Eye movement classification method based on multi-channel EOG signal airspace and inter-channel feature fusion

Abstract

The invention belongs to the technical field of bio-electricity signal processing, and particularly relates to an eye movement classification method based on multi-channel EOG signal airspace and inter-channel feature fusion. Through the multi-channel electrode layout design, the spatial resolution of eye movement signals is enhanced, and noise interference of a traditional dual-electrode layout is avoided; a CSP spatial domain feature extraction method is adopted, 28 groups of CSP projection matrixes are constructed for 8 types of eye movement tasks, and the distinguishing capability of spatial domain features on tiny eye movement modes is improved; through arranging a time domain and frequency domain traditional feature extraction module, a space domain feature extraction module and an inter-channel correlation feature extraction module, correlation features are extracted so as to improve classification accuracy and stability. According to the method, a complete technical chain of multi-channel electro-oculogram signal acquisition layout, preprocessing, feature extraction and classification decision is adopted, so that the bottleneck of insufficient utilization of spatial information and missing of channel collaboration in a traditional method is solved. Experimental verification shows that the stability under individual difference and the average accuracy rate are 88.23%.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioelectric signal processing, and in particular relates to an eye movement classification method. Background Art

[0002] Currently, the main eye movement classification algorithms based on EOG signals include thresholding, template matching, traditional machine learning (ML), and deep learning (DL). The choice of algorithm depends largely on the research objectives and data characteristics.

[0003] Thresholding methods typically require collecting EOG signals from users with different eye movements. By calculating various features, such as signal amplitude and saccade duration, classification thresholds are then defined in feature space. Thresholds are typically set based on the researcher's own experience. While this method is straightforward, it often lacks robustness and requires frequent calibration.

[0004] Similar to the idea of the threshold method, the template matching method is essentially a process of matching the obtained EOG signal with eye movement templates in different directions. Generally speaking, there are large differences between the EOG signal waveforms generated by eye movements in different directions. Therefore, researchers judge the type of eye movement by calculating the similarity between the signal and different templates. Ding et al. used different EOG signal eye movement templates to correspond to different Chinese strokes, and then used the Dynamic Time Warping (DTW) algorithm to calculate the similarity between the input signal and the template to realize the input of the strokes, and developed a Chinese eye writing input system. In addition, commonly used template matching similarity metrics include Pearson correlation coefficient, root mean square error, etc.

[0005] In addition to thresholding and template matching, a combination of feature engineering and traditional machine learning methods is also widely used in the field of eye movement classification. Machine learning methods extract discriminant features from EOG signals and use machine learning classifiers to classify different eye movements. Researchers typically extract features from the time domain, frequency domain, or a combination of both, such as the maximum amplitude, minimum amplitude, mean, variance, sample entropy, power spectral density, wavelet coefficients, etc. of the EOG signal. The EOG signal is then classified using commonly used machine learning classifiers, such as SVM, KNN, and RF.

[0006] In the existing technology, the classification of electrooculogram signals mainly relies on time domain features (such as amplitude and zero crossing rate) and frequency domain features (such as power spectral density). However, this type of classification method based on traditional features has some defects. First, the spatial distribution characteristics of electrooculogram signals are ignored and the spatial information in the electrooculogram signals is not utilized; second, the existing feature extraction methods do not fully exploit the inter-channel correlation features such as phase synchronization and amplitude correlation between multiple channels. Studies such as reference [1] have confirmed that the lack of spatial information will reduce the ability to recognize eye movement patterns. Therefore, there is an urgent need for a feature extraction method that can integrate spatial domain features and inter-channel correlation. Summary of the Invention

[0007] The purpose of the present invention is to provide an eye movement classification method based on the fusion of multi-channel EOG signal spatial domain and inter-channel features, so as to effectively improve the classification performance of eye movement classification tasks.

[0008] The eye movement classification method based on the fusion of multi-channel EOG signal spatial domain and inter-channel features provided by the present invention has the following specific steps.

[0009] (1) A multi-channel electrooculogram acquisition system [2] is used to collect electrooculogram signals of different eye movements; during the acquisition process, a new layout of 14 electrodes is used, such as Figure 2 As shown in the figure, 12 electrodes are evenly distributed above and below the eyes, with three electrodes on each eye to capture detailed eye movement information; the other two electrodes are placed on the left and right outer canthi of the subject, specifically for capturing horizontal eye movement signals; the ground electrode and the reference electrode are placed on the mastoid process behind the ear and the top of the forehead respectively; after differential processing, the original electrooculogram signal S containing 7 channels is obtained n (n=1,2,…,7) to complete the eye movement classification task in 8 directions (up, down, left, right, upper left, lower left, upper right, and lower right); this step aims to obtain multi-channel electrooculogram data and is not limited to the layout of 14 electrodes.

