Auditory directional attention decoding method and device

Abstract

The present invention relates to the field of brain-computer interface technology, and discloses an auditory orientation attention decoding method and device, the method comprising: collecting an EEG signal of a listener to obtain a segment of EEG signals; preprocessing the segment of EEG signals to obtain a preprocessed EEG signal; performing time-frequency transformation on the preprocessed EEG signal to obtain a time-frequency energy spectrum of the EEG signal; performing frame processing on the obtained time-frequency energy spectrum to obtain a decision window set; inputting the decision window set into a trained auditory orientation attention decoding model in sequence to obtain the classification probability of the spatial orientation attention of the EEG signal; the auditory orientation attention decoding model comprises: a feature extractor, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function. Because the time-frequency energy spectrum of the EEG signal is used as input, it is easier for the decoding model to capture the EEG energy distribution clues that are more important for auditory orientation attention, so it is beneficial to improve the accuracy of decoding cross-trial EEG data.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and in particular to an auditory orientation attention decoding method and device. Background Art

[0002] At present, auditory orientation attention decoding technology mainly decodes the listener's auditory attention information from the listener's EEG signals. It is an important part of the field of brain-computer interface technology and has broad application prospects in the next generation of hearing aids.

[0003] Existing auditory azimuth attention decoding methods usually use the collected EEG time domain signals as training data directly after preprocessing. However, since EEG signals have significant trial characteristics, that is, the EEG signals of each trial have their inherent statistical characteristics that are different from other trials, the decoding model is easily affected by these characteristics during training and cannot effectively learn the features related to auditory attention. Therefore, existing methods have the problem of low decoding accuracy of cross-trial EEG data. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the prior art, and a method for auditory position attention decoding is proposed. The present invention also discloses an auditory position attention decoding device.

[0005] An auditory directional attention decoding method, comprising:

[0006] Collect the listener's EEG signal to obtain a segment of EEG signal;

[0007] Preprocessing the section of EEG signal to obtain a preprocessed EEG signal;

[0008] Performing time-frequency transformation on the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal;

[0009] The obtained time-frequency energy spectrum is framed to obtain a decision window set;

[0010] The decision window set is sequentially input into the trained auditory orientation attention decoding model to obtain the classification probability of the spatial orientation attention of the EEG signal;

[0011] The auditory position attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function. The trained auditory position attention decoding model processes the processed time-frequency energy spectrum in the following way:

[0012] The feature extractor extracts the spatiotemporal features in the processed time-frequency energy spectrum;

[0013] The average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum and obtains the reduced-dimensional spatial features;

[0014] The classifier classifies the spatial features after dimensionality reduction and obtains the classification probability of spatial orientation attention.

[0015] In an improved auditory orientation attention decoding method, the preprocessed EEG signal is subjected to time-frequency transformation to obtain the time-frequency energy spectrum of the EEG signal, including:

[0016] The preprocessed EEG signal is subjected to wavelet transformation, and the absolute value is taken and then logarithmic transformation is performed to obtain the time-frequency energy spectrum of the EEG signal.

[0017] In an improved auditory orientation attention decoding method, the obtained time-frequency energy spectrum is framed to obtain an original decision window set, including:

[0018] The obtained time-frequency energy spectrum is segmented in the time dimension to obtain a set of decision windows.

[0019] In an improved auditory orientation attention decoding method, the average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum to obtain the reduced-dimensional spatial features, including:

[0020] The average pooling layer averages the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

[0021] In an improved auditory position attention decoding method, the auditory position attention decoding model is trained in the following manner to obtain a trained auditory position attention decoding model:

[0022] Construct a typical sample set: collect the EEG signals of the listener to obtain EEG signals of N trials; preprocess the EEG signals of N trials to obtain EEG signals of N trials after preprocessing; perform time-frequency transformation on the EEG signals of N trials after preprocessing to obtain the time-frequency energy spectrum of the EEG signals of N trials; perform frame processing on the time-frequency energy spectrum of the EEG signals of N trials to obtain the original decision window set, perform random sampling and average superposition on the original decision window set to obtain a typical sample set containing M samples, where M and N are positive integers greater than 1;

