High-dimensional electroencephalogram clustering decoding method under complex visual stimulation

Abstract

The invention discloses a high-dimensional electroencephalogram clustering decoding method under complex visual stimulation. The method comprises the following steps: designing an electroencephalogram signal acquisition experiment normal form under rapid sequence visual presentation; electroencephalogram signals are collected and processed; the electroencephalogram signals are further processed, high-quality electroencephalogram signals from sufficient subjects are obtained, and an electroencephalogram signal data set under rapid sequence visual presentation is formed; performing clustering analysis on the high-dimensional electroencephalogram signal data; the electroencephalogram signals are accurately identified; and the identification effect is evaluated. According to the method, potential characteristics of electroencephalogram signals under different visual stimuli are mined, and effective identification and analysis of electroencephalogram modes under different stimuli are achieved.

Description

Technical Field

[0001] The invention relates to a high-dimensional electroencephalogram clustering decoding method under complex visual stimulation, and belongs to the technical field of intelligent information processing. Background Art

[0002] The main methods of target recognition and detection at present are machine vision and manual interpretation. Machine vision is a study that enables computers to simulate human visual mechanisms to understand the objective world. With the help of the efficient operation speed of computers, machine vision can achieve rapid detection of target images. In a relatively complex background environment, it is necessary to accurately and quickly identify the current image based on relevant theoretical knowledge. Simple machine vision methods often cannot achieve ideal recognition results. The human visual system is controlled by the human brain, which can overcome the fact that machine vision is easily affected by factors such as complex image background and semantic gap. It has high accuracy and strong robustness when identifying targets, and does not require high image quality. It has the characteristics of robustness and reliability. However, traditional manual target recognition is often limited by body movements, and its visual cognitive ability is affected by subjective uncertainty. It cannot fully utilize the human brain's ability to quickly understand images, resulting in low efficiency, and the interpretation staff are prone to visual fatigue under high-intensity work. Therefore, it is necessary to combine human brain cognitive intelligence and machine computational intelligence, based on brain-computer interface technology, and comprehensively utilize the respective advantages of the two to convert the human brain's judgment and response to the target based on experience and intuition into characteristic signals, thereby achieving a significant improvement in the ability to recognize massive image target information and the speed of reaction, and solving the problem of target detection in complex environments.

[0003] Identification and analysis of the brain's visual system has important guiding significance for in-depth understanding of the brain's cognitive mechanism for complex target recognition. Studies have shown that the human brain's visual system processes in a hierarchical manner, extracting features from low-level system areas, to high-level shapes or target parts, and then to higher-level entire targets or target behaviors. High-level features are a combination of low-level features, and the feature representation from low to high levels becomes stronger and stronger, and increasingly expresses semantics or intentions. Starting from the perspective of neuroelectrophysiology, the present invention studies the evolution mechanism of EEG signals under complex visual stimulation, and uses cutting-edge artificial intelligence knowledge to decode the brain's visual cognitive process. It can provide new ideas for understanding the visual cognitive mechanism, and provide a theoretical basis and solution for intelligent image processing technology for complex target recognition. Summary of the invention

[0004] The technical problem solved by the present invention is: aiming at the problem of identifying and decoding brain nerve information under complex visual stimulation, a high-dimensional EEG clustering decoding method under complex visual stimulation is provided.

[0005] The present invention relates to a high-dimensional EEG clustering decoding method under complex visual stimulation, comprising the following steps:

[0006] S1. Experimental paradigm design of EEG signal acquisition under rapid sequence visual presentation;

[0007] In the rapid serial visual presentation experiment, the subject's task is to pay attention to the target picture and press a button as soon as possible after it appears. By analyzing the subject's EEG response, it is determined whether the subject has noticed the target picture, thereby achieving information transmission. The experimental setting parameters include stimulation type, stimulation duration, stimulation interval, target stimulation, target stimulation frequency, and response method.

[0008] S2. The experiment starts, the EEG signal acquisition system acquires EEG signals and processes the signals;

[0009] S3. further processing the EEG signals in the above step S2 to obtain high-quality EEG signals from a sufficient number of subjects to form an EEG signal data set under rapid sequence visual presentation;

[0010] S4. performing cluster analysis on the high-dimensional EEG signal data formed in the above step S3, including a step of formulating a high-dimensional clustering strategy and a step of analyzing the performance of a high-dimensional clustering algorithm;

[0011] S5.Under unsupervised conditions, high-dimensional EEG signal decoding and target classification based on high-dimensional spectral clustering are used to mine the potential characteristics of EEG signals under different visual stimuli in order to accurately identify EEG signals under different stimuli;

[0012] S6. Based on the EEG signal clustering result obtained in the above step S5, the recognition effect is evaluated.

[0013] In step S1, the rapid serial visual presentation experimental paradigm and parameters with a presentation frequency of 12.5 Hz are as follows: the visual stimulus image is selected according to the task requirements; the stimulus duration is set to 80 ms; the stimulus interval is set to 0; the target stimulus frequency is set to 5%; and the response mode is set to the subject counting the number of targets.

