Autonomous blink recognition method and system based on time-frequency feature fusion of single-channel electroencephalogram signal

Abstract

The application belongs to the field of brain-computer interface and electroencephalogram technology, and relates to an autonomous blinking recognition method and system based on time-frequency feature fusion of single-channel electroencephalogram signals. The method comprises the following steps: acquiring single-channel EEG signals and performing data preprocessing; determining and segmenting the blinking action in the preprocessed EEG signals by using a sliding window-based blinking interval detection algorithm; performing autonomous blinking feature extraction and optimization on the EEG signal interval determined as the blinking action in the time-frequency domain; and inputting the extracted features into a pre-trained MLP classifier to realize accurate recognition of autonomous blinking. The application utilizes the portability advantage of single-channel EEG devices, realizes accurate recognition of autonomous blinking, lays a key technical foundation for constructing an interactive system with low false detection rate and high response speed, and is especially suitable for scenes such as barrier-free interaction and virtual reality control that have strict requirements on intention recognition reliability.

Description

Technical Field

[0001] This invention belongs to the fields of brain-computer interface and electroencephalography (EEG) technology, and specifically relates to an autonomous blinking recognition method and system based on time-frequency feature fusion of single-channel EEG signals. Background Technology

[0002] The need for lightweight and practical applications of Brain-Computer Interface (BCI) technology has driven research into intent decoding based on single-channel electroencephalograms (EEGs). Involuntary blinking, as a conscious behavior controlled by the cerebral cortex, has significant application value in neurorehabilitation and human-computer interaction. Blinking actions encompass various types of active and involuntary movements, including voluntary blinking (such as single or double blinks triggered by the user to execute interactive commands), conditioned reflex blinking (such as protective blinking in the face of strong light or when an object approaches), and spontaneous blinking (such as blinking to maintain eye moisture and comfort). These movements are characterized by high frequency, high dynamism, and the coexistence of reflexes and autonomy. Accurately identifying blinking actions carrying the interface interaction intent from complex eye movements is a core challenge in building an efficient and reliable BCI system.

[0003] However, existing research focuses on recognizing whether blinking occurs. While it can detect blinking behavior, it cannot distinguish between conscious, involuntary blinking (such as in interactive commands) and unconscious, spontaneous or reflexive blinking. This leads to ambiguity in intent when blinking signals are used as interactive input channels. Furthermore, existing methods for recognizing intentional blinking suffer from poor interference resistance.

[0004] Interactive coding based on blinking as input relies on detecting voluntary and involuntary blinking states through electroencephalogram (EEG) signals. Existing methods use single-channel EEG signals to reduce system hardware costs and lower the barrier to entry, but this also requires addressing the issues of feature homogeneity and decreased classification accuracy due to the reduced number of channels.

[0005] Disadvantages of existing technology:

[0006] (1) Multi-channel EEG devices are expensive, complicated to wear, and have low practicality.

[0007] (2) Existing blink detection methods cannot distinguish between voluntary and involuntary blinks, resulting in ambiguity in interaction intent.

[0008] (3) Existing intentional blinking recognition schemes, such as “Intentional blinking recognition method and device based on single-channel EEG signal” (application number 201911242401.7, publication number CN111046779A), recognize spontaneous blinking by simple amplitude threshold comparison. The experimental sample is small (90 samples) and no relevant timeliness analysis is performed. It has the problems of being unable to resist interference from other eye movements and low reliability of experimental results.

[0009] Therefore, this invention conducts a deeper study on autonomous blink recognition based on single-channel EEG signals. Summary of the Invention

[0010] This invention aims to provide a method and system for autonomous blinking interaction driven by a single-channel EEG signal (EEG-Blink). This method not only leverages the portability of single-channel EEG devices to design a lightweight intent recognition scheme for BCI systems, achieving accurate recognition of autonomous blinking, but also lays a key technological foundation for building an interactive system with low false detection rate and high response speed. It is particularly suitable for scenarios with stringent requirements for the reliability of intent recognition, such as barrier-free interaction and virtual reality control.

[0011] The technical solution adopted in this invention is as follows:

[0012] A method for autonomous blink recognition based on time-frequency feature fusion of single-channel EEG signals includes the following steps:

[0013] Acquire single-channel EEG signals and perform data preprocessing;

[0014] A blinking interval detection algorithm based on a sliding window is used to determine and segment blinking actions in preprocessed EEG signals;

[0015] Autonomous blinking feature extraction and optimization are performed in the time-frequency domain for EEG signal intervals identified as blinking actions;

[0016] The extracted features are input into a pre-trained MLP classifier to achieve accurate recognition of autonomous blinking.