[0023] Use a typical sample set for training: select a sample from the typical sample set, input it into the auditory orientation attention decoding model, obtain the classification probability of spatial orientation attention, use the cross entropy function to calculate the model loss based on the true label of the classification, and use the gradient back propagation algorithm to update the parameters of the convolution module and the fully connected module. Repeat the above process of selecting a sample from the typical sample set and inputting it into the auditory orientation attention decoding model until the auditory orientation attention decoding model converges to obtain a trained auditory orientation attention decoding model.

[0024] The embodiment of the present invention further provides an auditory orientation attention decoding device, comprising:

[0025] The EEG signal acquisition module is used to collect the EEG signal of the listener and obtain a segment of EEG signal;

[0026] an EEG signal processing module, used for preprocessing the EEG signal to obtain a preprocessed EEG signal; performing time-frequency transformation on the preprocessed EEG signal to obtain a time-frequency energy spectrum of the EEG signal; performing frame processing on the obtained time-frequency energy spectrum to obtain a decision window set; and

[0027] The auditory orientation attention decoding module is used to input the decision window set into the trained auditory orientation attention decoding model in sequence to obtain the classification probability of the spatial orientation attention of the EEG signal; wherein the auditory orientation attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function; the feature extractor of the trained auditory orientation attention decoding model extracts the spatiotemporal features in the processed time-frequency energy spectrum, the average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum to obtain the reduced-dimensional spatial features, and the classifier classifies the reduced-dimensional spatial features to obtain the classification probability of the spatial orientation attention.

[0028] In an improved auditory orientation attention decoding device, the EEG signal processing module is specifically used to perform wavelet transform on the preprocessed EEG signal, take the absolute value and then perform logarithmic transform to obtain the time-frequency energy spectrum of the EEG signal.

[0029] In an improved auditory orientation attention decoding device, the EEG signal processing module is specifically used to segment the obtained time-frequency energy spectrum in the time dimension to obtain an original decision window set.

[0030] In an improved auditory orientation attention decoding device, the auditory orientation attention decoding module is specifically used for an average pooling layer to average the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

[0031] In an improved auditory position attention decoding device, the trained auditory position attention decoding model includes:

[0032] A construction module is used to construct a typical sample set: collect the EEG signals of the listener to obtain EEG signals of N trials; preprocess the EEG signals of the N trials to obtain the preprocessed EEG signals of the N trials; perform time-frequency transformation on the preprocessed EEG signals of the N trials to obtain the time-frequency energy spectrum of the EEG signals of the N trials; perform frame processing on the time-frequency energy spectrum of the EEG signals of the N trials to obtain an original decision window set, perform random sampling and average superposition on the original decision window set to obtain a typical sample set containing M samples, wherein M and N are positive integers greater than 1; and

[0033] The training module is used for training using a typical sample set: a sample is selected from the typical sample set and input into the auditory orientation attention decoding model to obtain the classification probability of spatial orientation attention, the model loss is calculated using the cross entropy function in combination with the true label of the classification, and the parameters of the convolution module and the fully connected module are updated using the gradient back propagation algorithm. The above process of selecting a sample from the typical sample set and inputting into the auditory orientation attention decoding model is repeated continuously until the auditory orientation attention decoding model converges to obtain a trained auditory orientation attention decoding model.