[0014] In this experiment, the subjects wore an electrode cap and sat 1 meter away from a 1920×1080 resolution monitor. A fixed cross mark was added in the middle of the visual stimulus image as a fixation point. The subjects pressed the space bar on the keyboard to start viewing the image sequence. Each experimental stage had 10 sequences, each sequence contained 200 pictures, and was repeated 5 rounds; the interval between two sequences was controlled by the subjects, and pressing the space bar started the presentation of the next sequence of images.

[0015] In step S2, the EEG signal acquisition system includes an amplifier circuit, a filter circuit and an analog-to-digital converter; the amplifier circuit includes three stages of amplification, namely, pre-amplification, intermediate stage amplification and final stage amplification, and a filter circuit is interspersed in the three-stage amplifier circuit. The pre-amplification adopts an instrument amplifier and a common-phase parallel differential amplifier.

[0016] The step S3 includes a reference electrode resetting step and a filtering step; in the reference electrode resetting step, the reference electrode is reset with the left mastoid channel M1 and the right mastoid channel M2 as reference; in the filtering step, a cutoff frequency is set, the sampling frequency is set to at least 3 times the cutoff frequency, and a low-pass filter is selected.

[0017] The cut-off frequency is 0.3 Hz, and a 65 Hz band-pass FIR digital filter is selected.

[0018] In step S4, the high-dimensional clustering strategy is selected from one of dimensionality reduction clustering, hypergraph-based clustering, subspace clustering and joint clustering.

[0019] In step S5, an attribute weighted spectral clustering algorithm based on knowledge entropy is adopted. The algorithm uses the relevant concept of knowledge entropy in rough set to evaluate the importance of each EEG feature, and then assigns it as a weight to the corresponding feature, and then applies the spectral clustering method to cluster the data points.

[0020] The specific steps of the above algorithm are as follows:

[0021] S51. Construct a similarity matrix: construct a similarity matrix based on the data, where each element in the matrix represents the similarity between two data points. Commonly used similarity metrics include Gaussian kernel function and k nearest neighbors, where k is the number of clusters;

[0022] S52. Construct a graph: construct a graph according to the similarity matrix, where each data point represents a node and the weight of each edge represents the similarity between two nodes;

[0023] S53. Calculate the Laplacian matrix: Calculate the Laplacian matrix of the graph; the Laplacian matrix can be used to represent the structural information of the graph;

[0024] S54. Calculate eigenvalues ​​and eigenvectors: Calculate the eigenvalues ​​and eigenvectors of the Laplace matrix, and select the eigenvectors corresponding to the first k smallest eigenvalues;

[0025] S55. Clustering: Use K-Means algorithm to cluster feature vectors.

[0026] In the evaluation, the silhouette coefficient, Dunn index, and CH index are introduced; if the true clustering label of the data is already known, one or more of the adjusted Rand index, adjusted mutual information, homogeneity, completeness, and V-measure score are introduced in the review.

[0027] Compared with the prior art, the present invention has the following beneficial effects: the present invention organically combines brain science and information science, fully utilizes the scientific achievements in the field of human brain visual cognition and biological neural electrophysiology, and condenses the relevant laws and knowledge information of neural electrophysiology; studies the clustering and recognition strategy of EEG signals under certain visual stimulation conditions, and transfers visual cognitive skills to the field of complex target recognition; constructs an image processing and recognition system based on brain cognition to solve problems such as real-time environmental perception and target recognition in complex backgrounds, which is expected to break through the traditional image recognition model based on computer vision, and can provide new ideas for building high-precision intelligent perception with high accuracy and generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of the high-dimensional EEG clustering decoding method under complex visual stimulation of the present invention.

[0029] Figure 2 Schematic diagram of the RSVP-based EEG signal acquisition experimental paradigm flow chart

[0030] Figure 3 Flowchart for EEG signal acquisition

[0031] Figure 4 Comparison of waveforms before and after channel filtering

[0032] Figure 5 The original acquisition signal FP1 channel time domain waveform and frequency domain power spectrum

[0033] Figure 6 The time domain waveform and frequency domain power spectrum of the FP1 channel after processing

[0034] Figure 7 Flowchart for EEG signal preprocessing under complex visual stimulation

[0035] Figure 8 Schematic diagram of the clustering results of EEG signals of a certain subject DETAILED DESCRIPTION

[0036] See also Figure 1 The present invention relates to a high-dimensional EEG clustering decoding method under complex visual stimulation, comprising the following steps:

[0037] S1. Design of experimental paradigm for EEG signal acquisition under rapid sequence visual presentation.

[0038] Aiming at the problem of rapid and accurate recognition of complex images, the present invention designs an effective electroencephalogram (EEG) signal acquisition experimental paradigm in combination with complex target visual stimulation background, and uses EEG signal acquisition equipment to complete the acquisition of EEG signals of the subjects. By analyzing the theoretical technology and methods of the design of neural signal acquisition experimental paradigm under typical rapid presentation (Rapid Serial Visual Presentation, RSVP) visual stimulation, a rapid serial visual presentation EEG signal acquisition experimental paradigm that meets the project requirements is designed.