[0010] (2) In order to ensure the quality of the data sent to the subsequent feature extraction step, the original electrooculogram signal S is obtained in step (1). n Perform signal preprocessing operations, the specific process is as follows Figure 3As shown in the figure, first, the original electrooculogram signal is filtered by a 0.1 to 10 Hz bandpass filter. This filter uses a fourth-order Butterworth filter to retain the effective eye movement energy part of 0.1-10 Hz and remove high-frequency noise interference. Secondly, in order to eliminate the 50 Hz interference caused by the power line, a 50 Hz notch filter is used for filtering. Then, due to the non-stationary characteristics of the electrooculogram signal, its specific time characteristics, such as the start and end of singular points, and the instantaneous features, cannot be fully analyzed by Fourier transform or windowed Fourier transform alone. Therefore, wavelet transform is further used to reduce the noise of the electrooculogram signal. This process includes three main stages: first, wavelet decomposition of the noise signal; second, threshold estimation and thresholding of the wavelet coefficients; finally, the denoised signal is reconstructed by inverse wavelet transform to obtain the denoised electrooculogram signal data; finally, in order to effectively adapt to and correct various shapes of baseline drift, a fourth-order polynomial fitting method is used to remove these non-eye movement related signal components; finally, the pre-processed pure electrooculogram signal data s is obtained. n ; Capital S n Indicates raw data, lowercase s n Indicates preprocessed data.

[0011] (3) For the pre-processed electrooculogram data s obtained in step (2), n ,The present invention proposes a feature extraction method of spatial features and inter-channel features to extract relevant features;

[0012] Specifically, a time domain and frequency domain traditional feature extraction module, a spatial domain feature extraction module, and an inter-channel correlation feature extraction module are set up respectively, where:

[0013] The traditional time-domain and frequency-domain feature extraction module is used to extract the time-domain and frequency-domain features commonly used in the field of eye movement classification research based on electrooculogram signals, and construct the traditional electrooculogram feature vector F0;

[0014] The inter-channel correlation feature extraction module calculates the phase lock value (PLV), cosine distance and logarithmic Euclidean distance between the two channels of the multi-channel electrooculogram signal S1, and calculates the inter-channel correlation feature vector F1; Figure 4 As shown, for each eye movement segment, PLV(s1,s2),PLV(s1,s3),…,PLV(s2,s3),…,PLV(s6,s7), that is, the phase locking value between each two channels, is calculated to obtain the phase locking value feature vector f PLV Similarly, the cosine distance and logarithmic Euclidean distance between each two channels are calculated to obtain the feature vectors and The three feature vectors are concatenated to obtain the inter-channel feature vector F1, that is, Where concat(·) means concatenating vectors to form a vector;

[0015] The spatial feature extraction module uses a spatial filter to extract the multi-channel electrooculogram signal s n Perform spatial projection and calculate the spatial feature F2 through the independent projection matrix of each classification category; Figure 5 As shown in the figure, for the 8-direction eye movement classification task, the 8 categories are combined in pairs, and a binary classification CSP filter is trained for each group to obtain 28 projection matrices; therefore, for each eye movement sample, after filtering using 28 projection matrices, the first 4 and last 4 components with the largest variance are extracted to form a 224-dimensional spatial feature vector F2.

[0016] (4) Through the above feature extraction, the traditional electrooculogram feature vector F0, the spatial domain feature F1 and the inter-channel correlation feature vector F2 are obtained respectively. These three feature vectors are concatenated head to tail to obtain a high-dimensional joint feature vector F. The minimum redundancy maximum correlation algorithm (mRMR) is used to filter out the optimal feature subset f from the high-dimensional joint feature vector F. Finally, the optimal feature subset f is used as the input of subsequent SVM, KNN and RF for classification.

[0017] In step (1):

[0018] The electrode configuration method adopts a new layout of 14 electrodes; among them, 12 electrodes are evenly distributed above and below the eyes, with three positions on each eye to capture detailed eye movement information; the other two electrodes are placed on the left and right outer canthi of the subjects, specifically for capturing horizontal eye movement signals; the ground electrode and reference electrode are placed on the mastoid process behind the ear and the top of the forehead, respectively, to ensure the stability and accuracy of signal acquisition.