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] On the one hand, the embodiment of the present invention takes the time-frequency energy spectrum of the EEG signal as input, so that the decoding model can more easily capture the EEG energy distribution clues that are more important for auditory orientation attention; on the other hand, by using samples from different trials to reconstruct typical samples as a training set, compared with the original data, the trial characteristics of the typical samples are significantly weakened, while the auditory attention characteristics are retained, which is more conducive to the training of the auditory orientation attention decoding model, and further helps to improve the accuracy of decoding EEG data across trials. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the auditory orientation attention decoding method according to embodiment 1 of the present invention;

[0037] Figure 2 is a flow chart of a training method for an auditory orientation attention decoding model according to Embodiment 1 of the present invention;

[0038] Figure 3 It is a structural diagram of an auditory orientation attention decoding device provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0040] Example 1

[0041] Figure 1A flowchart of an auditory attention decoding method, comprising:

[0042] Step 101: Collect the EEG signal of the listener to obtain a segment of EEG signal.

[0043] Step 102: preprocessing the section of EEG signal to obtain a preprocessed EEG signal.

[0044] Specifically, the following preprocessing can be performed:

[0045] The sampling rate of the collected EEG signal is reduced from 1000Hz to 128Hz, and then filtered using a Butterworth filter with a frequency range of 1-50Hz. Finally, for the data at each time sampling point, the average value of the data of all channels at that time point is subtracted, that is:

[0046]

[0047] Where E represents the EEG signal, t is the time index, i is the channel index, and C is the total number of channels.

[0048] Step 103: Perform time-frequency transformation on the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal.

[0049] Specifically, the time-frequency function can be wavelet transform and logarithmic transform, or short-time Fourier transform and other time-frequency transforms, the purpose of which is to obtain the time-frequency energy spectrum of the EEG signal.

[0050] The embodiment of the present invention uses Morlet wavelet transform to process the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal. The sampling times per second of the wavelet transform can be set to 10, and the frequency range is set to 2-50Hz. The wavelet transform result is logarithmically transformed after taking the absolute value to obtain the time-frequency energy spectrum of the EEG signal. The purpose of the logarithmic transformation is to compress the dynamic range of the time-frequency energy spectrum.

[0051] Step 104: Perform frame processing on the obtained time-frequency energy spectrum to obtain a decision window set.

[0052] Specifically, the frame processing here can be to segment the obtained time-frequency energy spectrum in the time dimension so that the length of each segment signal after segmentation is 1s in time. These signals with a length of 1s are the decision windows E w .

[0053] Step 105: Input the decision window set into the trained auditory orientation attention decoding model in sequence to obtain the classification probability of the spatial orientation attention of the EEG signal.

[0054] Here, each decision window enters the model and corresponds to a decoding result, which is the classification probability of the spatial orientation attention of the EEG signal.

[0055] In actual use, the subject can wear a brain computer to collect EEG signals in real time, continuously form decision windows, and the trained auditory orientation attention decoding model can decode these decision windows in real time.

[0056] In this step 105, the auditory position attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function. The trained auditory position attention decoding model processes the processed time-frequency energy spectrum in the following manner:

[0057] The feature extractor extracts the spatiotemporal features in the processed time-frequency energy spectrum;

[0058] The average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum and obtains the reduced-dimensional spatial features;

[0059] The classifier classifies the spatial features after dimensionality reduction and obtains the classification probability of spatial orientation attention.

[0060] Specifically, the average pooling layer can average the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

[0061] The above trained auditory orientation attention decoding model is passed Figure 2 The training shown is obtained, and the training method of the auditory orientation attention decoding model is explained below.

[0062] like Figure 2 As shown, it is a flow chart of the training method of the auditory orientation attention decoding model according to an embodiment of the present invention, comprising the following steps:

[0063] Step 201: Collect the EEG signals of the listener to obtain N trial EEG signals.

[0064] Specifically, an adult with normal hearing is selected as the listener, who is made to sit in a chair in the middle of the room in a comfortable posture and wear an EEG signal acquisition device. A speaker is placed 1.5 meters to the left and right of the listener to play the speech. The acquisition is divided into N trials. In each trial, the speakers on the left and right sides of the listener will play different speech respectively. The listener needs to focus his auditory attention on the speech in the specified direction throughout the trial. The specified direction of each trial is random, and N is a positive integer.