[0039] Rapid serial visual presentation is a commonly used experimental paradigm to measure attention and cognitive processing ability. In this experimental paradigm, the subjects need to watch a series of rapidly presented visual stimulus images in a short period of time and look for the target stimulus among them. In the RSVP experiment, the subject's task is to pay attention to the target image and press the button as soon as possible after it appears. By analyzing the subject's EEG response, it can be determined whether the subject has noticed the target image, thereby achieving information transmission.

[0040] In order to study the visual cognition of the visual system under the rapid presentation of visual stimuli, the present invention designs a rapid serial visual presentation experimental paradigm in which each image is presented for a certain time (for example, the image presentation time is 80ms, that is, the presentation frequency is 12.5Hz). The visual stimulus image adopts a complex target image to study the perception ability of the human visual system to the relevant visual stimulus images. The advantages of the RSVP experimental paradigm are simplicity, efficiency, and ease of operation, and it is suitable for a variety of application scenarios, such as image retrieval, safety detection, cognitive assessment, etc. When designing the RSVP experimental paradigm, it is usually necessary to consider factors such as stimulus type, stimulus duration, stimulus interval, target stimulus, target stimulus frequency, and reaction mode, which are specifically described as follows:

[0041] Stimulus type: The stimulus type should be selected according to the research purpose and hypothesis. For example, if you want to study language processing, you can choose words or sentences as stimuli; if you want to study visual processing, you can choose pictures or graphics as stimuli. The stimulus type will also affect the characteristics of the EEG signal. For example, language stimulation will cause components such as N400 and P600, while visual stimulation will cause components such as P1, N1 and P3.

[0042] Stimulus duration: Stimulus duration refers to the time each stimulus is presented on the screen, which determines the depth and difficulty of the subject's processing of the stimulus. Generally speaking, the shorter the stimulus duration, the more difficult the processing, the higher the attention, and the stronger the EEG signal. However, if the stimulus duration is too short, it may cause the subject to be unable to recognize the stimulus or produce a visual aftereffect, affecting the experimental effect. Therefore, the stimulus duration should be appropriately adjusted according to the stimulus type and task requirements, usually between 100-500 milliseconds.

[0043] Interstimulus interval: The interstimulus interval refers to the time interval between two consecutive stimuli, which affects the subject's memory and comparison of the previous and subsequent stimuli. Generally speaking, the shorter the interstimulus interval, the more difficult the memory and comparison, the higher the attention, and the stronger the EEG signal. However, if the interstimulus interval is too short, it may cause the subject to be unable to distinguish the previous and subsequent stimuli or produce visual fusion, affecting the experimental effect. Therefore, the interstimulus interval should also be appropriately adjusted according to the stimulus type and task requirements, usually between 0-200 milliseconds.

[0044] Target stimulus: Target stimulus refers to the stimulus that the subject needs to pay special attention to or respond to in RSVP. It can be used to measure the subject's attention, reaction speed, accuracy and other indicators. Target stimulus can be set with different characteristics according to different tasks, such as color, shape, size, position, semantics, etc. Target stimulus will also cause specific EEG components, such as P3, N2pc and LPC.

[0045] Target stimulation frequency: The target stimulation frequency refers to the probability of the target stimulation appearing in RSVP, which affects the subject's expectations and degree of surprise. Generally speaking, the lower the target stimulation frequency, the lower the subject's expectations, the higher the surprise, and the stronger the EEG signal. However, if the target stimulation frequency is too low, it may cause the subject to lose interest or attention, affecting the experimental effect. Therefore, the target stimulation frequency should be selected according to the research purpose and hypothesis, usually between 5-40%.

[0046] Response mode: Response mode refers to the way the subject responds to the target stimulus. It can be used to measure indicators such as the subject's response speed and accuracy. Response modes can take many forms, such as key presses, eye movements, and spoken language. Response modes will also affect the characteristics of EEG signals. For example, key presses will cause motor evoked potentials (MEPs), eye movement reactions will cause electrooculograms (EOGs), and spoken language reactions will cause electromyograms (EMGs).

[0047] Aiming at the problem of rapid visual stimulation image target recognition to be solved by the invention, an RSVP experimental paradigm and key parameters with a presentation frequency of 12.5Hz are designed. The visual stimulation images are selected according to the task requirements; the stimulation duration is set to 80ms; the stimulation interval is set to 0 according to the conventional RSVP paradigm, that is, there is no interval between the two visual stimulation images, and the next visual stimulation image appears immediately after the presentation of one image; the target stimulation frequency is set to 5%, that is, there are only 5 target images in every 100 visual stimulation images, and the others are non-target images; the response mode is set to the subject counting the number of targets to avoid interference from motion signals generated by key presses.

[0048] In this experiment, the subjects wore electrode caps and sat 1 meter away from a 1920*1080 resolution monitor. In order to reduce the influence of the subjects' eye movements during the experiment, a fixed cross mark was added in the middle of the visual stimulus image as the fixation point. The subjects pressed the space bar on the keyboard to start viewing the image sequence. Each experimental stage included 10 sequences, each sequence contained 200 pictures, and was repeated 5 times. The interval between two sequences was controlled by the subjects, and pressing the space bar started the presentation of the next sequence of images. See the experimental process for details. Figure 2 .

