SSVEP-BCI device based on frequency space multiplexing and target recognition method thereof

Abstract

The present invention discloses a SSVEP-BCI device based on frequency space multiplexing and a target recognition method thereof. The SSVEP-BCI device of the present invention comprises N array-shaped F The stimulation interface is composed of eight intermediate units, each of which is composed of two rows and two columns of minimum units. Each minimum unit is a stimulation module that flashes at a single frequency. The stimulation frequency within the intermediate units is the same, and the difference in stimulation frequency between any adjacent intermediate units is the same. Within the effective area of the stimulation interface, the eight neighboring minimum units of any target minimum unit form differentiated competitive stimulation neighbors. This invention is based on the multi-target SSVEP-BCI paradigm of frequency-spatial multiplexing, utilizing a certain number of available frequencies to achieve effective stimulation exceeding the number of frequencies. Based on frequency-spatial multiplexing within a limited space, it can achieve more separable response signals, thereby improving interaction efficiency.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and in particular to a SSVEP-BCI device based on frequency space multiplexing and a target recognition method thereof. Background Art

[0002] Brain-computer interface (BCI) technology is a communication system that allows the human brain to communicate directly with the external world. It extracts features from electroencephalogram (EEG) signals and transmits identified brain commands or information to controlled external devices. One strategy for improving BCI performance is to design effective BCI paradigms that can maintain a wider range of targets while eliciting sufficiently strong brain responses. New paradigms aim to establish a larger repertoire of targets and improve target selection efficiency. Steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (SSVEP-BCIs) offer a high information transfer rate and offer a larger repertoire than other paradigms. SSVEPs are periodic electroencephalogram (EEG) responses to visual frequency stimuli. The visual stimulus frequency range that can elicit SSVEP components can be roughly divided into three ranges: low frequency (6-12 Hz), mid-frequency (12-30 Hz), and high frequency (30-60 Hz). However, due to the harmonic nature of SSVEP, SSVEP-BCI frequency selection must avoid frequencies with multiple relationships. On the other hand, the smaller the difference between adjacent stimulation frequencies, the higher the requirements for the recognition algorithm. Therefore, the frequency band available for SSVEP-BCI is very limited. If a single frequency is used for each target, the large number of options requires a wider frequency band, which conflicts with the limited available frequencies. A new SSVEP-BCI that utilizes limited available frequencies to generate more stimulation targets has important research significance and application value for improving the performance of BCI systems. Summary of the Invention

[0003] Technical problem to be solved by the present invention: In response to the above-mentioned problems of the prior art, a SSVEP-BCI device based on frequency-space multiplexing and a target recognition method thereof are provided. The present invention is based on a multi-target SSVEP-BCI paradigm of frequency-space multiplexing, and utilizes a certain number of available frequencies to achieve effective stimulation greater than the number of frequencies. It can obtain more separable response signals based on frequency-space multiplexing within a limited space, thereby improving interaction efficiency.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A SSVEP-BCI device based on frequency space multiplexing includes N F A stimulation interface is formed by intermediate units, each intermediate unit is composed of two rows and two columns of minimum units, each minimum unit is a stimulation module flashing at a single frequency, the stimulation frequencies of the minimum units inside the intermediate unit are the same, and the stimulation frequencies of different intermediate units are different, so that in the effective area of the stimulation interface except for the outermost circle of minimum units, three neighboring minimum units with different stimulation frequencies among the eight neighboring minimum units of any target minimum unit form differentiated competitive stimulation neighbors to obtain a more separable response signal. Each target minimum unit has a different competitive stimulation neighbor, thereby realizing the distinction between different minimum units of the same frequency in each intermediate unit under the condition of frequency space multiplexing.

[0006] Optionally, the single-frequency flickering stimulation module uses a sine wave to control the sinusoidal change of brightness to present stimulation, and the sine waves of each minimum unit within the intermediate unit have the same phase, and there is a phase difference between the sine waves of adjacent different intermediate units.

[0007] Optionally, the phase difference is 0.5π.

[0008] Optionally, the stimulation frequency of each minimum unit within each intermediate unit in the stimulation interface ranges from 8 to 15.8 Hz.