[0065] It should be noted that before each trial of the experiment, the experimenter will give the subject an instruction to keep the subject's attention on the sound in a certain direction (left or right) during the trial. This pre-given instruction is the true label of the trial. When collecting data, this true label is added to the beginning of the EEG signal of the trial to facilitate subsequent extraction and processing.

[0066] Step 202: Preprocess the EEG signals of N trials to obtain preprocessed EEG signals of N trials.

[0067] Specifically, for the EEG signal of the listener, the sampling rate of the collected EEG signal is first reduced from 1000 Hz to 128 Hz, and then filtered using a Butterworth filter with a frequency range of 1-50 Hz. Finally, for the data at each time sampling point, the average value of the data of all channels at that time point is subtracted, that is:

[0068]

[0069] Where E represents the EEG signal, t is the time index, i is the channel index, and C is the total number of channels.

[0070] Step 203: Perform time-frequency transformation on the preprocessed N trial EEG signals to obtain the time-frequency energy spectra of the N trial EEG signals.

[0071] Specifically, the preprocessed EEG signal is processed using Morlet wavelet transform to obtain the time-frequency energy spectrum of the EEG signal, the sampling times per second of the wavelet transform is set to 10, and the frequency range is set to 2-50Hz. The time-frequency energy spectrum of the EEG signal is logarithmically transformed to compress its dynamic range.

[0072] Step 204: Perform frame processing on the time-frequency energy spectra of the EEG signals of N trials to obtain an original decision window set.

[0073] Specifically, the EEG time-frequency energy spectra from different trials are segmented in the time dimension so that the length of each segmented signal is 1 second. These signals with a length of 1 second are the decision windows E w .

[0074] Step 205: Randomly sample and average the original decision window set to obtain a new typical sample set containing M samples.

[0075] Specifically, random sampling and average superposition of the original decision window set can be achieved through the following three steps:

[0076] Step 1: Randomly sample K numbers from a uniform distribution between 0.1 and 1 and normalize them to obtain K normalized weights;

[0077] Step 2: Randomly sample a number from 0 and 1 as the target direction y, where 0 and 1 represent different targets, including left and right. For example: Randomly sample a number from 0 and 1 as the target direction y, where 0 represents the target is left and 1 represents the target is right.

[0078] Step 3: Randomly sample K decision window samples from all decision windows with true labels y, and use the above K normalized weights to perform weighted summation on the K decision window samples to obtain the processed time-frequency energy spectrum and obtain a typical sample E p , where K is an adjustable positive integer parameter.

[0079] Repeat steps 1 to 3 above until enough typical samples are obtained and used as training data.

[0080] Step 206: Taking the typical sample as input, a feature extractor composed of a convolutional neural network is used to extract the spatiotemporal features of the typical sample.

[0081] Specifically, a typical sample is randomly sampled from the typical sample set, and the typical sample is processed using a convolutional neural network to obtain the spatiotemporal features E of the typical sample. st The transposed convolutional neural network consists of a two-dimensional convolutional layer, a batch normalization layer, and a parametric rectified linear unit activation function. The number of filters in the two-dimensional convolutional layer is 9, the convolution kernel size is (3,3), and the convolution step size is set to (1,1).

[0082] Step 207: Use average pooling to reduce the dimension of the spatiotemporal features to obtain the reduced-dimensional spatial features.

[0083] Specifically, the spatiotemporal features E of typical samples are st Average in the time dimension to obtain the reduced-dimensional spatial feature E s .

[0084] Step 208: Use a classifier consisting of a fully connected network and a Softmax activation function to classify the spatial features and obtain the classification probability of the spatial orientation attention.

[0085] Specifically, a fully connected neural network with an output dimension of 2 can be used to process the spatial feature E s , and use the Softmax function for normalization to get the orientation classification probability p.

[0086] Step 209: Calculate the model loss using the cross entropy function in combination with the real label of the classification, and train the parameters of the convolution module and the fully connected module in combination with the back propagation method to obtain the final auditory orientation attention decoding model step.