[0049] S2. The experiment starts, the EEG signal acquisition system acquires EEG signals and processes the signals.

[0050] The characteristics of general EEG signals are:

[0051] ①The signal amplitude is weak and needs to be amplified by tens of thousands times;

[0052] ② The signal frequency range is low (within 100Hz, and extremely low frequency drift also needs to be filtered out), and high-frequency noise needs to be filtered out;

[0053] ③ The signal source impedance is high (there is a resistance of several thousand ohms between the scalp and the skull), which requires the preamplifier circuit to have a very high input impedance, generally greater than 10 megohms;

[0054] ④The power frequency interference of the power supply is common mode interference, and the amplifier common mode suppression needs to be relatively high;

[0055] ⑤Low signal-to-noise ratio and strong randomness.

[0056] Considering these characteristics, the EEG signal acquisition system mainly includes: amplification, filtering circuit design, analog-to-digital converter design, etc. It is generally set to 3-stage amplification: the first stage (preamplification): requires higher input impedance, higher common-mode rejection ratio, low offset voltage and drift, low noise and other performance, and can use instrument amplifiers, in-phase parallel differential amplifier circuits. The second stage (intermediate stage amplification) and the third stage (final stage amplification): need to meet high gain and low noise, low offset voltage and drift, good stability and output swing. The amplification process is interspersed with filtering processes: high-pass, low-pass, and band-stop filtering.

[0057] The experiment uses an electroencephalograph specially used for measuring and recording electroencephalograms to collect electroencephalogram signals. The working process is as follows: the electrodes placed on the scalp can detect weak EEG signals, usually 5-100μV, which are coupled to the differential amplifier through the electrode leads for amplification by about 10000-100000 times, and then converted into digital signals through the data acquisition card in the computer host, and finally the signal data is recorded by the recording system on the matching PC.

[0058] The EEG acquisition equipment of the present invention adopts the 64-lead device of NeuroScan SynAmps of Australia, which is specially designed for the acquisition of weak electrophysiological signals and shielded from the outside world. It uses standardized and reasonable operation methods to collect EEG signals in a laboratory environment that is specially designed for the acquisition of weak electrophysiological signals. The following are the following: ① To ensure good conductivity between the scalp and the sensor, the subject should first wash his hair before acquisition; ② The EEG acquisition equipment is connected and debugged, and the amplifier, AD converter, data transmission optical fiber and data recording notebook are correctly connected so that the EEG signals can be effectively collected; ③ Assist the subject to wear the EEG cap to ensure that the electrode position matches the corresponding area of ​​the scalp; ④ Explain the experimental process to the subject to confirm that the subject has a full understanding of the experimental process and operation; ⑤ Start the experiment. During the signal recording process, the experimenter always pays attention to the reaction of the subject, prevents and promptly handles and resolves emergencies such as equipment failure and subject discomfort during the experiment, and supervises the smooth progress of the entire experiment; ⑥ After the experiment, organize the equipment and keep good experimental records. See the signal acquisition flow chart Figure 3 .

[0059] S3. Further process the EEG signals in the above step S2 to obtain high-quality EEG signals from a sufficient number of subjects to form an EEG signal data set under rapid sequence visual presentation.

[0060] The present invention designs and implements a method for further processing of EEG signals, obtains high-quality EEG signal data from a sufficient number of subjects, and forms a rapid serial presentation EEG signal data set, providing a key data basis for subsequent EEG signal processing under complex visual stimulation.

[0061] 1) Reset the reference electrode

[0062] The electrophysiological amplifier used for EEG acquisition is a differential amplifier, which generally uses a unipolar lead. Each position electrode is connected to one input terminal of the amplifier, and the other input terminal is connected to a common reference electrode, so the potential of each electrode should be the result of subtracting its absolute potential from the absolute potential of the same reference electrode. The ideal reference electrode point should be a constant potential, but because the human body is a volume conductor, the bioelectricity of each part is changing at any time, so theoretically an ideal reference electrode does not exist. The present invention uses the left and right mastoid channels (M1 and M2 channels) as a reference to reset the reference electrode.

[0063] 2) Filtering

[0064] There are two main reasons for filtering the collected raw EEG data:

[0065] ① From the Nyquist theorem, we can know that if the collected original signal contains a signal with a frequency higher than 2 times the sampling frequency, frequency aliasing will occur, so a lower frequency needs to be selected as the cutoff frequency. Generally, in practical applications, the sampling frequency is at least 3 times the cutoff frequency.

[0066] ②Some noise contained in EEG is quite different from useful signals in frequency. In cognitive neuroscience experiments, the frequency range of most experimentally relevant useful components in EEG signals is between 0.01Hz and 60Hz, so a low-pass filter with a cutoff frequency of around 60Hz can be set to filter out noise information and retain effective components. The filter should be used reasonably, otherwise it will cause obvious signal distortion in severe cases. In addition to filtering out high-frequency signals, filters are also required in many experiments to attenuate extremely low-frequency signals to remove slow voltage changes caused by non-neural activities during the recording process. For example, sweating, electrode resistance drift, etc. will cause slow and continuous changes in the EEG baseline voltage, resulting in slow and long-term voltage drift, causing large distortion of the waveform. A high-pass filter with a cutoff frequency of about 0.1Hz can be used to filter out those extremely low frequencies.