[0019] The advantages of the present invention are:

[0020] Compared with traditional eye movement recognition technology, the present invention systematically solves the two problems of neglect of spatial information and lack of dynamic correlation between channels in the existing technology through the dual innovation of spatial information enhancement and multi-channel information extraction. Traditional methods only rely on time-frequency features or single channel analysis, ignoring the phase synchronization and morphological correlation of multiple channels during eye movement. The present invention innovatively integrates the three-dimensional inter-channel features of PLV phase synchronization (capturing the coordinated timing of the electric field around the eye), cosine distance and logarithmic Euclidean distance (characterizing statistical distribution differences) to ensure the stability of the algorithm under individual differences (average accuracy of 88.23%), and effectively improves the classification performance of eye movement classification tasks based on multi-channel electrooculogram signals, which provides a valuable reference for the subsequent feature processing of electrooculogram signals from channel information and spatial information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 11 is a flowchart of the eye movement classification method of the present invention.

[0022] Figure 2 This is a schematic diagram of the electrode setup for multi-channel electrooculogram data collection in the present invention.

[0023] Figure 3 This is a structural block diagram of the electrooculogram signal data preprocessing module in the present invention.

[0024] Figure 4 This is a structural diagram of the inter-channel feature calculation module in the present invention.

[0025] Figure 5 It is a structural diagram of the spatial feature calculation module in the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below through embodiments with reference to the accompanying drawings.

[0027] The eye movement classification method proposed in the present invention mainly includes: a traditional feature extraction module in the time and frequency domains, a spatial feature extraction module and an inter-channel correlation feature extraction module. The spatial feature extraction module uses a spatial filter to perform spatial projection on the multi-channel EOG signal, extracts the spatial features related to the eye movement direction, and constructs a spatial feature set. The inter-channel correlation feature extraction module includes phase measurement: calculating the inter-channel phase locking value (PLV) coefficient to quantify phase synchronization; performing distance measurement: cosine distance and logarithmic Euclidean distance to characterize the amplitude correlation between channels; and constructing an inter-channel feature set. The specific steps are:

[0028] Step 1. Data Collection: In this example, electrooculogram (EOG) data from 20 healthy subjects were collected, including an equal number of males and females. The experimental design required the subjects to complete the same experimental paradigm on two days with an interval of more than one week to ensure sufficient recovery time, thereby avoiding interference from fatigue and changes in physiological state on the results. All experiments were scheduled to be conducted in the same time period to reduce the impact of intraday physiological fluctuations. In addition, all subjects fully understood the purpose and procedures of the study before the experiment and underwent sufficient training to ensure the accuracy and standardization of the eye movement task.

[0029] The entire data collection process was conducted in a relatively dark environment to ensure that the projection screen displaying the saccadic reference points was bright and clear. First, the area where the participant's face would be positioned was cleaned, and the electrodes and ground electrode were installed. The participant was then guided to the experimental area. To standardize posture, participants were instructed to sit comfortably at a table with their feet flat on the floor, knees and hips bent at 90 degrees, and supported on a chair. For smaller participants, a seat cushion was provided to ensure their feet were firmly fixed to the floor. After adjusting the seating position, the participant's head was placed on a chin rest for stability, and their gaze was aligned horizontally, ensuring a distance of 70 cm from the center of their brow to the graphical participant interface and feedback window. This setup helped standardize the data collection process, ensuring consistency for each participant throughout the experiment. Finally, participants wore the wearable EOG system described above. The sampling frequency was set to 250 Hz, with a bandpass filter ranging from 0 to 30 Hz and a notch filter at 50 Hz to eliminate potential high-frequency noise and power line interference.

[0030] During the experiment, participants followed a guided procedure to sequentially perform saccades in eight different directions. The saccade locations were cued by a series of target dots displayed on a computer screen. The location of each target dot was designed to indicate the specific location to which the participant's gaze should be directed, simulating saccades in different directions during daily visual tasks.

[0031] 2. Data Preprocessing: First, the data segments must be segmented. To ensure accurate data segmentation, an automated marker synchronization method is employed. Using marker signals recorded by the Bluetooth module during data acquisition, the start and end of each eye movement event can be precisely located. These marker signals provide precise timestamps for each eye movement event, ensuring accurate alignment and segmentation of the data during subsequent processing.