[0017] Furthermore, the single-channel EEG signal is acquired using a portable single-channel EEG device equipped with TGAM; the data preprocessing includes power frequency interference suppression and bandwidth optimization.

[0018] Furthermore, the blink interval detection algorithm based on a sliding window includes:

[0019] The blink threshold is determined through adaptive calculation;

[0020] A sliding window is used to detect EEG signals in real time. By jointly judging whether the peak-to-peak value in the time domain exceeds the blink threshold and whether the frequency domain energy meets the verification conditions, the effective blink interval is accurately identified and saved.

[0021] Furthermore, the determination of the blink threshold through adaptive calculation includes:

[0022] Baseline windows of n blink events are randomly selected from the original signal, and the peak-to-peak value D = {d1, d2, ..., dn} is calculated for each window. n};

[0023] The blink threshold D was determined based on Tukey's Fence criterion. th :

[0024] D th =Q1-k·(Q3-Q1)

[0025] Where Q3 is the upper quartile of D, Q1 is the lower quartile, and k is the sensitivity factor.

[0026] Furthermore, the method of using a sliding window to detect EEG signals in real time, and by jointly judging whether the peak-to-peak value in the time domain exceeds the blink threshold and whether the frequency domain energy meets the verification conditions, accurately identifies and saves valid blink intervals, including:

[0027] The signal is traversed using a sliding window, and the peak-to-peak value d within the current window is extracted. If d > D th If so, the signal within the current window is marked as a candidate blink event;

[0028] A frequency domain energy verification mechanism is introduced. Based on the characteristic law that the energy of high-frequency components in blink events exhibits significant attenuation, the low-frequency energy is constrained to be higher than the high-frequency energy. Through dual verification of amplitude threshold and frequency domain energy, blink signals can be accurately detected in single-channel EEG signals.

[0029] An event boundary compensation mechanism is introduced, which sets the compensation windows for the preceding and following delays based on the sampling rate to ensure the integrity of the blinking event.

[0030] Furthermore, the extraction and optimization of autonomous blinking features in the time-frequency domain for the EEG signal intervals identified as blinking actions includes:

[0031] The EEG data was decomposed using a 5-band segmentation method to obtain datasets of five different frequency bands, corresponding to Delta, Theta, Alpha, Beta and Gamma bands respectively.

[0032] The time-domain and frequency-domain features of the EEG signals were extracted from the five frequency bands, and the cross-band ratio features of Delta and Theta bands were added in addition.

[0033] The extracted features are standardized, and feature importance is evaluated using Radio Frequency (RF) to select highly discriminative features. At the same time, principal components with a cumulative variance contribution rate reaching a set threshold are extracted using PCA.

[0034] After normalizing the PCA variance contribution and RF importance, a weighted average score is calculated, and a feature subset is selected based on the average score.

[0035] Furthermore, the time-domain features include: average value, maximum value, minimum value, peak-to-peak value, variance, and zero-crossing rate; the frequency-domain features include: Psd_mean: the average value of the power spectral density over the entire frequency band; Psd_max: the maximum value of the power spectral density over the entire frequency band; f_max: the frequency corresponding to when the power spectral density reaches its maximum value; and the differential entropy feature, used to measure the complexity or uncertainty of the distribution of a continuous random variable.

[0036] An autonomous blink recognition system based on time-frequency feature fusion of single-channel EEG signals, characterized in that it includes:

[0037] The EEG signal acquisition and preprocessing module is used to acquire single-channel EEG signals and perform data preprocessing.

[0038] The blinking action detection module is used to determine and segment blinking actions in the preprocessed EEG signal using a blinking interval detection algorithm based on a sliding window.

[0039] The autonomous blinking feature extraction and optimization module is used to extract and optimize autonomous blinking features in the time-frequency domain for EEG signal intervals identified as blinking actions.

[0040] The autonomous blinking recognition module is used to input the extracted features into a pre-trained MLP classifier to achieve accurate recognition of autonomous blinking.

[0041] The beneficial effects of this invention are as follows:

[0042] This invention proposes a portable and efficient blink interaction method (EEG-Blink) based on single-channel EEG signals, reconstructing blink artifacts into a user-initiated interactive command carrier. This invention utilizes the TGAM EEG module to construct a blink dataset and employs a sliding window-based blink interval detection algorithm, achieving a 100% blink detection rate.