[0009] Optionally, the number of intermediate units in the stimulation interface is 40, and the minimum stimulation frequency difference between the smallest units in any two intermediate units is 0.2 Hz.

[0010] The present invention also provides a target recognition method of the SSVEP-BCI device based on frequency space multiplexing, comprising:

[0011] S1, applying stimulation to the user through the frequency-space multiplexing SSVEP-BCI device and collecting the user's EEG signals synchronized with the stimulation;

[0012] S2, perform multiple frequency recognition on the EEG signal of a time segment, and calculate the EEG signal and N F The frequency response intensity of the stimulation frequency of the intermediate unit is obtained as N F The frequency response characteristics of the mid-level unit X m ;

[0013] S3: For each minimum unit, classify it according to the frequency response characteristic distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units to obtain the global response score of the minimum unit, and take the minimum unit with the highest global response score as the minimum unit that the user is looking at.

[0014] Optionally, step S2 includes: using a filter bank canonical correlation analysis algorithm and calculating the frequency response intensity of the stimulation frequency of a middle-level unit at intervals of a specified time length, thereby obtaining N F The frequency response characteristics of the mid-level unit X m .

[0015] Optionally, the specified time length is 0.2s.

[0016] Optionally, in step S3, when classifying based on the frequency response feature distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units, the classifier used is a dual-scale graph attention network model DSGAT, where the frequency response feature X of all intermediate units is input to the dual-scale graph attention network model DSGAT. m , the output is the global response score of all minimum units. Optionally, the dual-scale graph attention network model DSGAT includes two scale graphs: intermediate unit graph and minimum unit graph; the intermediate unit graph uses each intermediate unit as a node, called an intermediate node, and there is an edge between an intermediate node and the 8 neighboring intermediate nodes of its 8 neighborhoods; the minimum unit graph uses the minimum unit as a node, called a minimum node, and there is a corresponding relationship between the minimum node and the intermediate node of the intermediate unit graph, and the intermediate node of one intermediate unit graph corresponds to the 4 minimum nodes in the minimum unit graph; the operation of the dual-scale graph attention network model DSGAT on the intermediate unit graph includes 4 sub-networks with the same structure, each sub-network includes a temporal embedding layer and a multi-head graph attention layer, and the inputs of the 4 sub-networks are all the frequency response features X of all intermediate units. m After the aggregation operation of the four sub-networks, each intermediate node obtains four new features, which are respectively passed to the four minimum nodes corresponding to the minimum unit graph as the initial features of the minimum unit graph; the dual-scale graph attention network model DSGAT operates on the minimum unit graph including a fully connected layer and a softmax output layer, which is used to obtain the global response score of each minimum unit according to the initial features of the input minimum unit graph.

[0017] Compared with existing technologies, the present invention offers the following key advantages: It eliminates the need to design individual stimulus targets. Instead, it leverages the positional relationships between different stimuli to enhance the differences in properties between targets at the same frequency. This approach matches neighboring stimuli of different frequencies to targets at the same frequency, encoding targets using complex positional relationships. All visual stimuli act as targets and competing stimuli for each other. In particular, compared with frequency time-division multiplexing, frequency-space multiplexing allows for simultaneous utilization of information from different frequencies, improving interaction efficiency, without increasing stimulation duration. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1Schematic diagram of the stimulation interface of the SSVEP-BCI device according to an embodiment of the present invention.

[0019] Figure 2 This is a flow chart of the target recognition method of the SSVEP-BCI device according to an embodiment of the present invention.

[0020] Figure 3 Schematic diagram of the structure of the dual-scale graph attention network model DSGAT in an embodiment of the present invention.

[0021] Figure 4 Schematic diagram of a specific paradigm interface of the stimulation interface in an embodiment of the present invention.

[0022] Figure 5 In the embodiment of the present invention, the stimulation interface is Figure 4 Variant 1 is generated on the basic paradigm interface shown.

