A method for selecting EEG channels based on PLI-Relief

Abstract

The present invention relates to the field of intelligent transportation and EEG signal channel selection technology, and specifically to an EEG channel selection method based on PLI-Relief. The method comprises the following steps: collecting fatigue driving EEG data and preprocessing it to remove artifact components; constructing a correlation matrix based on the phase lag index between each channel signal; binarizing the correlation matrix and constructing a brain function network, and realizing channel preliminary selection according to network attributes; extracting the power spectral density characteristics of the EEG signal of the preliminary selected channel and using the Relief algorithm to sort the channels by weight; using the accuracy selection method to determine the optimal channel and perform fatigue state identification based on it. The advantages of the present invention are: the designed channel selection method comprehensively considers the time-frequency characteristics of the EEG signal and the functional connection relationship between the signals of each channel, and greatly reduces the number of channels while maintaining a high recognition accuracy, providing a feasible solution for the design of wearable EEG fatigue driving monitoring equipment.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation, and in particular to a method for selecting an electroencephalogram (EEG) channel based on PLI-Relief. Background Art

[0002] The popularity of cars has brought great convenience to our lives, but the resulting road traffic safety problems are becoming increasingly serious. Studies have shown that fatigue driving has become one of the main reasons affecting traffic safety. Traffic accidents caused by fatigue driving usually occur when the driver does not take risk avoidance measures, which often leads to chain traffic accidents and even more serious consequences, seriously threatening the country's road traffic safety, economic development and people's quality of life. Therefore, finding a method that can accurately monitor the driver's fatigue state in real time and prompt or warn the driver when he feels tired or sleepy is of great practical significance for reducing casualties and economic losses. At present, the detection method of driving fatigue is mainly based on the detection of the driver's physiological characteristics, especially the physiological characteristics based on EEG signals are considered to be an effective and objective means of judging the fatigue state.

[0003] Identifying driver fatigue based on EEG signals has been a research hotspot in recent years. This approach primarily involves collecting the driver's multi-channel EEG signals using an EEG cap, preprocessing the data to extract EEG signal features, and then using modeling to identify driver fatigue. Common EEG signal analysis methods include time domain analysis, frequency domain analysis, time-frequency analysis, and nonlinear analysis. EEG signals are complex physiological signals with spatially discrete and nonstationary characteristics. Analyzing only in the time-frequency domain or spatial connectivity will lose some valid information, affecting the final recognition accuracy.

[0004] The ultimate goal of fatigue driving monitoring research is the development and application of wearable devices. This places high demands on the real-time performance of fatigue monitoring systems, the portability of acquisition equipment, and operational complexity. Currently, EEG-based fatigue driving monitoring studies mostly use 32-, 64-, or 128-channel devices to collect EEG signals from subjects during simulated driving tasks. In many real-world application scenarios, it's not feasible to deploy a large number of electrodes on the head. This necessitates minimizing the number of EEG acquisition channels while minimizing the loss of fatigue driving monitoring accuracy, thereby better meeting the needs of real-world scenarios. Summary of the Invention

[0005] In response to the above situation, the present invention proposes an EEG channel selection method based on PLI-Relief. When selecting channels, it considers not only the spatial relationship between each EEG channel, but also the time-frequency characteristics of the channel signal itself. While ensuring the recognition accuracy, a large number of redundant channels are eliminated. This method has certain reference value for the development and use of wearable devices.

[0006] The present invention provides a method for selecting EEG channels based on PLI-Relief, and the technical solution includes the following steps:

[0007] Step 1: Preprocess the original EEG data and use the phase lag index to calculate the correlation between channels to obtain the adjacency matrix;

[0008] Step 2: Select a threshold to binarize the adjacency matrix, map the binarized matrix into a brain functional network, and select channels with higher phase synchronization as the primary channels based on network properties;

[0009] Step 3: For each subject, calculate the power spectral density characteristics of the preliminary selected channels, treat each channel as a feature, and then use the Relief algorithm to obtain the weight of each preliminary selected channel. The weights of all preliminary selected channels are normalized between [-1, 1] and the weights of the channel for all subjects are summed to obtain the weight of each preliminary selected channel that is independent of the individual.