[0043] Meanwhile, the RF-PCA method employed in this invention optimizes the time-frequency domain features of single-channel EEG signals, significantly improving the generalization ability of each model. Based on an individualized scheme with adaptive parameter tuning, an average spontaneous blink recognition rate of 95.44% was achieved on the improved MLP classifier. In terms of real-time performance, EEG-Blink achieved an average latency of 20.28ms for single blink recognition.

[0044] Finally, this invention provides a scalable application framework for brain-computer interface systems. By combining autonomous blinking with motor control, it enhances the naturalness and convenience of human-computer interaction, provides important assistive tools for people with disabilities, and enhances the immersive experience of technologies such as virtual reality and augmented reality, thereby promoting the development of eye-tracking interaction towards humanization and efficiency. Attached Figure Description

[0045] A more complete understanding of the invention and its accompanying advantages and features will be more readily apparent from the accompanying drawings and the following detailed description, wherein:

[0046] Figure 1 This is a schematic diagram of a single-channel EEG signal-driven autonomous blinking interaction method (EEG-Blink) in this invention.

[0047] Figure 2 The waveform comparison diagram after the average superposition of the two blink signals provides a reliable basis for the implementation of EEG-Blink;

[0048] Figure 3 This is a diagram showing the wearing method of the single-channel EEG acquisition device used in this invention;

[0049] Figure 4 This is a superimposed power spectral density graph of the blink signal in one experiment. The energy of the blink signal decreases with increasing frequency. A frequency domain energy verification mechanism is set up using 30Hz as the dividing line.

[0050] Figure 5 As an example of blink interval detection, the blink interval detection algorithm built into this invention successfully detected 8 blink events in one test.

[0051] Figure 6 This describes the process of acquiring single-channel EEG blink signals in this invention.

[0052] Figure 7 This is a network architecture diagram of the multilayer perceptron (MLP) used in this invention;

[0053] Figure 8 The average processing time for blink recognition on each subject was 9.22 ms, and the optimized EEG-Blink reduced the average processing time by 9.22 ms. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0055] The main contents of this invention include:

[0056] (1) Using a Mindband portable device equipped with a brainwave sensor TGAM (ThinkGear ASIC Module), the present invention conducted neurophysiological experiments to collect brainwave signals in response to generated visual stimuli, thereby constructing a single-channel blink dataset.

[0057] (2) In order to achieve efficient detection of blinking events, this invention proposes a double-verification blinking interval detection algorithm based on a sliding window. The algorithm first determines the blinking threshold through adaptive calculation, and then uses a sliding window to detect the EEG signal in real time. By jointly judging whether the peak-to-peak value in the time domain exceeds the threshold and whether the energy in the frequency domain meets the verification conditions, the algorithm can accurately identify and save the valid blinking interval.

[0058] (3) To address the problems of model overfitting and low computational efficiency caused by the high-dimensional features of EEG signals, this invention proposes RF-PCA, a two-stage feature optimization framework that integrates Random Forest (RF) and Principal Component Analysis (PCA). By using RF to calculate feature importance and PCA contribution rate statistics, combined with the time-frequency domain features of single-channel EEG signals, it provides highly discriminative features for the recognition of spontaneous blinking.

[0059] Involuntary and involuntary blinking exhibit different signal characteristics on electroencephalograms (EEGs). Figure 2 The average superposition comparison results of two blink signals under the same subject were presented. It was found that the spontaneous blink signal has a significantly higher amplitude and a longer peak interval. This characteristic difference provides a reliable basis for spontaneous blink recognition based on single-channel EEG signals. Based on this, the present invention provides a novel blink interaction method, EEG-Blink, such as... Figure 1 As shown, the steps include:

[0060] S1. Acquire single-channel EEG signals and perform data preprocessing;

[0061] S2. Using a blink interval detection algorithm based on a sliding window, blinking actions in the preprocessed EEG signal are determined and segmented;

[0062] S3. Perform autonomous blink feature extraction and optimization in the time-frequency domain for the EEG signal intervals identified as blinking actions;

[0063] S4. Input the extracted features into the pre-trained MLP classifier to achieve accurate recognition of autonomous blinking.

[0064] In step S1, a portable single-channel EEG device equipped with TGAM is used. This device is compatible with both dry and wet electrodes and uses the Fp1 lead in the frontal lobe, positioned according to the international 10-20 system standard, to acquire EEG signals. The wearing method is described in [link to instructions]. Figure 3 Meanwhile, the TGAM module uses Bluetooth 4.0 low-power transmission protocol (transmission rate 57600bps) to transmit data to the PC in real time for recording and analysis.