[0023] Figure 6 In the embodiment of the present invention, the stimulation interface is Figure 4 Variant 2 is generated on the basic paradigm interface shown. DETAILED DESCRIPTION

[0024] like Figure 1 As shown, the SSVEP-BCI device based on frequency space multiplexing in this embodiment includes N F A stimulation interface is formed by intermediate units, each intermediate unit is composed of minimum units in two rows and two columns (2×2 matrix structure), each minimum unit is a stimulation module flashing at a single frequency, the stimulation frequencies of the minimum units inside the intermediate unit are the same, and the stimulation frequencies of different intermediate units are different, so that in the effective area of the stimulation interface except for the outermost circle of minimum units, three neighboring minimum units with different stimulation frequencies among the eight neighboring minimum units of any target minimum unit form differentiated competitive stimulation neighbors to obtain a more separable response signal. Each target minimum unit has different competitive stimulation neighbors, thereby realizing the distinction between different minimum units of the same frequency in each intermediate unit under the condition of frequency-space multiplexing.

[0025] In this embodiment, the SSVEP-BCI device based on frequency-space multiplexing uses the positional relationship between different stimuli to increase the attribute differences of targets with the same frequency. Different targets with the same frequency are matched with neighboring stimuli of different frequencies. The target is encoded using a composite positional relationship, and all visual stimuli are targets and competitive stimuli for each other. In this embodiment, stimuli of different frequencies that are less than 4.5°-5° away from a target in space are called competitive stimulus neighbors of the target. When the user looks at a target, its competitive stimulus neighbors will also induce corresponding frequency components in the EEG signal; each target has differentiated competitive stimulus neighbors, and a target can be uniquely determined based on the response distribution pattern of its differentiated competitive stimulus neighbors; N F The stimulation frequency can form 4×N F The smallest unit, the outermost N O The minimum unit lacks enough differentiated neighbors and is therefore not separable. The effective target area with separability is the internal (4×N F -N O ) minimum units, and all visual stimuli are targets and competing stimuli for each other.

[0026] In this embodiment, the stimulation module with a single frequency flashing uses a sine wave to control the sinusoidal change of brightness to present stimulation, and the sine wave phases of each minimum unit inside the intermediate unit are the same, and there is a phase difference between the sine wave phases of the minimum units in adjacent different intermediate units. In this embodiment, the phase difference is 0.5π. In this embodiment, the stimulation frequency of each minimum unit inside each intermediate unit in the stimulation interface ranges from 8 to 15.8 Hz. In this embodiment, the number of intermediate units in the stimulation interface is 40 (N F =40), the minimum stimulation frequency difference between any two intermediate units is 0.2 Hz, each minimum unit is replicated into a 2×2 matrix as an intermediate unit, and then the intermediate units are tiled to form a 10×16 paradigm interface, which contains a total of 160 minimum units. Figure 1 , use c ij Indicates the jth smallest unit in the i-th intermediate unit. When looking at a smallest unit, Figure 1 Objective-C in i1 For example, the 8 smallest units closest to the center of the visual field are the neighbor stimuli that can induce a stronger SSVEP response. i1 With the same frequency target c i2 、c i3 、c i4 Competing stimuli are distinguished as Figure 1 The L-shaped area marked in i1 .

[0027] The stimulation arrangement of the present invention makes all visual stimuli mutually targeted and competitive. Each target has different competitive stimulus neighbors, so that different minimum units of the same frequency can be distinguished under the condition of frequency space multiplexing. However, since the outermost targets lack competitive stimulus neighbors, they do not have sufficient discrimination. Therefore, their role is to serve as neighbors to assist in locating internal units. Therefore, the actual effective separable targets are Figure 1 The 8×14 (112 targets) area inside the bold frame. The user adopts a target fixation method with regional attention to focus on the target he intends to select; the four smallest units in the same intermediate unit have different competitive stimulus neighbors in their direction. The regional fixation plus attention method is to fixate on the target c ij At the same time give L ij Region-selective attention for more separable response signals.

