A brain-computer interface auditory spatial attention detection system driven by a kind of auditory system

Abstract

The application discloses a kind of brain-computer interface auditory space attention detection system driven by auditory system, including: signal acquisition module, signal processing module and integration module;The signal acquisition module is used to collect electroencephalogram and the electroencephalogram is preprocessed, obtains electroencephalogram discrimination window;The signal processing module is used to decode the electroencephalogram discrimination window, obtains time characteristic and spatial characteristic;The integration module is used to integrate the time characteristic and the spatial characteristic, obtains attention detection result.The application can simulate the information coding mechanism of human auditory pathway to sound signal and the information processing mechanism of left and right brain cortex to coded information, solve the weak explainability and poor real-time of current auditory attention detection model, so as to realize low delay, short decision window under high recognition accuracy, provide effective method for developing cognitive hearing aid based on brain-computer interface.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of human-computer interaction science and brain-computer interface, and particularly relates to a brain-computer interface auditory spatial attention detection system driven by an auditory system. BACKGROUND

[0002] In a "cocktail party" scene, that is, in a complex acoustic environment with multiple sound sources, people can focus on and amplify a certain sound source by invoking the attention mechanism of the auditory system, thereby ignoring other non-attended sound sources, achieving the effect of filtering noise and resisting interference.

[0003] However, the hearing-impaired and some elderly people have difficulty filtering noise interference in a noisy environment, making them unable to communicate and perceive like normal people, and often can only be assisted and rehabilitated by wearing hearing aids. However, traditional hearing aids can only amplify all sound sources in the environment indiscriminately, and even with noise reduction technology and methods to improve speech intelligibility, the problem of poor device experience and low acceptance cannot be improved. Therefore, scholars have proposed a "cognitive hearing aid", which can decode auditory attention from the brain, obtain the sound source information that the user is focusing on, and amplify the specified sound source, thereby realizing intelligent auditory assistance, enabling the hearing-impaired to communicate and perceive in a complex acoustic environment like normal people. Traditional auditory attention decoding technology relies on the correlation between the sound source and the electroencephalogram signal, and by constructing a mapping model, the sound modality is converted into the electroencephalogram modality or the electroencephalogram modality is converted into the sound modality, and then the correlation between the converted modality data and the modality acquisition data is calculated to obtain the information of auditory attention. However, due to the complex nonlinear relationship between the real brain cortex signal and the original sound information, the linear mapping model is difficult to reproduce the complex modality conversion process of the brain, resulting in problems such as weak correlation, poor real-time performance, and low detection accuracy of traditional methods. With the continuous development of neural network algorithms, in recent years, non-linear decoding models based on deep learning (such as auditory spatial attention detection based on convolutional neural networks and auditory attention detection based on long short-term memory networks) have further improved the detection accuracy, but due to the simple design of existing auditory attention decoding models, the physiological mechanisms in the brain cognitive process are ignored, and there is a lack of explainability, resulting in the "black box problem" of low performance.

[0004] Under this background, in order to improve the explainability of the auditory spatial attention detection model, improve the real-time performance of the model, and solve the "black box problem" of the existing decoding model, it is urgent to propose an auditory spatial attention detection method for brain-computer interface driven by an auditory system. SUMMARY

[0005] The application aims to solve the problems in the prior art and provides an auditory spatial attention detection system of a brain-computer interface driven by an auditory system, which is characterized in that the auditory spatial attention detection system of the brain-computer interface driven by the auditory system comprises a signal acquisition module, a signal processing module and an integration module.

[0006] To achieve the above object, the application provides the following scheme.

[0007] An auditory spatial attention detection system of a brain-computer interface driven by an auditory system, comprising a signal acquisition module, a signal processing module and an integration module.

[0008] The signal acquisition module is configured to acquire an electroencephalogram (EEG) signal and pre-process the EEG signal to obtain an EEG signal discrimination window.

[0009] The signal processing module is configured to decode the EEG signal discrimination window to obtain time characteristics and spatial characteristics.

[0010] The integration module is configured to integrate the time characteristics and the spatial characteristics to obtain an attention detection result.

[0011] Preferably, the signal acquisition module comprises an acquisition unit and a pre-processing unit.

[0012] The acquisition unit is configured to acquire the EEG signal of a brain-computer interface operator in real time.