[0065] After data acquisition, data preprocessing was performed to address environmental noise interference and signal passband selection issues.

[0066] (1) Power frequency interference suppression: The original EEG signal is susceptible to environmental noise interference, especially 50Hz mains power frequency noise. Therefore, a zero-phase notch filter is used to suppress it.

[0067] (2) Frequency band optimization: In order to extract the characteristic frequency bands related to blinking activity, a finite-length unit impulse response (FIR) filter is used to filter the signal, and the signal passband range is set to 0.3–70Hz.

[0068] In step S2, this invention proposes a dual-verification blink interval detection algorithm based on a sliding window. The algorithm first determines the blink threshold through adaptive calculation, and then uses a sliding window to detect the EEG signal in real time. By jointly judging whether the peak-to-peak value in the time domain exceeds the threshold and whether the energy in the frequency domain meets the verification conditions, the algorithm can accurately identify and save the valid blink interval.

[0069] Blinking events in single-channel EEG signals typically manifest as transient high-amplitude fluctuations, with typical characteristics including: (1) the EEG waveform exhibits a biphasic (positive-negative or negative-positive) peak morphology, lasting approximately 500–800 ms; (2) the peak-to-peak value (PP) is significantly higher than the background EEG activity, with amplitudes reaching 30–80 μV under this device. Analysis of the power spectral density superposition plot of the blinking signal in the experiment, such as... Figure 4 As shown, the energy of the blink signal is concentrated in the lower frequency band (<30Hz), and the energy of the high frequency components decreases significantly, which provides support for the identification of blink events in the frequency domain.

[0070] To adapt to individual differences and fluctuations in environmental noise, this invention proposes an adaptive blink threshold calculation method, the core steps of which are as follows:

[0071] Baseline windows of n blink events are randomly selected from the original signal, and the peak-to-peak value D = {d1, d2, ..., dn} is calculated for each window. n Based on Tukey's Fence criterion, robust statistics are used to determine the threshold D. th :

[0072] D th =Q1-k·(Q3-Q1) (1)

[0073] Where Q3 is the upper quartile of D, Q1 is the lower quartile, and k is the sensitivity factor with a value of 1.5. This formula (1) is used to determine the lower threshold of the blinking action.

[0074] During the blink event detection phase, a sliding window is used to traverse the signal and extract the peak-to-peak value d of the signal within the current window. If d > D th If the signal in the current window is marked as a candidate blink event, see Table 1 for window parameters.

[0075] Table 1. Sliding window parameter settings

[0076] parameter Value Window size 150ms sliding step 50ms Overlap rate 66.67％

[0077] To reduce the false detection rate, a frequency domain energy verification mechanism was introduced. Based on the significant attenuation of high-frequency components in blink events, the low-frequency (<30Hz) energy was constrained to be higher than the high-frequency (>30Hz) energy. Through dual verification of amplitude threshold and frequency domain energy, blink signals can be accurately detected in single-channel EEG signals. Specifically, an event boundary compensation mechanism was introduced, setting a compensation window of 0.2s before and 0.8s after the sampling rate to ensure the integrity of the blink event. The blink interval detection results of a single experiment are shown in [link to experimental data]. Figure 5 .

[0078] In step S3, a multi-dimensional feature extraction method is used to combine the time-domain features, frequency-domain features, and cross-band ratio features of the single-channel EEG signal to comprehensively characterize the dynamic characteristics of blink artifacts.

[0079] (1) Temporal characteristics

[0080] Time-domain features are extracted directly from the original signal, reflecting the instantaneous waveform characteristics of the signal. Based on... Figure 2 Based on the analysis, this invention extracts multiple statistical features of EEG signals, including average value, maximum value, minimum value, peak-to-peak value, variance, and zero-crossing rate (ZCR). These features together constitute a multi-dimensional statistical index set describing the waveform characteristics of EEG signals.

[0081] (2) Frequency domain characteristics

[0082] Frequency domain analysis reveals the frequency components and energy distribution characteristics of a signal by transforming it from the time domain to the frequency domain.

[0083] In this invention, the Welch method is used to estimate the power spectral density (PSD), and the following frequency domain features are extracted:

[0084] Psd_mean: The average power spectral density over the entire frequency band;

[0085] Psd_max: The maximum value of the power spectral density across the entire frequency band;

[0086] f_max: The frequency at which the power spectral density reaches its maximum value.