[0087] Specifically, the cross entropy function is used to calculate the loss of the output classification probability p and the true orientation label y:

[0088] L = -[ylog p + (1-y)log (1-p)]

[0089] Combined with the loss function, the parameters of the convolution module and the fully connected module are updated using the gradient back propagation algorithm. The process from step 206 to step 209 is repeated until the model converges, thereby obtaining the final auditory orientation attention decoding model.

[0090] The embodiment of the present invention provides an auditory attention decoding method that combines EEG signals with typical sample training technology. When training the model, EEG time-frequency features are constructed through time-frequency transformations such as wavelet transform, and more important EEG energy distribution clues are obtained. A new data set is constructed using the typical sample generation method to train the model; when using the model, the collected EEG signals are also transformed in time and frequency to construct EEG time-frequency features. Since typical samples are reconstructed as training sets using samples from different trials, the trial characteristics of typical samples are significantly weakened compared to the original training data, and the auditory attention characteristics are retained, which is more conducive to the training of the auditory orientation attention decoding model, and then the trained model is used for auditory attention decoding, which is conducive to improving the accuracy of statistical attention decoding of EEG data across trials.

[0091] Example 2

[0092] like Figure 3 As shown, it is a structural diagram of an auditory position attention decoding device provided by Example 2 of the present invention, comprising: an EEG signal acquisition module 31, an EEG signal processing module 32 and an auditory position attention decoding module 33; wherein:

[0093] The EEG signal acquisition module 31 is used to acquire the EEG signal of the listener to obtain a segment of EEG signal.

[0094] The EEG signal processing module 32 is used to preprocess the EEG signal to obtain a preprocessed EEG signal; perform time-frequency transformation on the preprocessed EEG signal to obtain a time-frequency energy spectrum of the EEG signal; and perform frame processing on the obtained time-frequency energy spectrum to obtain a decision window set.

[0095] The auditory orientation attention decoding module 33 is used to input the decision window set into the trained auditory orientation attention decoding model in sequence to obtain the classification probability of the spatial orientation attention of the EEG signal; wherein the auditory orientation attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function; the feature extractor of the trained auditory orientation attention decoding model extracts the spatiotemporal features in the processed time-frequency energy spectrum, the average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum to obtain the reduced-dimensional spatial features, and the classifier classifies the reduced-dimensional spatial features to obtain the classification probability of the spatial orientation attention.

[0096] Preferably, the EEG signal processing module 32 is specifically used to perform wavelet transform on the preprocessed EEG signal, take the absolute value and then perform logarithmic transform to obtain the time-frequency energy spectrum of the EEG signal.

[0097] Preferably, the EEG signal processing module 32 is specifically used to segment the obtained time-frequency energy spectrum in the time dimension to obtain an original decision window set.

[0098] Preferably, the auditory orientation attention decoding module 33 is specifically used to use an average pooling layer to average the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

[0099] Preferably, the auditory orientation attention decoding device further includes: a sample set construction module 34 and a training module 35;

[0100] The sample set construction module 34 is used to construct a typical sample set: collect the EEG signals of the listener to obtain EEG signals of N trials; preprocess the EEG signals of N trials to obtain EEG signals of N trials after preprocessing; perform time-frequency transformation on the EEG signals of multiple trials after preprocessing to obtain the time-frequency energy spectrum of the EEG signals of N trials; perform frame processing on the time-frequency energy spectrum of the EEG signals of N trials to obtain an original decision window set, and perform random sampling and average superposition on the original decision window set to obtain a typical sample set containing M samples.