[0013] The pre-processing unit is configured to pre-process the EEG signal to obtain the EEG signal discrimination window.

[0014] Preferably, the working process of the pre-processing unit comprises band-pass filtering, downsampling, common average reference and independent component analysis to filter out artifacts, and then the EEG signal is segmented into the EEG signal discrimination window with a preset window length.

[0015] Preferably, the signal processing module comprises a left-brain time information decoder and a right-brain spatial information decoder.

[0016] The left-brain time information decoder is configured to decode the EEG signal discrimination window to obtain the time characteristics.

[0017] The right-brain spatial information decoder is configured to decode the EEG signal discrimination window to obtain the spatial characteristics.

[0018] Preferably, the working process of the left brain time information decoder comprises:

[0019] The brain electrical information discrimination window is input into a left brain information attention module, weight initialization is performed on a channel mask of left brain information, and a first initialization mask of left brain channel attention is obtained;

[0020] Three-dimensional convolution is performed on the first initialization mask, and a first three-dimensional feature of the first initialization mask is obtained;

[0021] The first three-dimensional feature is input into a linear layer and normalized, and a left brain channel mask of the left brain information is obtained;

[0022] Inner product operation is performed on the left brain channel mask and the brain electrical information discrimination window, and a left brain brain electrical information discrimination window of left brain channel attention is obtained;

[0023] The left brain brain electrical information discrimination window is input into a unidirectional gate recurrent unit for decoding, and the time feature is obtained.

[0024] Preferably, the working process of the right brain spatial information decoder comprises:

[0025] The brain electrical information discrimination window is input into a right brain information attention module, weight initialization is performed on a channel mask of right brain information, and a second initialization mask of right brain channel attention is obtained;

[0026] Three-dimensional convolution is performed on the second initialization mask, and a second three-dimensional feature of the second initialization mask is obtained;

[0027] The second three-dimensional feature is input into a linear layer and normalized, and a right brain channel mask of the right brain information is obtained;

[0028] Inner product operation is performed on the right brain channel mask and the brain electrical information discrimination window, and a right brain brain electrical information discrimination window of right brain channel attention is obtained;

[0029] The right brain brain electrical information discrimination window is frequency filtered, and a filtered brain electrical feature is obtained;

[0030] The filtered brain electrical feature is transposed and spatially filtered, and the spatial feature is obtained.

[0031] Preferably, the integration module comprises a splicing unit, an extraction unit and a fusion unit;

[0032] The splicing unit is configured to splice and fuse the time feature and the spatial feature in the last dimension to obtain a fusion feature of auditory spatial attention;

[0033] The extraction unit is configured to perform feature extraction on the fusion feature to obtain a compressed feature;

[0034] The fusion unit is configured to perform spatio-temporal information fusion on the compressed features to obtain the attention detection result.

[0035] Preferably, the extraction unit comprises a first convolutional layer, a first max-pooling layer, a second convolutional layer and a second max-pooling layer.

[0036] The fusion unit comprises a first linear layer, a second linear layer and an activation function.

[0037] Compared with the prior art, the present application has the following advantages:

[0038] The present application can simulate the information encoding mechanism of the human auditory pathway for sound signals and the processing mechanism of the left and right brain cortex for the encoded information, and solve the weak interpretability and poor real-time performance of the current auditory attention detection model, thereby achieving high recognition accuracy under low delay and short decision window, and providing an effective method for developing a cognitive hearing aid based on a brain-computer interface. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings described in the following embodiments are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 The figure is a schematic diagram of the system structure of the embodiment of the present application.

[0041] Figure 2 The figure is a schematic diagram of the system running process of the embodiment of the present application.

[0042] Figure 3 The figure is a schematic diagram of the structure of the signal processing module and the integration module of the embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0045] In the present embodiment, as Figure 1 , Figure 2 ,Figure 3 The application discloses an aural system driven brain-computer interface auditory space attention detection system.

[0046] The signal acquisition module is configured to acquire and pre-process the EEG signals to obtain EEG signal discrimination windows.