[0087] Meanwhile, the differential entropy feature is used to measure the complexity or uncertainty of the distribution of continuous random variables, and its expression is as follows:

[0088]

[0089] Where h(X) is the calculated differential entropy, X is a random variable, and f(x) is the probability density function of X. First, the EEG signal is divided into five frequency bands: Delta, Theta, Alpha, Beta, and Gamma (see Table 2). Since they statistically approximate a Gaussian distribution, the differential entropy formula for each frequency band can be simplified as follows:

[0090]

[0091] Among them, H DE Let σ be the differential entropy, e be the natural constant, and σ be the differential entropy. 2 Let X be the variance.

[0092] (3) Cross-band ratio characteristics

[0093] To characterize the energy interaction characteristics of EEG signals across different frequency bands, the cross-band ratio feature is constructed by quotienting the features of Delta and Theta bands.

[0094] To address the issues of model overfitting and low computational efficiency caused by the high-dimensional features of EEG signals, this invention proposes RF-PCA, a two-stage feature optimization framework that integrates random forest and principal component analysis. The specific process is as follows:

[0095] First, the preprocessed EEG data was decomposed using a 5-band segmentation method to obtain datasets in five different frequency bands, corresponding to Delta (δ), Theta (θ), Alpha (α), Beta (β), and Gamma (γ) bands, respectively. The specific frequency band divisions are shown in Table 2.

[0096] Table 2. Five-band division table

[0097] brainwaves Frequency range Delta(δ) 0.5-4Hz Theta(θ) 4-8Hz Alpha(α) 8-13Hz Beta(β) 13-30Hz Gamma(γ) 30-64Hz

[0098] The time-frequency domain features of the EEG signals were extracted from the five frequency bands, and the ratio features of the Delta and Theta bands were added in addition, forming a total of 10×5+10=60-dimensional features.

[0099] Then, in the feature selection stage, the 60-dimensional features are first standardized, and feature importance is evaluated using Randomized Framework (RF) (supervised stage) to screen out highly discriminative features. Simultaneously, principal components with a cumulative variance contribution rate reaching a set threshold (preferably 95% in this method) are extracted using PCA (unsupervised stage). The calculation formula is as follows:

[0100]

[0101] Where n is the number of principal components to be retained, r is the number of principal components currently being tried, λ1, λ2, λ3, ..., λ d Let be the eigenvalues ​​of the covariance matrix, and d be the original feature dimension. Projection matrix. The data is mapped to a low-dimensional space, and the sum of the absolute weights of each original feature in all principal components is calculated, i.e., the PCA variance contribution (Var). j :

[0102]

[0103] Among them, w ij For (projection matrix) (the element in the i-th row and j-th column).

[0104] Finally, the variance contribution of PCA (Var) is calculated. j Importance of RF j After normalization and weighting, the overall score of the features is calculated. j Here, the weight α = 0.82. The feature subset is selected based on the comprehensive score. This method selects the feature subset of the top 1 / 3 of the comprehensive scores.

[0105] Score j =α·Importance j +(1-α)·Var j (6)

[0106] This method enhances feature discriminative power through RF and reduces dimensionality redundancy through PCA, constructing a lightweight feature engineering framework for a real-time blink detection system for single-channel EEG signals. This significantly improves computational efficiency and model generalization ability, which will be validated in the next stage.

[0107] In step S4, structured experiments are used to evaluate the effectiveness of EEG-Blink in recognizing spontaneous blinks. Initially, personalized model tuning is explored through adaptive threshold calculation. Subsequently, a timeliness analysis of EEG-Blink is performed to evaluate the feasibility of its real-time interactive applications.

[0108] Dataset: Experimental data were collected using a single-channel EEG device equipped with TGAM. The subjects were five university students (three males and two females), with an average age of 22, all in good health and right-handed. During data collection, simple, repetitive visual stimuli were used, providing participants with a visual interface and prompting them every 4 seconds. Data collection for each participant consisted of two phases: an involuntary blinking phase without motor intention, and a voluntary blinking phase simulating a mouse click, where participants were asked to place their right hand on the mouse and imagine clicking. Data was collected between the voluntary and involuntary blinking phases with a 1-minute rest in between, for a total of 40 trials. For details on the single-channel EEG signal acquisition paradigm, see [link to documentation]. Figure 6 .