[0032] Next, a 0.1 to 10 Hz bandpass filter using a fourth-order Butterworth design was applied to the raw EOG signal, and a 50 Hz notch filter was used to remove 50 Hz interference caused by the power line. The EOG signal was further denoised using a wavelet transform. This process involves three main stages: first, wavelet decomposition of the noise signal; second, threshold estimation and thresholding of the wavelet coefficients; and finally, reconstruction of the denoised signal using an inverse wavelet transform. In this example, the Bior5.5 wavelet was selected, combined with a discrete wavelet transform method using minimax and soft thresholding, successfully improving the signal-to-noise ratio to 16.96 dB.

[0033] To effectively adapt to and correct various shapes of baseline drift, a fourth-order polynomial fitting method is adopted to remove these non-eye movement related signal components. After performing the above preprocessing steps, each group of experiments will generate 56 eye movement segments. Throughout the experiment, for each subject, a total of 560 eye movement segments are generated, totaling 11,200 data samples. Among them, the number of valid samples is 10,865.

[0034] 3. Use the time-domain and frequency-domain traditional feature extraction module to select the commonly used time-domain and frequency-domain features in the field of eye movement classification research based on electrooculogram (EOG) signals, and construct the common EOG feature set in this invention. These features include Power Spectral Density (PSD), Variance (Var), and Mean Value (Mean), etc. The detailed feature list is shown in Table 1.

[0035] Table 1. EOG Signal Features

[0036]

[0037] 4. The spatial domain feature extraction module uses the Common Spatial Pattern (CSP) algorithm to extract the spatial domain features of EOG signals. The CSP algorithm uses matrix diagonalization to find a set of optimal spatial filters for projection, maximizing the variance value difference between two types of signals, so as to extract feature vectors with high discrimination from the signals. The CSP algorithm is a spatial filtering method for two-category discrimination. For the multi-class eye movement task, a "one-to-one" CSP strategy is adopted to classify the 8-direction eye movement task. That is, for the EOG signals of 8 categories, the CSP algorithm is implemented pairwise.

[0038] Suppose X1 and X2 are the multi-channel evoked response spatio-temporal signal matrices of two types of eye movements respectively, and their dimensions are both N×T, where N is the number of channels and T is the number of samples collected for each channel. To calculate their covariance matrices, now assume N < T. Generally, a mathematical model of composite sources is used to describe EOG signals, and X1 and X2 can be written as:

[0039]

[0040] where S1 and S2 represent two types of eye movement tasks respectively. Without loss of generality, assume that these two source signals are linearly independent of each other; S mRepresents the signal source shared by the two types of eye movement tasks, which can be understood as a shared source, such as blinking, environmental noise, etc. Assume that S1 is composed of m1 sources and S2 is composed of m2 sources. Then C1 and C2 are composed of m1 and m2 common spatial patterns related to S1 and S2. Since each spatial pattern is an N×1 dimensional vector, this vector is now used to represent the distribution weight of the signal caused by a single source signal on N leads. m Indicates that S m The goal of the CSP algorithm is to design spatial filters F1 and F2 to obtain the spatial factor W.

[0041] The covariance matrices R1 and R2 after normalization of X1 and X2 are:

[0042]

[0043] Where trace(·) represents the matrix trace. Adding the two together yields the mixed spatial covariance matrix R. R is then subjected to eigenvalue decomposition and the eigenvalues are sorted in descending order to calculate the whitened value matrix P.

[0044]

[0045] Where U is the eigenvector matrix of R, and λ is the diagonal matrix of the corresponding eigenvalues. Applying the matrix P to R1 and R2 yields:

[0046] S1=PR1P T ,S2=PR2P T , (4)

[0047] S1 and S2 have a common eigenvector, and there exist two diagonal matrices λ1, λ2 and the same eigenvector matrix B, satisfying:

[0048] S1=Bλ1B T , S2=Bλ2B T , (5)

[0049] λ1+λ2=I, (6)

[0050] For the eigenvector matrix B, when one category S1 has the largest eigenvalue, the other category S2 has the smallest eigenvalue. Therefore, the eigenvectors V corresponding to the largest and smallest m eigenvalues are selected to form the projection matrix W of the current two categories, that is, the spatial filter for the eye movements of S1 and S2. The eigenvector f of each sample data can be calculated using the projection matrix. i :

[0051]

[0052] Among them, Z i Each row of VAR(Z) corresponds to a CSP filter component (i.e., spatial mode), and each column corresponds to a time sample. i ) indicates Z i The variance is calculated for each row.