[0047] The signal acquisition module comprises an acquisition unit and a preprocessing unit. The acquisition unit is configured to acquire real-time EEG signals of a brain-computer interface operator. In this embodiment, the acquisition unit can be an EEG acquisition device, and the sampling frequency is set to 1000 Hz to acquire 64-channel EEG signals of the brain-computer interface operator. The preprocessing unit is configured to pre-process the EEG signals to obtain EEG signal discrimination windows. The working process of the preprocessing unit comprises the following steps: band-pass filtering, downsampling, common average reference, and independent component analysis to remove artifacts, and then segmenting the EEG signals into EEG signal discrimination windows with a preset window length. In this embodiment, in order to reduce the influence of interference components such as line noise in the original EEG signals, the EEG signals are band-pass filtered. In this embodiment, a FIR filter is used to band-pass filter the EEG signals, and the lower limit cutoff frequency is 1 Hz and the upper limit cutoff frequency is 32 Hz. In order to reduce the calculation amount, the EEG signals are downsampled to 128 Hz, and the EEG signals are truncated to EEG discrimination windows with a preset window length of 1 s for subsequent analysis, that is, each EEG discrimination window contains 128x1=128 sampling points. Then, the common average reference is used to remove the interference signals commonly existing in all channels. Specifically, the EEG signals of each channel are subtracted from the average value of all channels at the same time. Finally, during the process of wearing the EEG acquisition cap, the user will blink from time to time, which will generate strong bioelectric interference near the forehead electrode, which is called electrooculogram artifact. Moreover, due to the frequency characteristics of the waveform, band-pass filtering cannot effectively remove the electrooculogram artifact. Therefore, the independent component analysis is used to remove the electrooculogram artifact in this embodiment. Specifically, the EEG signals and the electrooculogram artifact are regarded as two different kinds of signal sources, the signal source separation is realized through independent component analysis, and mathematically, the C-channel EEG signals are projected into C independent signal sources through the calculation of the demixing matrix W. Then, the signal source corresponding to the electrooculogram artifact is determined according to the projection vector (the projection vector of the electrooculogram artifact has a large component near the forehead) and the activity amplitude (the waveform of the electrooculogram artifact is a sharp oscillation with a large peak amplitude) of each signal source, and the signal source is set to 0. Subsequently, the original EEG signals are projected back through inverse transformation, and the calculation method is as follows:

[0048] Y(t))=W*X(t)

[0049] X1(t)=W -1 *Y1(t)

[0050] In the formula, X(t) is the brain electrical signal of the C channel, W is the demixing matrix, Y(t) is the decomposed C independent components, Y1(t) is the independent component after the artifact is set to 0, and X1(t) is the brain electrical signal after the artifact is removed.

[0051] The signal processing module is configured to decode the brain electrical signal discrimination window to obtain the time feature and the space feature. The signal processing module comprises a left brain time information decoder and a right brain space information decoder.

[0052] The left brain time information decoder is configured to decode the brain electrical information discrimination window to obtain the time feature. The working process of the left brain time information decoder comprises: inputting the brain electrical information discrimination window into the left brain information attention module, initializing the weight of the channel mask of the left brain information to obtain a first initialization mask of the left brain channel attention; performing three-dimensional convolution on the first initialization mask to obtain a first three-dimensional feature of the first initialization mask; inputting the first three-dimensional feature into a linear layer and performing normalization to obtain the left brain channel mask of the left brain information; performing inner product operation on the left brain channel mask and the brain electrical information discrimination window to obtain the left brain electrical information discrimination window of the left brain channel attention; and inputting the left brain electrical information discrimination window into a unidirectional gated recurrent unit for decoding to obtain the time feature.

[0053] Specifically, (1) initialization of the weight of the channel mask of the left brain information:

[0054] In order to better extract the time information contained in the left brain cortex brain electrical signal in the embodiment, first, different initialization weights are given to different channels of the brain electrical signal from the left brain: a higher weight is given to the left brain area channel of the mask of the brain electrical signal input into the time decoder.