[0067] The present invention uses a bandpass FIR digital filter with cutoff frequencies of 0.3 Hz and 65 Hz to filter out low-frequency and high-frequency noise. The waveforms of the EEG data collected by a subject before and after channel filtering are shown in the figure below. Figure 4 As shown, by comparison, the high-frequency noise components are obviously filtered out.

[0068] The original EEG signal FP1 channel time domain waveform and frequency domain power spectrum are as follows Figure 5 As shown in the figure, it can be seen that the original signal collected contains complex noise interference, and the signal amplitude is large, with an order of magnitude of 10 4 Microvolt, wide frequency band, occupying [0, 100Hz], effective weak electrophysiological signals are submerged.

[0069] The time domain waveform and frequency domain power spectrum of the corresponding channel signal after filtering are shown in Figure 6 It can be seen that the amplitude of the extracted weak electrophysiological signal is on the order of tens of microvolts, and the frequency band occupies [0, 20Hz], which is consistent with the characteristics of electrophysiological signals and verifies the effectiveness of the weak signal processing method.

[0070] This step S3 refers to Figure 7 .

[0071] S4. Perform cluster analysis on the high-dimensional EEG signal data formed in the above step S3, including the steps of formulating a high-dimensional clustering strategy and analyzing the performance of a high-dimensional clustering algorithm.

[0072] The selected EEG features still have high dimensions. The traditional feature decoding model takes a long time to train and is prone to overfitting, making it difficult to identify. Therefore, it is necessary to propose a target classification method for high-dimensional EEG features, fully condense and integrate multi-scale EEG signal features and perform clustering analysis to provide a technical basis for accurate classification and recognition of visual stimulus targets based on high-dimensional neural information. Clustering algorithms are unsupervised learning methods, which are significantly different from supervised classification methods such as discriminant analysis. In supervised classification, the system provides a set of labeling patterns in advance, and the learning task is to label a newly encountered, unlabeled data. Usually, a given training pattern is used to learn the characteristics of the class, and then the trained model is used to process new objects. In unsupervised clustering, there is generally no specific labeling information. Its purpose is to split the unlabeled data set into multiple non-overlapping clusters, so that the data points within the class are as similar as possible, and the data points between classes are as different as possible. The category labels of these clusters are obtained only from the data and are data-driven. Clustering can be considered as a concise model for explaining data, which describes the basic structure of the data in summary. There are many methods currently used to explore the relationship between data objects. However, traditional clustering algorithms can no longer meet the requirements of today's data analysis. Some algorithms are only suitable for processing data sets with fewer attributes. When the dimension of the data is very high, the distribution of samples is too dispersed and different objects cannot be effectively distinguished, so the clustering effect becomes very poor. Effective measures need to be taken to convert high-dimensional data into low-dimensional data while retaining the main features of the data to make it easier to cluster the data.

[0073] (1) Develop high-dimensional clustering strategies

[0074] The high-dimensional target object in cluster analysis may need more than dozens of aspects to be represented, and each attribute becomes a dimension of the data object. Cluster analysis of high-dimensional data has become an important research field. In the application field related to high-dimensional data, the curse of dimensionality is a very common phenomenon, which refers to a series of problems caused by too many variables in data analysis. The research on clustering algorithms in the field of data mining has achieved certain results, but many traditional clustering algorithms can usually obtain more accurate results in clustering general low-dimensional data, while the processing effect of high-dimensional data is often affected by the curse of dimensionality. If these traditional algorithms are used for clustering, the expected results are often not obtained. In order to meet the needs of many users in different fields, researchers have proposed many clustering methods for high-dimensional data, mainly including clustering based on dimensionality reduction, clustering based on hypergraph, subspace clustering and joint clustering, etc.:

[0075] Dimensionality reduction clustering: Dimensionality reduction is a technique that seeks data compactness by mapping data points to a lower-dimensional space. This compact representation of low-dimensional space is conducive to further processing of data. Dimensionality reduction is one of the important research branches in many research fields. Its methods are diverse. According to different dimensionality reduction methods, many clustering methods based on dimensionality reduction have been formed, such as Kohonen self-organizing feature mapping (SOFM), principal component analysis (PCA), multidimensional scaling (MDS), and fractal dimensionality reduction. No matter what dimensionality reduction method is used to cluster high-dimensional data, its basic purpose is to first seek an equivalent low-dimensional representation of high-dimensional data according to the corresponding method, and then use the existing traditional clustering method to cluster the data after dimensionality reduction, that is, to use the clustering results of data in low-dimensional space to represent the clustering characteristics of high-dimensional data. Different clustering methods have different ways of seeking low-dimensional representation of high-dimensional data, and the degree of approximation between the data after dimensionality reduction and the original data is also different, so their clustering performance is also different.