[0053] 5. Inter-channel correlation feature extraction module: Saccadic eye movements in different directions produce different signal patterns on different electrodes. For example, horizontal saccades mainly produce changes on horizontal electrodes, while vertical saccades are more obvious on vertical electrodes. In the 8-direction eye movement classification task, it is assumed that each task will have a unique synchronization pattern associated with it. Phase Locking Value (PLV) is a statistic used to quantify the phase synchronization between two signals [3]. PLV is used to evaluate the degree of phase synchronization between different channels at a specific frequency. The basic idea of PLV is to evaluate the synchronization of two signals by comparing their phase difference.

[0054] The phase definition used is based on the Hilbert Transformation (HT) [4]. The HT of a signal can be used to calculate the instantaneous amplitude and instantaneous phase by converting the real signal into a complex signal, thereby extracting the instantaneous phase and instantaneous amplitude. The HT formula is defined as follows:

[0055]

[0056] Among them, x(t) represents the electrooculogram signal in a single channel, so the analytical signal corresponding to the original signal is:

[0057]

[0058] j represents the imaginary unit, so the electrode phase can be defined as:

[0059]

[0060] For each pair of electrodes, their phase difference can be further obtained by calculating their respective instantaneous phases:

[0061] Δφ = (φ1-φ2) mod 2π, (11)

[0062] Among them, mod means modulo, φ x (t) represents the electrode phase, φ1(t) and φ2(t) represent the instantaneous phases of the two electrodes (electrode 1 and electrode 2) at time point t, and the instantaneous phase difference between electrode 1 and electrode 2 is calculated.

[0063] Finally, the phase lock value is quantified by calculating the absolute value of the exponential sum of the phase difference, which is:

[0064]

[0065] Where N is the total number of sample points, t is the time point, and φ1 and φ2 are the instantaneous phases of the two electrodes at time t. The PLV value typically ranges from 0 (no phase synchronization) to 1 (high phase synchronization).

[0066] To measure the correlation between different channels in different types of eye movements (such as horizontal and vertical movements), we calculated the distance metric between each two channels. We used the cosine distance and log-Euclidean distance, two classic distance metrics used in academia.

[0067] The cosine distance between channels is defined as:

[0068]

[0069] Among them, X i and X j Represents the electrooculogram data of two channels respectively, x ik and x jk Respectively represent the kth sample point in the electrooculogram data of the two channels.

[0070] The log-Euclidean distance is defined as:

[0071]

[0072] Among them, x ik and x jk Respectively represent the kth sample point in the electrooculogram data of the two channels.

[0073] 6. Feature fusion and classification

[0074] The traditional feature extraction module in the time and frequency domains extracts 13-dimensional features for each channel, thereby generating a total of 91-dimensional features for each eye movement sample.

[0075] Analyze any two categories of the 8-direction eye movement categories. Calculate an independent projection matrix for each pair of categories, and generate projection matrices. For each projection matrix, the four directions with the largest and smallest weights are selected to form an 8-dimensional feature. Ultimately, each pair of categories generates an 8-dimensional feature, for a total of 8 × 28 = 224-dimensional features.

[0076] Inter-channel features: Extract the correlation features between two channels from the multi-channel electrooculogram data of 7 channels. For 7 channels, 3 different correlation indicators can be extracted from each pair of channel combinations, totaling Dimensional features.

[0077] The spatial feature set, inter-channel feature set and traditional time-frequency domain features are spliced into a high-dimensional joint feature vector, and the traditional time-frequency domain feature set is used as a baseline for comparison; feature selection: the minimum redundancy maximum correlation algorithm (mRMR) is used to screen the optimal feature subset to reduce dimensional redundancy.

[0078] 1. Classifier design:

[0079] (a) SVM kernel function selection:

[0080] Using radial basis kernel function (RBF), penalty factor C = 1.0

[0081] (b) KNN dynamic parameters:

[0082] k=5, the distance metric is Euclidean distance

[0083] (c) Random Forest Optimization:

[0084] 100 decision trees, maximum depth 15

[0085] The validation scheme adopts leave-one-subject cross-validation (LOOCV): each time, all the data of one subject are left out as the test set, and the data of the remaining N-1 subjects are used for training, and the cycle is repeated until all subjects are tested.