[0055] (2) convolution layer and linear layer:

[0056] In order to better obtain the contribution degree of different channels to the decoding of auditory attention, the initialization weight is input into the convolution layer and the linear layer to obtain the channel mask of the left brain information, and the calculation method is:

[0057] M c =elu(Conv(M0))

[0058] M l =elu(w2(elu(w1*M c +b1)+b2)

[0059] Wherein, M0∈R C*TFor initialization of the channel mask of the weight, C is the number of channels of the electroencephalogram signal, T is the number of time sampling points of the electroencephalogram information discrimination window; Conv(·) is a convolution module, in which the number of input feature map channels is 1, the number of output feature map channels is 1, and the convolution kernel size is 1*N*1; elu(·) is an exponential linear unit, which is used as an activation function in an artificial neural network; w1 and w2 are weight parameters of linear layer one and linear layer two, and b1 and b2 are bias parameters of linear layer one and linear layer two; M l ∈R C*1 is the output left brain information channel mask.

[0060] (3) Normalization

[0061] In order to map the weight value of the output left brain information channel mask to the space of (0, 1), the channel mask is normalized, and the calculation method is:

[0062]

[0063] In which, max(·) is a function for finding the maximum value of a vector, min(·) is a function for finding the minimum value of a vector, M left ∈R C*1 is the normalized left brain information channel mask.

[0064] (4) Time information decoding:

[0065] The output mask is used to give different weights to different channels of the left brain electroencephalogram signal discrimination window, so that the electroencephalogram data from the left brain channel is more used for time information decoding, and the calculation method is:

[0066]

[0067] In which, E∈R C*T is the preprocessed electroencephalogram information discrimination window, C is the number of channels of the electroencephalogram signal, and T is the number of time sampling points of the electroencephalogram discrimination window; is a dot product operation, which broadcasts M left to the dimension of C*T, and then performs dot product operation with the electroencephalogram information discrimination window; E left ∈R C*T is the left brain electroencephalogram information discrimination window with left brain channel attention. The left brain electroencephalogram information discrimination window with left brain channel attention is extracted by a recurrent neural network module to obtain the time feature of auditory spatial attention decoding, and the calculation method is:

[0068] F t =GRU(E left )

[0069] In which, GRU(·) is a unidirectional gated recurrent unit, the input size is C, the number of channels of the electroencephalogram signal, and the number of hidden layers is n.h ;F t is the time feature of the output.

[0070] The right brain spatial information decoder is used for decoding the electroencephalogram information discrimination window to obtain the spatial feature. The working process of the right brain spatial information decoder includes: inputting the electroencephalogram information discrimination window into the right brain information attention module, initializing the weight of the channel mask of the right brain information to obtain a second initialized mask of the right brain channel attention; performing three-dimensional convolution on the second initialized mask to obtain a second three-dimensional feature of the second initialized mask; inputting the second three-dimensional feature into a linear layer and performing normalization to obtain the right brain channel mask of the right brain information; performing inner product operation on the right brain channel mask and the electroencephalogram information discrimination window to obtain the right brain electroencephalogram information discrimination window of the right brain channel attention; performing frequency filtering on the right brain electroencephalogram information discrimination window to obtain a filtered electroencephalogram feature; transposing the filtered electroencephalogram feature and performing spatial filtering to obtain the spatial feature.

[0071] Specifically:

[0072] (1) Channel mask weight initialization of the right brain information:

[0073] In order to better extract the time information contained in the right brain cortex electroencephalogram signal in the embodiment, different initialization weights are first given to different channels of the electroencephalogram signal from the right brain: a higher weight is given to the right brain area channel of the mask of the electroencephalogram signal input into the time decoder.

[0074] (2) Convolution layer and linear layer:

[0075] In order to better obtain the contribution degree of different channels to the decoding of auditory attention, the initialization weight is input into the convolution layer and the linear layer to obtain the channel mask of the right brain information, and the calculation method is:

[0076] M c = elu(Conv(M0))

[0077] M l = elu(w2(elu(w1*M c +b1)+b2)

[0078] Wherein, M0∈R C*T is the channel mask of the initialization weight, C is the channel number of the electroencephalogram signal, and T is the time sampling point number of the electroencephalogram information discrimination window; Conv(·) is a convolution module, wherein the input feature map channel number is 1, the output feature map channel number is 1, and the convolution kernel size is 1*N*1; elu(·) is an exponential linear unit, which is used as an activation function in an artificial neural network; w1 and w2 are weight parameters of linear layer one and linear layer two, b1 and b2 are bias parameters of linear layer one and linear layer two; M l ∈R C*1Masking the output right brain information channel.