[0109] After data preprocessing, the blink interval detection algorithm of this invention was applied, collecting 800 sets of voluntary blinking data and 800 sets of involuntary blinking data, achieving a 100% detection success rate. To comprehensively evaluate model performance, a hierarchical 5-fold cross-validation strategy was adopted, randomly dividing the dataset into 5 equal folds. For each fold of data, different classification models were trained, and the trained models were tested on the test set, with performance metrics calculated and saved. Finally, the most suitable model was selected based on the experimental results, and its parameters were further optimized to improve classification accuracy.

[0110] In practical applications, due to differences in physiological factors and behavioral patterns among individuals, this experiment calculated an adaptive blink threshold for each participant. This threshold was then applied to a blink interval detection algorithm to obtain the participant-specific EEG blink signal interval. Based on this, a 5-fold cross-validation experiment was designed using three classifiers: Random Forest (RF), Support Vector Machine (SVM), and Multi-layer Perceptron (MLP).

[0111] Random forest is an algorithm based on ensemble learning. Its core idea is to use the Bagging algorithm to combine multiple decision trees into a powerful predictive model. When building a random forest, each decision tree is trained independently based on a random subset of the original dataset. This randomness helps reduce the risk of overfitting the model to specific data. Random forest uses statistical voting to determine the final prediction result; that is, it aggregates the classification results of each decision tree and selects the category with the most votes as the model's final prediction output. For specific parameter settings of random forest, see Table 3; other parameters use default values.

[0112] Table 3. Random Forest Model Parameter Settings

[0113] Parameter name Parameter meaning Parameter settings n_estimators Number of weak classifiers 145 criterion Evaluation criteria for subtree partitioning Entropy max_depth Maximum depth, related to pruning 18

[0114] Support Vector Machines (SVMs) are a widely used machine learning algorithm in classification and regression tasks. The basic idea is to find a hyperplane or decision boundary that maximizes the margin between categorical data points; that is, to find an optimal separating plane between data points of different classes to effectively separate them. For specific SVM parameter settings, see Table 4; other parameters remain at their default values.

[0115] Table 4. Support Vector Machine Model Parameter Settings

[0116] Parameter name Parameter meaning Parameter settings kernel Kernel function type rbf C Regularization parameters 10 gamma Kernel function built-in coefficients 0.1

[0117] A Multilayer Perceptron (MLP) is a deep learning model consisting of multiple layers of neurons, where each neuron in a layer is fully connected to the layer preceding it. The MLP neural network structure includes an input layer, two hidden layers, and an output layer, and its general model is as follows: Figure 7 In the 0th layer (input layer), x1, x2 to xn are vectorized into X (n represents the number of features); between the 0th layer and the 1st layer (hidden layer 1), there are weights w1, w2...wm (m is related to the number of neurons set), which are vectorized into W[1], where [1] represents the weight of the 1st layer, and the bias b is similar;

[0118] Therefore, for the first layer, the calculation formula is:

[0119] Z [1] =W [1] ·X+b [1] (7)

[0120] A [1] =relu(Z) [1] (8)

[0121] Where Z is a linear combination of input values, and A is the value of Z through the ReLU activation function. For the input value of the first layer, X, the output value is A, which is also the input value of the next layer. The calculation process is similar to that between the 0th and 1st layers. Finally, we get y (the output value of this neural network).

[0122] The input layer is typically given by the shape of the input data and is not trained. The layers that need to be trained are the hidden and output layers. Each hidden layer uses ReLU as its activation function. The ReLU function, or Rectified LinearUnit, is a commonly used non-linear activation function in neural network structures. Its form is as follows:

[0123] f(x) = max(0, x) (9)

[0124] Meanwhile, this network uses CrossEntropyLoss as its loss function and Adam as its optimizer. Based on this, an improved MLP neural network model is proposed. By employing a regularization strategy and adding BatchNorm1d and Dropout layers, the model's complexity is controlled, improving performance and preventing overfitting. For specific MLP parameter settings, please refer to Table 5; other parameters are default values.

[0125] Table 5. Parameter settings for the multilayer perceptron model

[0126] Parameter name Parameter meaning Parameter settings solver Optimizer Adam lr Learning rate le-3 alpha Regularization parameters le-5 hidden_layer_sizes Hidden layer settings (128,64) activation Activation function relu dropout Proportion of randomly deleted neurons 0.3

[0127] Meanwhile, the recognition effect of the model was measured by accuracy and standard deviation, aiming to build a blink recognition model adapted to individual-specific signals. Table 6 shows the experimental results.