[0086] 2. Results

[0087] By comparing the performance of three basic classifiers on different feature sets, we analyzed the classification performance using the commonly used EOG feature set and the spatial and inter-channel fusion feature set. The results are shown in Table 2. As can be seen, the inclusion of inter-channel correlation features significantly improved the classification accuracy of all classifiers. The SVM classifier performed the best with the spatial and inter-channel fusion feature set, achieving an accuracy of 88.23%. In comparison, while the KNN-based classification results were weaker than those of the other two classifiers, they still showed improvement with the spatial and inter-channel fusion feature sets.

[0088] Table 2. Comparison of LOOCV accuracy of different feature sets in 8-direction eye movement classification (%)

[0089]

[0090] The present invention solves the following technical problems:

[0091] 1. Multi-channel electrode layout design

[0092] Enhance the spatial resolution of eye movement signals and avoid noise interference from traditional two-electrode layout;

[0093] 2. Improved CSP spatial feature extraction method based on electrooculogram signal multi-classification task

[0094] Construct 28 sets of CSP projection matrices (C(8,2) = 28 inter-class combinations) for 8 eye movement tasks; improve the ability of spatial features to distinguish subtle eye movement patterns (such as oblique eye movements);

[0095] 3. Inter-channel correlation feature extraction

[0096] Phase lock values and inter-channel distance metrics to measure amplitude correlation (cosine distance and logarithmic Euclidean distance) are jointly calculated (21 sets of phase lock values + 21 sets of Euclidean distances + 21 sets of cosine distances between 7 channels);

[0097] 4.Full-process technical solution combination

[0098] A complete technical chain from multi-channel electrooculogram signal acquisition layout → preprocessing → feature extraction → classification decision; systematically solving the bottlenecks of insufficient spatial information utilization and lack of channel coordination in traditional methods; verification scheme ensures the stability of the algorithm under individual differences (average accuracy of 88.23%).

[0099] References

[0100] [1]LV Z,WANG Y,ZHANG C,et al.An ica-based spatial filtering approach to saccadic eog signal recognition[J].Biomedical Signal Processing andControl,2018,43:9-17.

[0101] [2]ZENG Z,TAO L,ZHOU L,et al.A multi-channel eye-mask-based wearablesystem for accurate eye movement recognition and saccadic angle estimation[C] / / 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society(EMBC).IEEE,2024:1-4.

[0102] [3]MARTIN-CHINEA K, J,ACOSTA L.A new plv-spatialfiltering to improve the classification performance in bci systems[J].IEEETransactions on Neural Systems and Rehabilitation Engineering,2022,30:2275-2282.

[0103] [4]LACHAUX J P,RODRIGUEZ E,MARTINERIE J,et al.Measuring phasesynchrony inbrain signals[J].Human brain mapping,1999,8(4):194-208。