[0076] Clustering based on hypergraph: Hypergraph is an extension of conventional graph. Each edge in the graph can connect multiple vertices to become a hyperedge. Clustering methods based on hypergraph map the relationship of high-dimensional data to a hypergraph. Each edge in the graph expresses the relationship of these data, and the weight on the edge indicates the closeness of the corresponding relationship. On this basis, the clustering method based on hypergraph is actually to find a partition of the hypergraph vertex and make the data in the same partition as relevant as possible. The clustering steps based on hypergraph partition can be simply described as: ① Define the conditions for a point (as a vertex of the graph) to be connected to other points through the hypergraph; ② Define the measurement of the connection weight in the graph; ③ According to a certain graph partitioning algorithm, find the hyperedge with the smallest weight and disconnect it, so as to divide the hypergraph into two parts, each part as a cluster (class); ④ Repeat the above partitioning until the partitioned cluster reaches a certain value, or the quality of the new partition generated is lower than the preset threshold. In general, the key idea of ​​the clustering algorithm based on hypergraph partitioning is to convert the data processing problem in the high-dimensional data space into a graph partitioning problem, and seek the clustering of high-dimensional data by constructing the minimum spanning tree of a specific hypergraph. The time complexity of this method is only O(n*d*k), where n is the size of the data set, d is the dimension of the data points, and k is the number of clusters. For different application fields and backgrounds, researchers have also proposed many hypergraph-based clustering methods.

[0077] Subspace clustering: Subspace clustering is also called feature selection. It divides the original data space into different subspaces and examines the existence of clusters only on those relevant subspaces. These algorithms generally use search methods such as greedy strategies to search different feature subspaces, and then use some criteria to evaluate these subspaces to find the required clusters. Typical subspace clustering algorithms include CLIQUE, ENCLUS, and MAFIA. This type of algorithm uses the Apriori strategy to find and merge grids whose metrics are greater than a given threshold, generate candidate subspaces, and sort these candidate subspaces according to their coverage, that is, the number of points in the subspace; then use the minimum description length criterion to prune the subspaces with lower scale.

[0078] (2) Performance analysis of high-dimensional clustering algorithms

[0079] The high-dimensional clustering method based on dimensionality reduction is one of the most intuitive methods for clustering high-dimensional data. Its advantages are easy to understand and simple to implement, but its defects are also obvious: First, the presence of noise data in the data set is a key factor affecting the effect of dimensionality reduction clustering. In the process of mapping the original high-dimensional data to the low-dimensional space through dimensionality reduction, the distance between the noise data and the "clean" data will also be reduced, which will inevitably reduce the quality of clustering. In many application fields, it is usually difficult to remove the influence of noise data in the preprocessing process. In addition, clustering based on dimensionality reduction is fundamentally based on the distance or similarity evaluation between data. When the dimension of the data is not very high, these methods work better, but when the dimension of the data increases, clustering processing will be difficult to achieve the expected effect. The reasons are: ① It is difficult to define a distance metric in a very high-dimensional space; ② Distance-based methods usually need to calculate the mean distance between each cluster. When the dimension of the data is very high, the distance difference between different clusters will become very small.

[0080] The advantages of the hypergraph-based clustering method are mainly reflected in two aspects: ① Through this method, the calculation of similarity between high-dimensional data can be avoided during the clustering process, thereby reducing the impact of the dimensionality curse on high-dimensional clustering; ② This method can also be used to control the quality of clustering according to the needs of specific users or fields. The reason is that by using different levels of minimum support in the Apriori algorithm, the relationship between data expressed by the hypergraph model can be appropriately adjusted. Higher support values ​​correspond to higher-quality clusters with fewer data points, and lower support values ​​correspond to coarse clusters with more data points. However, it cannot be ignored that the quality of the clustering effect of this algorithm is closely related to the selection of corresponding parameters. First, when looking for frequent sets, the determination of the support level is closely related to the specific application field; second, for continuous variables, they must be discretized before the algorithm can be applied for processing. The discretization of continuous attributes will inevitably lead to the loss of certain relationships between data, which will make the clustering results deviate greatly from the actual situation.

[0081] Subspace clustering is somewhat similar to clustering based on dimensionality reduction, but the latter preprocesses high-dimensional data by direct dimensionality reduction, that is, clustering is performed in a specific low-dimensional space after dimensionality reduction; while the former divides high-dimensional data into several different subspaces, and then seeks data clustering in different subspaces as needed. Using the idea of ​​subspace clustering, data clustering can be sought from multiple angles and by comprehensively considering multiple attributes. However, in this type of algorithm, the division and selection of subspaces is also an important issue worthy of in-depth study. If there are too many subspace divisions, not only will the computational complexity be very large, but the clustering results will also be too complicated; if the subspace division is too rough, the influence of the dimensionality curse on clustering cannot be well avoided.

[0082] S5.Under unsupervised conditions, high-dimensional EEG signal decoding and target classification based on high-dimensional spectral clustering are used to explore the potential characteristics of EEG signals under different visual stimuli in order to accurately identify EEG signals under different stimuli.