[0128] Table 6. Classification accuracy of each participant's data on RF, SVM, and MLP models (unit: 100%)

[0129]

[0130] Based on the experimental results in Table 6, it can be found that the MLP model achieved the most optimal values, with an average accuracy of 95.44%, which exceeds the 93.33% of the existing intentional blink recognition scheme (CN111046779A), proving the effectiveness of the EEG-Blink method of this invention.

[0131] The time latency of real-time EEG signal processing is a core indicator for evaluating the feasibility of a system. The EEG-Blink proposed in this invention achieves low-latency, high-precision real-time processing performance through a two-stage optimization process: sliding window dynamic threshold detection and autonomous blink recognition. (Window length: 150ms, step size: 50ms)

[0132] Blink data from each participant were collected and tested separately. The average, standard deviation, and maximum processing time (the time difference from the start of blinking to the completion of blink recognition) were calculated, and the performance difference before and after feature optimization was compared. Based on the best-performing MLP model in the experiment, EEG-Blink was constructed, and the results are shown in Table 7.

[0133] Table 7. Processing time for single blink recognition before and after EEG-Blink optimization (unit: ms)

[0134] average value Standard deviation Maximum value Before improvement 29.50 0.64 30.41 Improved 20.28 0.18 20.40

[0135] The results show that after RF-PCA feature optimization, the average processing time of EEG-Blink decreased from 29.50 ms (±0.64) to 20.28 ms (±0.18), a reduction of 31.25%, and the standard deviation decreased by 71.9%, demonstrating a significant improvement in algorithm stability. Detailed statistics can be found in [link to statistics]. Figure 8 .

[0136] The EEG-Blink method, utilizing only a single-channel EEG signal, through data preprocessing, blink interval detection, feature extraction and optimization, and MLP classification, can accurately and efficiently identify users' voluntary and involuntary blink signals, thus enabling its application in blink-driven human-computer interaction.

[0137] Key points of this invention:

[0138] 1) An adaptive blink threshold calculation method based on robust statistics. Addressing the variability in individual EEG signals, this invention proposes an adaptive blink threshold calculation method based on robust statistics, establishing a personalized blink recognition model and providing a generalizable solution for individualized BCI adaptation.

[0139] 2) Feature extraction method for blink signals. To address the issue of limited features in single-channel EEG signals, this invention employs a multi-dimensional feature extraction method, combining the time-domain features, frequency-domain features, and cross-frequency ratio features of single-channel EEG signals to comprehensively characterize the dynamic properties of blink artifacts.

[0140] 3) RF-PCA-based time-frequency feature optimization strategy. This invention designs a feature selection method (RF-PCA) that extracts transient features of blink signals through time-frequency analysis and utilizes a PCA dimensionality reduction pipeline guided by random forest to enhance feature discriminativeness and effectively reduce redundant information interference.

[0141] 4) Application based on an improved Multilayer Perceptron (MLP) model. To further improve model performance and effectively prevent overfitting, this invention proposes an improved MLP neural network model. Its main feature is the introduction of regularization strategies, adding BatchNorm1d layers and Dropout layers after certain layers of the network. These regularization techniques effectively control the model's complexity, reduce its excessive dependence on training data, and thus improve the model's generalization ability on unknown data.

[0142] Another embodiment of the present invention provides an autonomous blink recognition system based on time-frequency feature fusion of single-channel electroencephalogram (EEG) signals, comprising:

[0143] The EEG signal acquisition and preprocessing module is used to acquire single-channel EEG signals and perform data preprocessing.

[0144] The blinking action detection module is used to determine and segment blinking actions in the preprocessed EEG signal using a blinking interval detection algorithm based on a sliding window.

[0145] The autonomous blinking feature extraction and optimization module is used to extract and optimize autonomous blinking features in the time-frequency domain for EEG signal intervals identified as blinking actions.

[0146] The autonomous blinking recognition module is used to input the extracted features into a pre-trained MLP classifier to achieve accurate recognition of autonomous blinking.

[0147] The above division of modules is merely illustrative. In practical applications, the functions described above can be assigned to different functional modules as needed to complete all or part of the functions described in the aforementioned method. The specific working process of each module can be found in the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0148] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing steps of the method of the present invention.

[0149] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) storing a computer program that, when executed by a computer, implements the steps of the method of the present invention.