Claims

1. An eye movement classification method based on the fusion of multi-channel EOG signal spatial domain and inter-channel features, characterized by: The specific steps are as follows: (1) A multi-channel electrooculogram (EOG) acquisition system was used to acquire EOG signals of different eye movements. During the acquisition process, a 14-electrode layout was used. Twelve electrodes were evenly distributed above and below the eyes, with three electrodes on each side, to capture detailed eye movement information. The other two electrodes were placed on the left and right outer canthi to capture horizontal eye movement signals. The ground electrode and the reference electrode were placed on the mastoid process behind the ear and the top of the forehead, respectively. After differential processing, the original EOG signal S containing 7 channels was obtained. n , n=1,2,…,7; complete the eye movement classification task of 8 directions: up, down, left, right, upper left, lower left, upper right, lower right; this step obtains multi-channel electrooculogram signal data; (2) Obtain the original electrooculogram signal S in step (1) n Signal preprocessing operations were performed. First, the original EOG signal was filtered with a bandpass filter of 0.1 to 10 Hz. This filter used a fourth-order Butterworth filter to retain the effective energy part of the eye movement in the range of 0.1-10 Hz and remove high-frequency noise interference. Secondly, a 50 Hz notch filter was used for filtering. Then, the EOG signal was denoised by wavelet transform, which included three stages: (a) wavelet decomposition of the noise signal; (b) threshold estimation and thresholding of the wavelet coefficients; (c) reconstructing the denoised signal through inverse wavelet transform to obtain the denoised EOG signal data. Finally, a fourth-order polynomial fitting method was used to remove these non-eye movement related signal components. The preprocessed pure EOG signal data was recorded as s n ; (3) The pre-processed electrooculogram data s obtained in step (2) n , the feature extraction method of spatial domain features and inter-channel features is used to extract relevant features; specifically, a time domain and frequency domain traditional feature extraction module, a spatial domain feature extraction module and an inter-channel related feature extraction module are set up respectively, wherein: The time-domain and frequency-domain traditional feature extraction module is used to extract the commonly used time-domain and frequency-domain features in the field of eye movement classification research based on electrooculogram (EOG) signals, and construct a traditional EOG feature vector, denoted as F0. The inter-channel correlation feature extraction module is used to calculate the phase locking value (PLV), cosine distance and logarithmic Euclidean distance between the two channels of the multi-channel electrooculogram signal S1, and obtain the inter-channel correlation feature vector, which is recorded as F1. For each eye movement segment, PLV(s1,s2),PLV(s1,s3),…,PLV(s2,s3),…,PLV(s6,s7), that is, the phase locking value between the two channels, is calculated respectively, and the phase locking value feature vector is obtained, which is recorded as f PLV Similarly, the cosine distance and logarithmic Euclidean distance between each two channels are calculated to obtain the feature vectors and The three feature vectors are concatenated to obtain the inter-channel feature vector F1, that is, Where concat(·) means concatenating vectors to form a new vector; The spatial feature extraction module uses a spatial filter to extract the multi-channel electrooculogram signal s n Perform spatial projection and calculate the spatial feature F2 through the independent projection matrix of each classification category. Specifically, for the 8-direction eye movement classification task, the 8 categories are combined in pairs, and a binary classification CSP filter is trained for each group, resulting in 28 projection matrices. Therefore, for each eye movement sample, after filtering using the 28 projection matrices, the first 4 and last 4 components with the largest variance are extracted to form a 224-dimensional spatial feature vector, recorded as F2. (4) Through the above feature extraction, the traditional EOG feature vector F0, the spatial domain feature F1, and the inter-channel correlation feature vector F2 are obtained respectively. The three feature vectors are concatenated head-to-tail to obtain a high-dimensional joint feature vector F. The minimum redundancy maximum relevance (mRMR) algorithm is used to select the optimal feature subset f from the high-dimensional joint feature vector F. Finally, the optimal feature subset f is used as the input for subsequent support vector machine (SVM), k-nearest neighbor (KNN), and random forest (RF) for classification.

2. The eye movement classification method according to claim 1, characterized in that In step (3), the construction of the traditional EOG feature vector F0 includes the following features: power spectral density (PSD), variance (Var), mean (Mean), maximum signal amplitude (Max), minimum signal amplitude (Min), maximum and minimum amplitude difference of EOG signal (AMP), root mean square (RMS), average amplitude of EOG signal (Mean), signal energy (Energy), kurtosis (Kurt) which measures the degree of deviation of the measured signal from a certain distribution, crest factor (CF), form factor (FF), power spectral density from 1 - 20 Hz (NLE), non-linear energy (NLE), sample entropy (SampleEn) which measures the unpredictability or randomness of a time series.

3. The eye movement classification method according to claim 2, characterized in that: In step (3), the spatial domain feature extraction module uses the common spatial pattern (CSP) algorithm to extract the spatial domain features of EOG signals. The CSP algorithm uses the diagonalization of matrices to find a set of optimal spatial filters for projection, maximizing the variance value difference between two types of signals, so as to extract a feature vector with high discrimination from the signals. For the multi-class eye movement task, the "one-versus-one" CSP strategy is adopted to classify the 8-direction eye movement task, that is, for the EOG signals of 8 categories, the CSP algorithm is implemented pairwise. Specifically: Assume that X1 and X2 are the multi-channel evoked response spatio-temporal signal matrices of two types of eye movements respectively, and their dimensions are both N×T, where N is the number of channels and T is the number of samples collected for each channel. To calculate their covariance matrices, assume N < T. The mathematical model of compound sources is used to describe the EOG signals, and X1 and X2 are written as: Among them, S1 and S2 represent two types of eye movement tasks respectively; let’s assume that these two source signals are linearly independent of each other; S m Represents the common signal sources under the two types of eye movement tasks; assuming that S1 is composed of m1 sources and S2 is composed of m2 sources; then C1 and C2 are composed of m1 and m2 common spatial patterns related to S1 and S2. Since each spatial pattern is an N×1 dimensional vector, this vector is used to represent the distribution weight of the signal caused by a single source signal on N leads; C m Indicates that S m The corresponding common spatial pattern; the goal of the CSP algorithm is to design spatial filters F1 and F2 to obtain the spatial factor W; The normalized covariance matrices R1 and R2 of X1 and X2 are respectively: where trace(·) represents the trace of the matrix; the sum of the two can obtain the mixed spatial covariance matrix R; then the eigenvalue decomposition of R is performed, and the eigenvalues are arranged in descending order to calculate the whitening value matrix P: where U is the eigenvector matrix of R, λ is the diagonal matrix of the corresponding eigenvalues, and the matrix P is applied to R1 and R2 to obtain: <h2 style=";text-align:left;direction:ltr">S1=PR1P<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> S2=PR2P<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (4) S1 and S2 have common eigenvectors, and there exist two diagonal matrices λ1, λ2 and the same eigenvector matrix B, satisfying: <h2 style=";text-align:left;direction:ltr">S1=Bλ1B<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> , S2=Bλ2B<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (5) λ1 + λ2 = I, (6) For the eigenvector matrix B, when one category S1 has the largest eigenvalue, the other category S2 has the smallest eigenvalue. Therefore, the eigenvectors V corresponding to the largest and smallest m eigenvalues are selected to form the projection matrix W of the current two categories, that is, the spatial filter for the eye movements of S1 and S2. The eigenvector f of each sample data can be calculated using the projection matrix. i : Among them, Z i Each row corresponds to a CSP filter component, and each column corresponds to a time sample; VAR(Z i ) indicates Z i The variance is calculated for each row.