[0083] According to the characteristics of key EEG information under different visual stimuli, the present invention proposes an unsupervised EEG information decoding and target classification method based on high-dimensional spectral clustering, and explores the potential characteristics of EEG signals under different visual stimuli, in order to accurately identify EEG signals under different stimuli. At present, graph theory has been proved to be a very useful tool for solving important combinatorial problems in the fields of geometry, number theory, operations research and optimization. In recent years, with the popularization of computers and the increase in the scale of integrated circuits, the research on hypergraph theory and its application has received more and more attention. Spectral clustering is a method of clustering using the principle of graph segmentation. It converts the data clustering problem into a graph partitioning problem for solution, which is particularly suitable for the case where the data set is non-convex. Each point in the data set is used as a vertex of the graph, and the similarity value between any two points is used as the weight of the edge connecting the two vertices, so that an undirected weighted graph is constructed. According to the objective function, the graph can be divided into several unconnected subgraphs, the connection weight within the subgraph is maximized, and the connection weight between each subgraph is minimized. The point set contained in the subgraph is the cluster generated after clustering. Spectral clustering theory provides a new approach to solving clustering problems, and it can effectively handle many practical problems faced in unsupervised learning.

[0084] In view of the fact that the traditional spectral clustering method performs well on some low-dimensional and small-scale data sets, it is easily disturbed by noise and irrelevant attributes in the application scenarios of high-dimensional data. In order to make full use of the information contained in each attribute and strengthen the role of important attributes, the present invention designs an attribute weighted spectral clustering algorithm based on knowledge entropy. The algorithm uses the relevant concept of knowledge entropy in the rough set to evaluate the importance of each EEG feature, and then assigns it as a weight to the corresponding feature, and then applies the spectral clustering method to cluster the data points. The algorithm can better handle high-dimensional data and has strong robustness and generalization ability. The basic idea of ​​the spectral clustering algorithm is to regard the data points as nodes in the graph, and then establish edges between the nodes according to the similarity between the data points; then divide the data into different clusters by cutting the graph. The specific steps of the algorithm are as follows:

[0085] 1) Construct a similarity matrix: Construct a similarity matrix based on the data, where each element in the matrix represents the similarity between two data points. Commonly used similarity metrics include Gaussian kernel function and k-nearest neighbor.

[0086] 2) Graph Construction: A graph is constructed based on the similarity matrix, where each data point represents a node and the weight of each edge represents the similarity between two nodes.

[0087] 3) Calculate the Laplacian matrix: Calculate the Laplacian matrix of the graph. The Laplacian matrix can be used to represent the structural information of the graph.

[0088] 4) Calculate eigenvalues ​​and eigenvectors: Calculate the eigenvalues ​​and eigenvectors of the Laplacian matrix, and select the eigenvectors corresponding to the first k smallest eigenvalues.

[0089] 5) Clustering: Use K-Means or other clustering algorithms to cluster the feature vectors.

[0090] S6. Based on the EEG signal clustering result obtained in the above step S5, the recognition effect is evaluated.

[0091] The attribute weighted spectral clustering algorithm based on knowledge entropy proposed in this invention is used to perform cluster analysis on EEG signals under complex image visual stimulation. The cluster recognition results are shown in Figure 2. Figure 8 As shown in the figure, the left figure shows the real distribution of real EEG data under target stimulation and non-target stimulation after mapping in two-dimensional space, and the right figure shows the clustering recognition effect of EEG signals under different types of visual stimulation using the proposed clustering recognition algorithm. It can be seen that the method proposed in the present invention can realize effective clustering recognition of EEG signals under different visual stimulations.

[0092] In order to further evaluate the effectiveness of the proposed method, the present invention analyzes common evaluation indicators of clustering and performs quantitative evaluation on the proposed method. Clustering, as an unsupervised learning algorithm, has no real value to verify the results. This makes it complicated to evaluate the quality of clustering. However, commonly used indicators such as silhouette coefficient, Dunn index, Calinski-Harabasz index, etc. can be used to evaluate the performance of clustering models. Different indicators are defined as follows:

[0093] Silhouette Coefficient: The silhouette coefficient measures the cohesion within a cluster and the separation between clusters. Its value is between -1 and 1, and the larger the value, the better the clustering quality. The silhouette coefficient is a commonly used clustering evaluation indicator.

[0094] Dunn's Index: Dunn's Index measures the compactness within a cluster and the separation between clusters. The larger its value, the better the clustering quality.

[0095] CH Index (Calinski-Harabasz Index): The Calinski-Harabasz Index is also called the variance ratio criterion. It measures the compactness within a cluster and the separation between clusters. The larger its value, the better the clustering quality.

[0096] In addition, if the true clustering labels of the data are already known, then indicators such as Adjusted Rand Index (ARD), Adjusted Mutual Information (AMI), Homogeneity (Homo), Completeness (Comp) and V-measure (V_m) can be used to evaluate the effectiveness of the clustering algorithm. They can all be used to measure the consistency between the clustering results and the true labels. The larger the value, the better the clustering effect.