[0028] like Figure 2 As shown, this embodiment further provides a target recognition method of the aforementioned SSVEP-BCI device based on frequency space multiplexing, including:

[0029] S1, applying stimulation to the user through the frequency-space multiplexing SSVEP-BCI device and collecting the user's EEG signals synchronized with the stimulation;

[0030] S2, multiple (N T times) frequency recognition, respectively calculate the EEG signal and N F The frequency response intensity of the stimulation frequency of the intermediate unit is obtained as N F The frequency response characteristics of the mid-level unit X m ;

[0031] S3: For each minimum unit, classify it according to the frequency response characteristic distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units to obtain the global response score of the minimum unit, and take the minimum unit with the highest global response score as the minimum unit that the user is looking at.

[0032] In this embodiment, step S2 includes: using the filter bank canonical correlation analysis (FBCCA) algorithm and calculating the frequency response intensity of the stimulation frequency of a middle-level unit at a specified time interval, thereby obtaining N F The frequency response characteristics of the mid-level unit X m It should be noted that the time length specified here can be selected according to actual needs. For example, in this embodiment, the time length specified is 0.2s, and the frequency response intensity of each frequency is calculated every 0.2s to obtain the frequency response characteristics X of 40 intermediate units.m .

[0033] It should be noted that the filter bank canonical correlation analysis algorithm is an existing method. This embodiment merely applies this method and does not involve any improvement of this method. Therefore, its specific implementation details will not be described in detail here.

[0034] for Figure 1 For the 8×14 (112 targets) area within the bold frame, the classification and recognition based on the different competitive stimulus neighbors of each target minimum unit in step S3 is equivalent to 112 classification tasks. In this embodiment, when classifying according to the frequency response feature distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units in step S3, the classifier used is the dual-scale graph attention network model DSGAT (Dual Scale Graph Attention Networks), and the input of the dual-scale graph attention network model DSGAT is the frequency response feature X of all intermediate units. m , the output is the global response score of all minimum units.

[0035] In this embodiment, the dual-scale graph attention network model DSGAT designs a dual-scale graph for two types of units (intermediate units and minimum units) with two types of stimulus units as graph nodes. The dual-scale graph attention network model DSGAT includes two scale graphs: intermediate unit graph and minimum unit graph; the intermediate unit graph uses each intermediate unit as a node, called an intermediate node, and there is an edge between an intermediate node and its 8 neighboring intermediate nodes in the 8 neighborhoods; the minimum unit graph uses the minimum unit as a node, called a minimum node, and there is a corresponding relationship between the minimum node and the intermediate node of the intermediate unit graph. One intermediate node of the intermediate unit graph corresponds to 4 minimum nodes in the minimum unit graph; as shown in FIG. Figure 3 As shown in the figure, the operation of the dual-scale graph attention network model DSGAT on the intermediate unit graph includes four sub-networks with the same structure (sub-network 1 to sub-network 4). Each sub-network contains a temporal embedding layer and an attention aggregation layer implemented by a multi-head graph attention layer (Velickovic P, Cucurull G, Casanova A, et al. Graph Attention Networks, ArXiv, 2018, abs / 1710.10903); the operation of the temporal embedding layer is to convert N m Input features Perform Hadamard product operation with a learnable temporal embedding matrix to obtain the corresponding features To compensate for the time imbalance of the SSVEP detection algorithm; the input features of the 4 sub-networks They are all frequency response features of all intermediate units, and after the aggregation operation of the attention aggregation layer of the four sub-networks, each intermediate node obtains 4 sets of new features, which are respectively passed to the 4 minimum nodes corresponding to the minimum unit graph as the initial features of the minimum unit graph; the operation of the dual-scale graph attention network model DSGAT on the minimum unit graph includes a fully connected layer FC and a softmax output layer, which is used to obtain the global response score of each minimum unit based on the initial features of the input minimum unit graph. The fully connected layer FC sends the 4 sets of new features output by the 4 sub-networks to the softmax output layer, and the softmax activation function is used to activate them through the softmax output layer to obtain the global response scores of Nc minimum units.

[0036] It should be noted that the dual-scale graph attention network model DSGAT adopted in this embodiment is only an example of a specific recognition method. In essence, it is to classify according to the frequency response characteristics of the intermediate unit where the target minimum unit is located and its neighboring intermediate units to obtain the global response score of the target minimum unit, that is, to achieve the mapping between the frequency response characteristics of the intermediate unit where the target minimum unit is located and its neighboring intermediate units and the global response score of the target minimum unit. Undoubtedly, any other machine learning classifier model that can achieve the above mapping can be adopted as needed, and should not be limited to the dual-scale graph attention network model DSGAT adopted in this embodiment or its specific dual-scale structure.