[0010] Step 4: Sort the T preliminarily selected channels by weight, and starting from the channel with the largest weight, take the first N channel features for combination, where N ranges from 1 to T. Then, use a common classification algorithm to classify the fatigue state and obtain the accuracy corresponding to each channel feature combination. As the number of channels increases, when the fatigue state recognition accuracy changes relatively little, the channel combination corresponding to the current accuracy is selected as the optimal channel selection result.

[0011] In step 2, the channels are initially selected through the brain functional network. The spatial functional connectivity of the EEG signals of each channel is considered during channel selection, which makes up for the deficiency of the traditional method of using only time-frequency domain features for channel selection research.

[0012] The key algorithm process is as follows:

[0013] (1) Brain network construction

[0014] In the brain functional network, the PLI value between two nodes is calculated as:

[0015]

[0016] Where N represents the number of time points, Δφ rel (t n) represents the phase difference between the two signals at time point n. Sign is a sign function that outputs 1 when the independent variable is positive, -1 when the independent variable is negative, and 0 when the independent variable is 0. The PLI range is [0, 1]. When the PLI value is 0, the phase difference distribution of the two EEG signals is completely symmetrical. Otherwise, the larger the PLI value, the more asymmetrical the phase difference distribution of the corresponding two sets of EEG signals.

[0017] When using the Relief algorithm for channel selection, each EEG signal channel is taken as a feature, and then the weight of each channel is calculated to perform EEG signal channel selection.

[0018] (2) Relief algorithm

[0019] The calculation of weights in the Relief algorithm is shown in formula (2).

[0020]

[0021] Among them, x ih is the i-th sample (i ranges from 1 to k) x i The nearest neighbor sample among the same sample, x im is the sample x i The nearest neighbor sample among the heterogeneous samples, Represents sample x i The value of feature j, Similarly, Represents sample x i and sample x ih The distance between them is calculated in two ways according to the type of feature j. For discrete features:

[0022]

[0023] For continuous features:

[0024] Represents sample x i and sample x im The distance between Similarly, the greater the weight of a feature, the more relevant the feature is to the sample category and the more suitable it is for sample classification. Conversely, the smaller the weight, the less relevant the feature is to the sample category and the less suitable it is for sample classification.

[0025] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0026] 1. This paper combines the spatial functional connectivity of EEG signals with the time-frequency characteristics of the signals themselves to construct a model for EEG channel selection during fatigue driving. The proposed framework considers both the correlation between channels and the characteristics of the channel signals themselves when selecting channels, eliminating a large number of redundant channels. It has certain reference value for the development and use of wearable devices.

[0027] 2. The present invention extracts a feature for each EEG channel, namely the PSD feature corresponding to the channel, and uses each EEG channel as a feature to perform feature selection, that is, channel selection, through the Relief algorithm, providing a new idea for channel selection.

[0028] 3. The present invention maintains a high fatigue state recognition accuracy while selecting the minimum number of channels, overcomes the data redundancy and computational complexity brought about by the use of full-channel EEG signals, and can achieve optimal selection of EEG channels, laying the foundation for the development and use of brain-computer interface wearable devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the main flow chart of EEG channel selection proposed by the present invention

[0030] Figure 2 Flowchart of the EEG channel selection algorithm based on PLI-Relief DETAILED DESCRIPTION

[0031] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0032] See also Figure 1 and Figure 2 ,A PLI-Relief-based EEG channel selection method includes four parts: brain network construction, channel preliminary selection, weight calculation of the ,preliminary selected channels, and selecting the optimal channel using the accuracy ,selection method. The specific steps are as follows:

[0033] Step 1: Collect EEG data in real time, or use a publicly available EEG dataset as raw EEG data. For example, the SEED-VIG dataset, a public dataset for fatigue driving research, can be used. It includes 17 channels of EEG signals from 23 subjects performing a simulated driving task.

[0034] Step 2: Preprocess the original EEG data and use the phase lag index to calculate the correlation between channels to obtain the adjacency matrix;

[0035] The PLI value between two nodes is calculated as:

[0036]

[0037] Where N represents the number of time points, Δφ rel (t n ) represents the phase difference between the two signals at time point n. Sign is a sign function that outputs 1 when the independent variable is positive, -1 when the independent variable is negative, and 0 when the independent variable is 0. The PLI range is [0, 1]. When the PLI value is 0, the phase difference distribution of the two EEG signals is completely symmetrical. Otherwise, the larger the PLI value, the more asymmetrical the phase difference distribution of the corresponding two sets of EEG signals.