[0079] (3) Normalization:

[0080] In order to map the weight value of the output right brain information channel mask to the space of (0, 1), the channel mask is normalized, and the calculation method is:

[0081]

[0082] Where, max(·) is the function of finding the maximum value of the vector, min(·) is the function of finding the minimum value of the vector, M right ∈R C*1 is the normalized right brain information channel mask.

[0083] (4) Spatial information decoding:

[0084] Using the mask output by the right brain attention module, different weights are given to different channels of the right brain electroencephalogram information discrimination window, so that more electroencephalogram data from the right brain channel is used for spatial information decoding, and the calculation method is:

[0085]

[0086] Where, E∈R C*T is the preprocessed electroencephalogram information discrimination window, C is the number of electroencephalogram signal channels, and T is the time sampling point number of the electroencephalogram discrimination window. is the dot product operation, which uses the broadcast mechanism to broadcast M right to the dimension of C*T, and then performs dot product operation with the electroencephalogram information discrimination window; E right ∈R C*T is the right brain electroencephalogram information discrimination window with right brain channel attention. Using the convolutional neural network module, the right brain electroencephalogram information discrimination window with right brain channel attention is extracted to obtain the spatial features of auditory spatial attention decoding, and the calculation method is:

[0087] F s0 = RELU(conv1(E right ))

[0088] F s = Max(RELU(conv2(F s0 )))

[0089] conv1 (·) is a convolution operation, which is used as a frequency filter of the EEG information discrimination window, the input feature map channel number is 1, the output feature map channel number is 16, the convolution kernel size is C*1, and C is the channel number of the EEG data; ReLU(·) is a linear rectifier function, which is used as an activation function in the spatial information decoder; conv2(·) is a convolution operation, which is used as a spatial filter of the EEG information discrimination window, the input feature map channel number is 1, the output feature map channel number is 4, and the convolution kernel size is 2*32; Max(·) is a max-pooling layer function, which is used to reduce redundancy and prevent overfitting; F s is the output spatial feature.

[0090] The integration module is used for integrating the time feature and the spatial feature to obtain an attention detection result. The integration module includes a splicing unit, an extraction unit and a fusion unit.

[0091] The splicing unit is used for splicing and fusing the time feature and the spatial feature in the last dimension to obtain a fusion feature of the auditory spatial attention.

[0092] In the embodiment, the splicing unit splices the left brain time feature and the right brain spatial feature of the auditory spatial attention in the last dimension to obtain the fusion feature of the auditory spatial attention, and the calculation method is:

[0093] F st = Cat((F s , F t ), dim=-1)

[0094] wherein F st is the output fusion feature, F s is the spatial feature, F t is the time feature, Cat is a splicing function used for fusing the spatial feature and the time feature, and dim=-1 means that the splicing is performed in the last dimension of the spatial feature and the time feature, i.e., the last dimension.

[0095] The extraction unit is used for feature extraction on the fusion feature to obtain a compressed feature; the extraction unit includes a first convolution layer, a first max-pooling layer, a second convolution layer and a second max-pooling layer. The input feature map channel number of the first convolution layer is 1, the output feature map channel number is 1, and the convolution kernel size is 4*8; the first max-pooling layer is 2*2; the input feature map channel number of the second convolution layer is 1, the output feature map channel number is 1, and the convolution kernel size is 2*4; and the second max-pooling layer is 2*2. The specific working process includes: the obtained fusion feature is further extracted through a convolution operation, and the extracted compressed feature is output, and the calculation method is:

[0096] F1 = Max(RELU(conv3(F st )))

[0097] F2 = Max(RELU(conv4(F1)))

[0098] wherein, conv3(·) is a convolution operation of the first convolution layer, compressing the fusion feature, the input feature map channel number is 1, the output feature map channel number is 1, and the convolution kernel size is 4*8; ReLU(·) is a linear rectifier function; conv4(·) is a convolution operation of the second convolution layer, compressing the fusion feature, the input feature map channel number is 1, the output feature map channel number is 1, and the convolution kernel size is 2*4; Max(·) is a max-pooling layer function, used to reduce redundancy and prevent overfitting; and F2 is the output compressed feature.