[0037] In order to verify the feasibility and effectiveness of the method of this embodiment, a real EEG experiment was conducted to verify this embodiment. The EEG data was collected using a 12-lead EEG acquisition system, and the paradigm was adjusted by arranging the structure to generate two other paradigm variants. The three paradigm interfaces are as follows: Figure 4 、 Figure 5 and Figure 6 In the experiment, users followed system prompts to fixate on the corresponding minimum unit. Each paradigm was followed by a set of experiments, each consisting of 224 selection tasks. Six users (aged 23-33) participated in the validation test. A five-fold cross-validation was performed on the data from each experimental set, resulting in the recognition accuracy rates shown in the table below.

[0038] Table 1: Recognition accuracy table.

[0039] user Basic paradigm accuracy (%) Accuracy of paradigm variant 1 (%) Paradigm variant 2 accuracy (%) S1 84.35±3.23 93.91±2.38 85.65±5.67 S2 90.00±3.30 93.48±2.67 87.39±3.89 S3 88.26±4.51 86.52±3.89 84.35±4.96 S4 96.52±4.77 96.52±1.19 83.48±3.95 S5 75.12±4.82 68.26±3.64 76.52±7.90 S6 92.17±5.87 93.91±4.18 94.35±3.30 Mean ± standard deviation 87.74±7.39 88.77±10.59 85.29±13.39

[0040] As can be seen from Table 1, the target recognition method of the SSVEP-BCI device based on frequency-space multiplexing in this embodiment is proven to be feasible. Although there are individual differences in the results of different users, this phenomenon is common in brain-computer interface interaction systems. In the verification test of the target recognition method of the SSVEP-BCI device based on frequency-space multiplexing in this embodiment, the degree of adaptation of different users to the target gaze method accompanied by regional attention will also have a certain impact on the test results. Overall, the real EEG experimental results verify the feasibility and effectiveness of the target recognition method of the SSVEP-BCI device based on frequency-space multiplexing in this embodiment.

[0041] In summary, the target recognition method of the SSVEP-BCI device based on frequency-space multiplexing in this embodiment does not require the design of a single stimulus target, but rather utilizes the positional relationship between different stimuli to increase the attribute differences of targets at the same frequency, matches different targets at the same frequency with neighboring stimuli of different frequencies, and utilizes the composite positional relationship to encode the target, so that all visual stimuli are mutually targeted and competitive stimuli. In particular, compared with the frequency time-division multiplexing method, there is no need to increase the stimulation time at the expense of frequency-space multiplexing. The frequency-space multiplexing method can simultaneously utilize different frequency information to improve interaction efficiency. The method of this embodiment utilizes the principle of competitive neuronal dynamics in the human brain's visual processing cortical network to increase the number of targets in the steady-state visual evoked potential brain-computer interface paradigm through frequency-space multiplexing. Unlike previous frequency multiplexing methods, the present invention does not design a single stimulus target, but instead uses the positional relationship between different stimuli to increase the attribute differences of targets at the same frequency. By designing the visual stimulus arrangement so that the same frequency target matches different competitive stimulus neighbors, a "neighbor coding" is formed, and all visual stimuli are mutually targeted and competitive stimuli, thereby making full use of the available frequency and realizing a brain-computer interface system with more targets.

[0042] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including the instruction device, which implements the function specified in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0043] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A SSVEP-BCI device based on frequency space multiplexing, characterized in that: Including array arrangement A stimulation interface is formed by intermediate units, each intermediate unit is composed of two rows and two columns of minimum units, each minimum unit is a stimulation module flashing at a single frequency, the stimulation frequencies of the minimum units inside the intermediate unit are the same, and the stimulation frequencies of different intermediate units are different, so that in the effective area of the stimulation interface except for the outermost circle of minimum units, three neighboring minimum units with different stimulation frequencies among the eight neighboring minimum units of any target minimum unit form differentiated competitive stimulation neighbors to obtain a more separable response signal. Each target minimum unit has a different competitive stimulation neighbor, thereby realizing the distinction between different minimum units of the same frequency in each intermediate unit under the condition of frequency space multiplexing.