[0150] Another embodiment of the present invention provides a computer program product, the computer program product including a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An autonomous blink recognition method based on single-channel electroencephalogram signal time-frequency feature fusion, characterized in that, The method comprises the following steps: acquiring a single-channel EEG signal and performing data preprocessing; determining and segmenting an eye blinking action in the preprocessed EEG signal through a sliding window-based eye blinking interval detection algorithm; extracting and optimizing independent eye blinking features of the EEG signal interval determined as the eye blinking action in a time-frequency domain; inputting the extracted features into a pre-trained MLP classifier to realize accurate recognition of the independent eye blinking; the sliding window-based eye blinking interval detection algorithm comprises: determining an eye blinking threshold through adaptive calculation; adopting a sliding window to perform real-time detection on the EEG signal, accurately recognizing and saving an effective eye blinking interval by jointly judging whether a time-domain peak-to-peak value exceeds the eye blinking threshold and whether a frequency-domain energy satisfies a verification condition; the method of determining the eye blinking threshold through adaptive calculation comprises: Randomly select n baseline windows of blink events from the original signal, calculate the peak-to-peak value of each window ; Determine blink threshold based on Tukey's Fence criterion : wherein is the upper quartile, is the lower quartile, is a sensitivity factor; the method of adopting the sliding window to perform real-time detection on the EEG signal, accurately recognizing and saving the effective eye blinking interval by jointly judging whether the time-domain peak-to-peak value exceeds the eye blinking threshold and whether the frequency-domain energy satisfies the verification condition comprises: The sliding window is used to traverse the signal, and a peak-to-peak value of the signal in the current window is extracted , if , the signal in the current window is marked as a candidate blink event. introducing a frequency-domain energy verification mechanism, based on the feature that high-frequency component energy in an eye blinking event presents significant attenuation, constraining the low-frequency energy to be higher than the high-frequency energy, and through dual verification of the amplitude threshold and the frequency-domain energy, accurately detecting the eye blinking signal in the single-channel EEG signal; introducing an event boundary compensation mechanism, setting a compensation window with a front delay and a rear delay based on a sampling rate to ensure the integrity of the eye blinking event.

2. The method of claim 1, wherein, The single-channel EEG signal is acquired by a portable single-channel electroencephalogram (EEG) device equipped with a TGAM. The data preprocessing includes power interference suppression and frequency band optimization.

3. The method of claim 1, wherein, the method of extracting and optimizing independent eye blinking features of the EEG signal interval determined as the eye blinking action in a time-frequency domain comprises: adopting a 5-frequency band segmentation method to decompose the EEG data to obtain five data sets of different frequency bands, respectively corresponding to Delta, Theta, Alpha, Beta and Gamma bands; extracting time-domain features and frequency-domain features of the EEG signal on the five frequency bands, and additionally adding cross-frequency band ratio features of the Delta and Theta bands; standardizing the extracted features, using RF to realize feature importance evaluation, and screening out high-discriminative features; at the same time, extracting principal components whose cumulative variance contribution rate reaches a set threshold through PCA; normalizing the PCA variance contribution and RF importance, weighting, calculating a comprehensive score, and selecting a feature subset according to the comprehensive score.

4. The method of claim 3, wherein, The time-domain features include: mean value, maximum value, minimum value, peak-to-peak value, variance, and zero-crossing rate. The frequency-domain features include: Psd_mean: average value of power spectral density in the entire frequency band; Psd_max: maximum value of power spectral density in the entire frequency band; f_max: frequency corresponding to the maximum value of power spectral density; differential entropy feature, used to measure the complexity or uncertainty of the distribution of continuous random variables.

5. An autonomous blink recognition system based on single-channel electroencephalogram signal time-frequency feature fusion using the method of any one of claims 1-4, characterized in that, The method comprises the following steps: an EEG signal acquisition and preprocessing module for acquiring a single-channel EEG signal and performing data preprocessing; The blink action detection module is configured to determine the blink action in the preprocessed EEG signal by using a sliding window-based blink interval detection algorithm. The autonomous blink feature extraction and optimization module is configured to extract and optimize autonomous blink features of the EEG signal interval determined as the blink action in the time-frequency domain. The autonomous blink recognition module is configured to input the extracted features into a pre-trained MLP classifier to achieve accurate recognition of the autonomous blink.

6. A computer device, comprising: A computer program is stored in the memory and configured to be executed by the processor, and the computer program comprises instructions for executing the method of any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by a computer to implement the method of any one of claims 1-4.

Images