[0099] The fusion unit is used for spatio-temporal information fusion of the compressed feature to obtain an attention detection result. The fusion unit includes a first linear layer, a second linear layer and an activation function. The first linear layer has an input of 72 dimensions and an output of 32 dimensions, the second linear layer has an input of 32 dimensions and an output of 2 dimensions, and the output result is mapped to a value range of 0 to 1 by using the activation function to obtain an auditory spatial attention detection result and a sound source direction focused by a brain-computer interface operator. The specific working process includes: inputting the compressed feature into the neural network linear layer, fusing the features, and outputting a model detection result, and the calculation method is:

[0100] Q = SoftMax(w4(Sigmoid(w3*F2+b3)+b4)

[0101] wherein, Sigmoid(·) is a nonlinear function, SoftMax(·) is a normalized exponential function, and is used as an activation function in a human auditory information integration module; w3 and w4 are weight parameters of the third linear layer and the fourth linear layer, b3 and b4 are bias parameters of the third linear layer and the fourth linear layer; and Q is the output model detection result.

[0102] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A brain-computer interface auditory spatial attention detection system driven by an auditory system, characterized in that, include: Signal acquisition module, signal processing module, and integration module; The signal acquisition module is used to acquire EEG signals and preprocess the EEG signals to obtain an EEG signal discrimination window; The signal processing module is used to decode the EEG signal discrimination window to obtain temporal and spatial features; The integration module is used to integrate the temporal features and the spatial features to obtain the attention detection result; The signal processing module includes: a left-brain time information decoder and a right-brain spatial information decoder; The left brain time information decoder is used to decode the EEG signal discrimination window to obtain the time feature; The right brain spatial information decoder is used to decode the EEG signal discrimination window to obtain the spatial features; The workflow of the left-brain time information decoder includes: The EEG signal discrimination window is input into the left brain information attention module to initialize the weight of the left brain information channel mask and obtain the first initialization mask of left brain channel attention. Perform a three-dimensional convolution on the first initialization mask to obtain the first three-dimensional feature of the first initialization mask; The first three-dimensional feature is input into the linear layer and normalized to obtain the left brain channel mask of left brain information; The inner product operation is performed on the left brain channel mask and the EEG signal discrimination window to obtain the left brain EEG information discrimination window for left brain channel attention. The left brain EEG information discrimination window is input into a one-way gated loop unit for decoding to obtain the time feature; The workflow of the right brain spatial information decoder includes: The EEG signal discrimination window is input into the right brain information attention module to initialize the weight of the right brain information channel mask and obtain the second initialization mask of right brain channel attention. Perform a 3D convolution on the second initialization mask to obtain the second 3D features of the second initialization mask; The second three-dimensional feature is input into the linear layer and normalized to obtain the right brain channel mask of right brain information; The inner product operation is performed on the right brain channel mask and the EEG signal discrimination window to obtain the right brain EEG information discrimination window for right brain channel attention. The right brain EEG information discrimination window is frequency filtered to obtain filtered EEG features; The filtered EEG features are transposed and then spatially filtered to obtain the spatial features.

2. The auditory spatial attention detection system driven by an auditory system as described in claim 1, characterized in that, The signal acquisition module includes: an acquisition unit and a preprocessing unit; The acquisition unit is used to acquire the real-time electroencephalogram (EEG) signals of the brain-computer interface operator. The preprocessing unit is used to preprocess the EEG signal to obtain the EEG signal discrimination window.

3. The auditory spatial attention detection system driven by an auditory system as described in claim 2, characterized in that, The workflow of the preprocessing unit includes: performing bandpass filtering, downsampling, common-average reference, and independent component analysis on the EEG signal to remove artifacts, and then segmenting the EEG signal into a discrimination window of a preset window length.

4. The auditory spatial attention detection system driven by an auditory system as described in claim 1, characterized in that, The integration module includes: a splicing unit, an extraction unit, and a fusion unit; The splicing unit is used to splice and fuse the temporal features and the spatial features in the last dimension to obtain the fused features of auditory spatial attention; The extraction unit is used to extract features from the fused features to obtain compressed features; The fusion unit is used to perform spatiotemporal information fusion on the compressed features to obtain the attention detection result.

5. The auditory spatial attention detection system driven by an auditory system as described in claim 4, characterized in that, The extraction unit includes: a first convolutional layer, a first max pooling layer, a second convolutional layer, and a second max pooling layer; The fusion unit includes: a first linear layer, a second linear layer, and an activation function.

Images