4. The eye movement classification method according to claim 3, characterized in that: In the inter-channel correlation feature extraction module described in step (3), for the 8-direction eye movement classification task, it is assumed that each task will have a unique synchronization pattern associated with it, and the phase locking value (PLV) is used to evaluate the phase synchronization degree between different channels at a specific frequency, that is, the synchronization of two signals is evaluated by comparing their phase difference; specifically: The phase definition used is based on the Hilbert transform (HT) for calculating instantaneous amplitude and instantaneous phase. It converts a real signal into a complex signal to extract the instantaneous phase and instantaneous amplitude. The HT formula is defined as follows: Among them, x(t) represents the electrooculogram signal in a single channel, so the analytical signal corresponding to the original signal is: j represents the imaginary unit, and the electrode phase is defined as: For each pair of electrodes, their phase difference is further obtained by calculating their respective instantaneous phases: Δφ=(φ1-φ2)mod 2π, (11) Among them, mod means modulo, φ x (t) represents the electrode phase, φ1(t) and φ2(t) represent the instantaneous phases of the two electrodes (electrode 1 and electrode 2) at time point t, so the instantaneous phase difference between electrode 1 and electrode 2 is calculated; Finally, the phase lock value is quantified by calculating the absolute value of the exponential sum of the phase difference, which is: Where N is the total number of sample points, t is the time point, φ1 and φ2 are the instantaneous phases of the two electrodes at time t; the value range of PLV is usually from 0 to 1, where 0 indicates no phase synchronization and 1 indicates high phase synchronization; At the same time, in order to measure the correlation between different channels in different types of eye movements, the distance metric between each two channels is calculated; the specific distance metric indicators used are cosine distance and logarithmic Euclidean distance; The cosine distance between channels is defined as: Among them, X i and X j Represents the electrooculogram data of two channels respectively, x ik and x jk Respectively represent the kth sample point in the electrooculogram data of the two channels; The log-Euclidean distance is defined as: Among them, x ik and x jk Respectively represent the kth sample point in the electrooculogram data of the two channels.

5. The eye movement classification method according to claim 4, characterized in that: Further feature fusion and classification; specifically: The traditional feature extraction module in the time and frequency domains extracts 13-dimensional features for each channel, and generates a total of 91-dimensional features for each eye movement sample; Analyze any two categories of the 8-direction eye movement categories; calculate an independent projection matrix for each pair of categories, and generate projection matrices; for each projection matrix, select the four directions with the largest and smallest weights to form an 8-dimensional feature; finally, each pair of categories generates an 8-dimensional feature, for a total of 8×28=224-dimensional features; Inter-channel features: Extract the correlation features between two channels from the multi-channel electrooculogram data of 7 channels; 7 channels, extract 3 different correlation indicators from each pair of channel combinations, totaling dimensional features; The spatial feature set, inter-channel feature set and traditional time-frequency domain features are spliced into a high-dimensional joint feature vector, and the traditional time-frequency domain feature set is used as a baseline for comparison; feature selection: the minimum redundancy maximum correlation algorithm (mRMR) is used to screen the optimal feature subset to reduce dimensional redundancy.

Images