[0038] Step 3: For each subject, calculate the power spectral density characteristics of the preliminary channel, treat each channel as a feature, and then use the Relief algorithm to obtain the weight of each preliminary channel. The weights of all preliminary channels are normalized between [-1, 1], and the weights of the channel for all subjects are summed to obtain the weight of each preliminary channel that is independent of the individual. The formula for calculating the weight of the preliminary channel using the Relief algorithm is:

[0039]

[0040] Among them, x ih is the i-th (i ranges from 1 to k) sample x i The nearest neighbor sample among the same sample, x im is the sample x i The nearest neighbor sample among the heterogeneous samples, Represents sample x i The value of feature j, and Similarly; Represents sample x i and sample x ih The distance on feature j is calculated in two ways according to the type of feature j. For discrete features:

[0041]

[0042] For continuous features:

[0043]

[0044] Represents sample x i and sample x im The distance between Similarly, the greater the weight of a feature, the more relevant the feature is to the sample category and the more suitable it is for sample classification. Conversely, the smaller the weight, the less relevant the feature is to the sample category and the less suitable it is for sample classification.

[0045] Step 4: Sort the T preliminarily selected channels according to their weights, and starting from the channel with the largest weight, take the first N channel features for combination, where N ranges from 1 to T. Then, use a common classification algorithm to classify the fatigue state and obtain the accuracy corresponding to each channel feature combination. As the number of channels increases, when the fatigue state recognition accuracy changes relatively little, the channel combination corresponding to the current accuracy is selected as the optimal channel selection result.

Claims

1. A method for selecting EEG channels based on PLI-Relief, characterized in that The following steps are involved: Step 1: Use EEG equipment to collect EEG signals of the subjects in real time while they are performing the driving task; Step 2: Preprocess the raw EEG data; Step 3: Use the Phase Lag Index (PLI) to measure the functional connectivity between channels. Select a threshold to binarize the PLI matrix and construct a brain functional network. Based on the network properties, select channels with higher phase synchronization as the primary channels. The PLI value between two channel signals is expressed as: Where N represents the number of time points, Δφ rel (t n ) represents the phase difference between the two signals at time point n. sign is a sign function. When the independent variable of the function is positive, its output is 1; when the independent variable is negative, its output is -1; when the independent variable is 0, its output is 0. The value range of PLI is [0, 1]. When the value of PLI is equal to 0, it means that the phase difference distribution of the two EEG signals is completely symmetrical. Otherwise, the larger the value of PLI, the more asymmetrical the phase difference distribution of the corresponding two sets of EEG signals. Step 4: For each subject, calculate the power spectral density characteristics of the preliminary selected channels, treat each channel as a feature, and then use the Relief algorithm to obtain the weight of each preliminary selected channel. The weights of all preliminary selected channels are normalized between [-1, 1], and the weights of the channel for all subjects are summed to obtain the weight of each preliminary selected channel that is independent of the individual. The formula for calculating the weight of the preliminary selected channel using the Relief algorithm is: Among them, x ih is the i-th (i ranges from 1 to k) sample x i The nearest neighbor sample among the same sample, x im is the sample x i The nearest neighbor sample among the heterogeneous samples, Represents sample x i The value of feature j, and Similarly; Represents sample x i and sample x ih The distance on feature j is calculated in two ways according to the type of feature j. For discrete features: For continuous features: Represents sample x i and sample x im The distance between Similarly, the greater the weight of a feature, the more relevant the feature is to the sample category and the more suitable it is for sample classification. Conversely, the smaller the weight, the less relevant the feature is to the sample category and the less suitable it is for sample classification. Step 5: Sort the T preliminarily selected channels by weight, and starting from the channel with the largest weight, take the first N channel features for combination, where N ranges from 1 to T. Then, use a common classification algorithm to classify the fatigue state and obtain the accuracy corresponding to each channel feature combination. As the number of channels increases, when the fatigue state recognition accuracy changes relatively little, select the channel combination corresponding to the current accuracy as the optimal channel selection result.

Images