[0097] The clustering effect of EEG signals of different subjects is evaluated by using the high-dimensional EEG clustering decoding method under complex visual stimulation of the present invention, as shown in the following table:

[0098] Table 1 Evaluation table of clustering effect of EEG signals of different subjects

[0099]

[0100] Parts not described in detail in the present invention belong to common knowledge of those skilled in the art.

Claims

1. A high-dimensional EEG clustering decoding method under complex visual stimulation, characterized in that: The following steps are involved: S1. Experimental paradigm design of EEG signal acquisition under rapid sequence visual presentation; In the rapid serial visual presentation experiment, the subject's task is to pay attention to the target picture and press a button as soon as possible after it appears. By analyzing the subject's EEG response, it is determined whether the subject has noticed the target picture, thereby achieving information transmission. The experimental setting parameters include stimulation type, stimulation duration, stimulation interval, target stimulation, target stimulation frequency, and response method. S2. The experiment starts, the EEG signal acquisition system acquires EEG signals and processes the signals; S3. further processing the EEG signals in the above step S2 to obtain high-quality EEG signals from a sufficient number of subjects to form an EEG signal data set under rapid sequence visual presentation; S4. performing cluster analysis on the high-dimensional EEG signal data formed in the above step S3, including a step of formulating a high-dimensional clustering strategy and a step of analyzing the performance of a high-dimensional clustering algorithm; S5.Under unsupervised conditions, high-dimensional EEG signal decoding and target classification based on high-dimensional spectral clustering are used to mine the potential characteristics of EEG signals under different visual stimuli in order to accurately identify EEG signals under different stimuli; S6. Based on the EEG signal clustering result obtained in the above step S5, the recognition effect is evaluated.

2. The method according to claim 1, characterized in that In step S1, the rapid serial visual presentation experimental paradigm and parameters with a presentation frequency of 12.5 Hz are as follows: the visual stimulus image is selected according to the task requirements; the stimulus duration is set to 80 ms; the stimulus interval is set to 0; the target stimulus frequency is set to 5%; and the response mode is set to the subject counting the number of targets.

3. The method according to claim 2, characterized in that In this experiment, the subjects wore an electrode cap and sat 1 meter away from a 1920×1080 resolution monitor. A fixed cross mark was added in the middle of the visual stimulus image as a fixation point. The subjects pressed the space bar on the keyboard to start viewing the image sequence. Each experimental stage had 10 sequences, each sequence contained 200 pictures, and was repeated 5 rounds; the interval between two sequences was controlled by the subjects, and pressing the space bar started the presentation of the next sequence of images.

4. The method according to claim 3, characterized in that In step S2, the EEG signal acquisition system includes an amplifier circuit, a filter circuit and an analog-to-digital converter; the amplifier circuit includes three stages of amplification, namely, pre-amplification, intermediate stage amplification and final stage amplification, and a filter circuit is interspersed in the three-stage amplifier circuit. The pre-amplification adopts an instrument amplifier and a common-phase parallel differential amplifier.

5. The method according to claim 4, characterized in that The step S3 includes a reference electrode resetting step and a filtering step; in the reference electrode resetting step, the reference electrode is reset with the left mastoid channel M1 and the right mastoid channel M2 as reference; in the filtering step, a cutoff frequency is set, the sampling frequency is set to at least 3 times the cutoff frequency, and a low-pass filter is selected.

6. The method according to claim 5, characterized in that The cut-off frequency is 0.3 Hz, and a 65 Hz band-pass FIR digital filter is selected.

7. The method according to claim 6, characterized in that: In step S4, the high-dimensional clustering strategy is selected from one of dimensionality reduction clustering, hypergraph-based clustering, subspace clustering and joint clustering.

8. The method according to claim 7, characterized in that In step S5, an attribute weighted spectral clustering algorithm based on knowledge entropy is adopted. The algorithm uses the relevant concept of knowledge entropy in rough set to evaluate the importance of each EEG feature, and then assigns it as a weight to the corresponding feature, and then applies the spectral clustering method to cluster the data points.

9. The method according to claim 8, characterized in that The specific steps of the above algorithm are as follows: S51. Construct a similarity matrix: construct a similarity matrix based on the data, where each element in the matrix represents the similarity between two data points. Commonly used similarity metrics include Gaussian kernel function and k nearest neighbors, where k is the number of clusters; S52. Construct a graph: construct a graph according to the similarity matrix, where each data point represents a node and the weight of each edge represents the similarity between two nodes; S53. Calculate the Laplacian matrix: Calculate the Laplacian matrix of the graph; the Laplacian matrix can be used to represent the structural information of the graph; S54. Calculate eigenvalues ​​and eigenvectors: Calculate the eigenvalues ​​and eigenvectors of the Laplace matrix, and select the eigenvectors corresponding to the first k smallest eigenvalues; S55. Clustering: Use K-Means algorithm to cluster feature vectors.

10. The method according to claim 9, characterized in that In the evaluation, the silhouette coefficient, Dunn index, and CH index are introduced; if the true clustering label of the data is already known, one or more of the adjusted Rand index, adjusted mutual information, homogeneity, completeness, and V-measure score are introduced in the review.

Images