2. The SSVEP-BCI device based on frequency space multiplexing according to claim 1, characterized in that: The single-frequency flickering stimulation module uses a sine wave to control the sinusoidal change of brightness to present stimulation, and the sine wave phases of each minimum unit within the intermediate unit are the same, and there is a phase difference between the sine waves of adjacent different intermediate units.

3. The SSVEP-BCI device based on frequency space multiplexing according to claim 2, characterized in that: The phase difference is 0.5π.

4. The SSVEP-BCI device based on frequency space multiplexing according to claim 1, characterized in that The stimulation frequency of each minimum unit within each intermediate unit in the stimulation interface ranges from 8 to 15.8 Hz.

5. The SSVEP-BCI device based on frequency space multiplexing according to claim 4, characterized in that The number of intermediate units in the stimulation interface is 40, and the minimum stimulation frequency difference between the smallest units in any two intermediate units is 0.2 Hz.

6. A target recognition method of a SSVEP-BCI device based on frequency space multiplexing according to any one of claims 1 to 5, characterized in that: include: S1, applying stimulation to a user by means of the frequency-space multiplexing SSVEP-BCI device according to any one of claims 1 to 5, and collecting EEG signals of the user synchronized with the stimulation; S2, perform multiple frequency recognition on the EEG signal of a time segment, and calculate the EEG signal and The frequency response intensity of the stimulation frequency of the intermediate unit is obtained Frequency response characteristics of the mid-level unit ; S3: For each minimum unit, classify it according to the frequency response characteristic distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units to obtain the global response score of the minimum unit, and take the minimum unit with the highest global response score as the minimum unit that the user is looking at.

7. The target recognition method of the SSVEP-BCI device based on frequency space multiplexing according to claim 6, characterized in that: Step S2 includes: using the filter bank canonical correlation analysis algorithm and calculating the frequency response intensity of the stimulus frequency of a middle-level unit at a specified time interval, thereby obtaining Frequency response characteristics of the mid-level unit .

8. The target recognition method of the SSVEP-BCI device based on frequency space multiplexing according to claim 7, characterized in that: The specified time length is 0.2s.

9. The target recognition method of the SSVEP-BCI device based on frequency space multiplexing according to claim 6, characterized in that: In step S3, when classifying based on the frequency response feature distribution of the intermediate unit where the minimum unit is located and its neighboring intermediate units, the classifier used is the dual-scale graph attention network model DSGAT, and the input of the dual-scale graph attention network model DSGAT is the frequency response feature of all intermediate units. , the output is the global response score of all minimum units.

10. The target recognition method of the SSVEP-BCI device based on frequency space multiplexing according to claim 9, characterized in that: The dual-scale graph attention network model DSGAT includes graphs of two scales: intermediate unit graph and minimum unit graph; the intermediate unit graph uses each intermediate unit as a node, called an intermediate node, and there is an edge between an intermediate node and the 8 neighboring intermediate nodes of its 8 neighborhoods; the minimum unit graph uses the minimum unit as a node, called a minimum node, and there is a corresponding relationship between the minimum node and the intermediate node of the intermediate unit graph, and an intermediate node of an intermediate unit graph corresponds to 4 minimum nodes in the minimum unit graph; the operation of the dual-scale graph attention network model DSGAT on the intermediate unit graph includes 4 sub-networks with the same structure, each sub-network includes a temporal embedding layer and a multi-head graph attention layer, and the inputs of the 4 sub-networks are all frequency response characteristics of all intermediate units , and after the aggregation operation of the four sub-networks, each intermediate node obtains four new features, which are respectively passed to the four minimum nodes corresponding to the minimum unit graph as the initial features of the minimum unit graph; The dual-scale graph attention network model DSGAT operates on the minimum unit graph including a fully connected layer and a softmax output layer, which is used to obtain the global response score of each minimum unit according to the initial features of the input minimum unit graph.

Images