaptive filter based on the least mean square error (LMS) algorithm, it dynamically eliminates environmental noise (such as 50Hz power frequency interference) and electromyographic artifacts (such as EMG signals generated by blinking or frowning); Bandpass filtering: Preserves the target frequency band (0.5-50Hz), filters out ultra-low frequency drift and high-frequency noise, and ensures that the signal is focused on the characteristic EEG frequency band (such as δ, θ, α, β, and γ waves); Baseline correction: Using polynomial fitting or sliding average methods to eliminate signal baseline drift and improve the accuracy of time-frequency analysis.

[0043] The feature extraction module extracts three key features from the preprocessed signal: Time-frequency features: This module uses the continuous wavelet transform (CWT) to generate a time-frequency distribution matrix of the signal, quantifying the dynamic changes in energy in different frequency bands over time, such as the energy increase of alpha waves when eyes are closed. Spatial component features: This module uses independent component analysis (ICA) to separate independent components in the mixed signal, identify EEG activity characteristics of specific brain regions (such as the prefrontal or occipital lobes), and remove eye movement or myoelectric artifacts. Nonlinear dynamic features: This module calculates metrics such as approximate entropy, sample entropy, and fractal dimension to assess signal complexity. It also combines multiscale entropy analysis and the Lyapunov exponent to quantify dynamic stability. Recursive feature elimination (RFE) is used to select the most discriminative feature subset, such as the combination of prefrontal beta wave power and sample entropy during anxiety.

[0044] The model analysis module integrates a machine learning model optimized by the attention mechanism to achieve multi-source feature fusion and pattern recognition. The core components include: Multi-head attention layer: composed of multiple parallel attention heads, each head independently learns the feature associations of signals from different brain regions (such as the frontal lobe and parietal lobe) or time segments (such as high-frequency spikes before epileptic seizures), and dynamically assigns weights to enhance the sensitivity of key information.

[0045] Feature Fusion and Classification: The outputs of each attention head are concatenated and fed into a fully connected layer, where they are trained end-to-end using a cross-entropy loss function. For example, in a sleep staging task, the model fuses the time-frequency energy of occipital alpha waves with the nonlinear features of parietal delta waves to output a sleep stage classification result.

[0046] The dataset construction module generates high-quality training datasets through diversity enhancement strategies, including: multi-source data acquisition: covering EEG samples from different geographic regions, age groups (20-80 years old), and gender groups to capture the influence of genetic, environmental, and physiological differences on the signal; interference simulation: artificially superimposing high-frequency environmental noise (such as equipment electromagnetic interference), low-frequency physiological noise (such as skin impedance fluctuations), and muscle activity simulation signals (such as EMG artifacts produced by frowning) to generate mixed signals with varying interference intensities (mild to severe); data balancing: evenly distributing samples by age, gender, and interference type to prevent the model from being biased towards specific conditions. For example, the anxiety disorder dataset contains an equal number of male and female samples, and the proportion of data under different noise levels is balanced.

[0047] The real-time monitoring module combines dynamic time window and sliding window technologies to achieve continuous analysis and immediate feedback of EEG signals: Dynamic time window: Non-stationarity is assessed by calculating signal entropy or variance in real time, and the window length is adaptively adjusted—shortening the window (e.g., 200ms) to capture transient events when non-stationarity is high, and extending the window (e.g., 2s) to improve feature calculation efficiency when stationarity is high. Sliding window update: Sliding the time window with a 50% overlap rate continuously updates the feature set to ensure real-time analysis results. Early warning and feedback: When the model detects an abnormal pattern, the real-time feedback unit triggers an audible and visual alarm and generates a visual report for clinical reference.

[0048] The dynamic time window in the real-time monitoring module is an adaptive and adjustable window, and its window length is dynamically adjusted according to the non-stationarity of the signal. Specifically, it evaluates the non-stationarity of the current time segment by calculating the entropy or variance of the signal in real time; if the non-stationarity exceeds the preset threshold, the time window length is shortened to capture instantaneous changes; if the signal tends to be stable, the window length is extended to improve the efficiency of feature calculation.

[0049] Specifically, the system first quantifies the non-stationarity of the signal by calculating the entropy or variance of the current time segment in real time. The entropy value reflects the complexity and randomness of the signal—high entropy indicates a highly fluctuating signal with a large amount of information, while low entropy corresponds to stable or periodic activity. The variance directly measures the intensity of the fluctuation in the signal amplitude. A sudden increase in variance may indicate a transient event. The system presets a non-stationarity threshold. When the calculated value exceeds the threshold, the signal is determined to be non-stationary. At this time, the time window length is automatically shortened to capture transient changes with high temporal resolution. Conversely, if the signal tends to be stable, the window length is extended to improve the efficiency of feature extraction and model inference.

[0050] Dynamic time windows are combined with sliding window technology to ensure continuous and real-time signal analysis. The sliding window advances in fixed steps. After each window slide, the dynamic time window reassesses non-stationarity based on the latest signal fragment and adjusts the window length.

[0051] Example 3 Reference Figure 3 , a multi-source EEG signal comprehensive processing device, comprising: Wearable sensor array for synchronous acquisition of scalp EEG, EOG, and EMG signals; An embedded processor configured to execute the steps of the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 5; Real-time feedback unit, used to output EEG pattern recognition results and warning information.

[0052] Specifically, the wearable sensor array is used to synchronously collect multi-source EEG signals, including scalp electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG). EEG electrodes can cover all or part of key brain regions including the frontal lobe, parietal lobe, temporal lobe and occipital lobe, and can comprehensively capture EEG activity in different regions; EOG electrodes are deployed around the eye sockets to monitor artifacts generated by eye movements; EMG electrodes are placed on facial or neck muscle groups to identify muscle activity noise. The sensor uses high-sensitivity electrodes and low-noise circuit design to ensure the synchronization and stability of signal acquisition, providing a multi-dimensional data foundation for subsequent analysis. Through multi-channel synchronous acquisition, the device can comprehensively reflect the complex signal characteristics of brain activity, physiological interference and environmental noise.

[0053] The embedded processor executes all steps of the multi-source EEG signal processing method. The preprocessing stage uses adaptive filtering to remove power-frequency noise and myoelectric artifacts. Bandpass filtering preserves the target frequency band (0.5-50 Hz), and baseline correction eliminates signal drift. The feature extraction module extracts time-frequency features (such as the frequency band energy distribution generated by wavelet transform), spatial component features (such as brain region signals separated by independent component analysis), and nonlinear dynamic features (such as multiscale entropy and Lyapunov exponent) from the preprocessed signals. The model analysis module integrates a machine learning model optimized by the attention mechanism. Through a multi-head attention layer, it dynamically assigns weights to different brain regions and time segments, fuses multi-source features, and outputs pattern recognition results. The real-time monitoring module combines dynamic time windowing and sliding window technology to adaptively adjust the window length to balance transient event capture and computational efficiency, ensuring real-time continuous signal analysis.

[0054] The real-time feedback unit is used to output EEG pattern recognition results and warning information. When an abnormal pattern is detected, the unit triggers an immediate warning through an audible and visual alarm, continuously updates the feature set with a sliding window with a 50% overlap rate, and outputs the analysis results every 1 second.

[0055] Example 4 The fourth embodiment of the present invention is based on the same inventive concept and proposes a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the multi-source EEG signal comprehensive processing method of the above embodiment are implemented.

[0056] Example 5 The fifth embodiment of the present invention is based on the same inventive concept and proposes a terminal, which includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory and execute the multi-source EEG signal comprehensive processing method of the above embodiment.

[0057] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0058] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for comprehensive processing of multi-source EEG signals, characterized in that: The following steps are involved: S1, collect multi-source EEG signals, including scalp EEG, electro-oculogram (EOG), and electromyogram (EMG) signals; S2. Preprocessing the multi-source signals, including filtering, denoising and baseline correction; S3. Extracting time-frequency features, spatial component features, and nonlinear dynamic features of the preprocessed signal, wherein the nonlinear dynamic features include at least one of approximate entropy, sample entropy, and fractal dimension; S4. Input the extracted features into a machine learning model that includes an attention mechanism, and perform fusion analysis by dynamically assigning feature weights to different brain regions and time points; S5. Training the model based on a diversity-enhanced EEG signal dataset, where the dataset covers signal samples from different geographic regions, ages, genders, and under simulated interference conditions; S6. Deploy the trained model to the real-time monitoring system, combine dynamic time window and sliding window technology to update the feature set, and realize the analysis and pattern recognition of continuous EEG signals.

2. The multi-source EEG signal comprehensive processing method according to claim 1, characterized in that: The extraction of nonlinear dynamic features in S3 further includes: The complexity of the signal is quantified through multi-scale entropy analysis, and the dynamic stability of the signal is evaluated in combination with the Lyapunov exponent.

3. The multi-source EEG signal comprehensive processing method according to claim 1, characterized in that: The attention mechanism in S4 is specifically a multi-head attention layer, which is used to dynamically adjust the model's attention weights on signals from different brain regions and time segments.

4. The multi-source EEG signal comprehensive processing method according to claim 1, characterized in that: The diversity-enhanced dataset in S5 also includes interference samples generated by superimposing noise and muscle activity simulation signals.

5. The multi-source EEG signal comprehensive processing method according to claim 1, characterized in that: The feature selection algorithm in S3 is recursive feature elimination, which is used to select the feature subset with the highest discrimination degree for the target EEG pattern from the time-frequency features, spatial component features and nonlinear dynamic features.

6. A multi-source EEG signal comprehensive processing system, according to the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 5, characterized in that: The system comprises: Signal acquisition module, used to obtain multi-source EEG signals; Preprocessing module, used to filter, denoise and baseline correct the signal; a feature extraction module configured to extract time-frequency features, spatial component features, and nonlinear dynamic features; Model analysis module, including machine learning models optimized with attention mechanisms for multi-source feature fusion and pattern recognition; Dataset construction module, used to generate training datasets with enhanced diversity; The real-time monitoring module, combined with dynamic time window and sliding window technology, deploys the model to achieve continuous signal analysis.

7. The multi-source EEG signal comprehensive processing system according to claim 6, characterized in that: The dynamic time window in the real-time monitoring module is an adaptively adjustable window, the window length of which is dynamically adjusted according to the non-stationary degree of the signal, specifically including: By calculating the entropy or variance of the signal in real time, the non-stationarity of the current time segment is evaluated; if the non-stationarity exceeds the preset threshold, the time window length is shortened to capture instantaneous changes; if the signal tends to be stable, the window length is extended to improve the efficiency of feature calculation.

8. A multi-source EEG signal comprehensive processing device, characterized in that: include: Wearable sensor array for synchronous acquisition of scalp EEG, EOG, and EMG signals; An embedded processor configured to execute the steps of the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 5; Real-time feedback unit, used to output EEG pattern recognition results and warning information.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 4 is implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-source EEG signal comprehensive processing method according to any one of claims 1 to 4 is implemented.

Images