[0101] The training module 35 is used for training using a typical sample set: a sample is selected from the typical sample set, and the sample is input into the auditory orientation attention decoding model to obtain the classification probability of the spatial orientation attention. The model loss is calculated using the cross entropy function in combination with the true label of the classification. The parameters of the convolution module and the fully connected module are updated using the gradient back propagation algorithm. The above process of selecting a sample from the typical sample set and inputting the sample into the auditory orientation attention decoding model is repeated continuously until the auditory orientation attention decoding model converges to obtain a trained auditory orientation attention decoding model.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. An auditory position attention decoding method, comprising: Collect the listener's EEG signal to obtain a segment of EEG signal; Preprocessing the section of EEG signal to obtain a preprocessed EEG signal; Performing time-frequency transformation on the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal; The obtained time-frequency energy spectrum is framed to obtain a decision window set; The decision window set is sequentially input into the trained auditory orientation attention decoding model to obtain the classification probability of the spatial orientation attention of the EEG signal; The auditory position attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function. The trained auditory position attention decoding model processes the processed time-frequency energy spectrum in the following way: The feature extractor extracts the spatiotemporal features in the processed time-frequency energy spectrum; The average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum and obtains the reduced-dimensional spatial features; The classifier classifies the spatial features after dimensionality reduction and obtains the classification probability of spatial orientation attention; The auditory position attention decoding model is trained in the following way to obtain a trained auditory position attention decoding model: Construct a typical sample set: collect the EEG signals of the listener to obtain EEG signals of N trials; preprocess the EEG signals of N trials to obtain EEG signals of N trials after preprocessing; perform time-frequency transformation on the EEG signals of N trials after preprocessing to obtain the time-frequency energy spectrum of the EEG signals of N trials; perform frame processing on the time-frequency energy spectrum of the EEG signals of N trials to obtain the original decision window set, perform random sampling and average superposition on the original decision window set to obtain a typical sample set containing M samples, where M and N are positive integers greater than 1; Use a typical sample set for training: select a sample from the typical sample set, input it into the auditory orientation attention decoding model, obtain the classification probability of spatial orientation attention, use the cross entropy function to calculate the model loss based on the true label of the classification, and use the gradient back propagation algorithm to update the parameters of the convolution module and the fully connected module. Repeat the above process of selecting a sample from the typical sample set and inputting it into the auditory orientation attention decoding model until the auditory orientation attention decoding model converges to obtain a trained auditory orientation attention decoding model.

2. The auditory position attention decoding method according to claim 1, characterized in that: Perform time-frequency transformation on the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal, including: The preprocessed EEG signal is subjected to wavelet transformation, and the absolute value is taken and then logarithmic transformation is performed to obtain the time-frequency energy spectrum of the EEG signal.

3. The auditory position attention decoding method according to claim 1, characterized in that: The obtained time-frequency energy spectrum is framed to obtain the original decision window set, including: The obtained time-frequency energy spectrum is segmented in the time dimension to obtain a set of decision windows.

4. The auditory position attention decoding method according to claim 1, characterized in that: The average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum and obtains the reduced-dimensional spatial features, including: The average pooling layer averages the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

5. An auditory orientation attention decoding device, characterized in that: include: The EEG signal acquisition module is used to collect the EEG signal of the listener and obtain a segment of EEG signal; An EEG signal processing module, used for preprocessing the EEG signal to obtain a preprocessed EEG signal; Performing time-frequency transformation on the preprocessed EEG signal to obtain the time-frequency energy spectrum of the EEG signal; The obtained time-frequency energy spectrum is framed to obtain a decision window set; and The auditory orientation attention decoding module is used to input the decision window set into the trained auditory orientation attention decoding model in sequence to obtain the classification probability of the spatial orientation attention of the EEG signal; wherein the auditory orientation attention decoding model includes: a feature extractor composed of a convolutional neural network, an average pooling layer, and a classifier composed of a fully connected network and a Softmax activation function; the feature extractor of the trained auditory orientation attention decoding model extracts the spatiotemporal features in the processed time-frequency energy spectrum, the average pooling layer reduces the dimension of the spatiotemporal features in the time-frequency energy spectrum to obtain the reduced-dimensional spatial features, and the classifier classifies the reduced-dimensional spatial features to obtain the classification probability of the spatial orientation attention; The trained auditory orientation attention decoding model includes: A construction module is used to construct a typical sample set: collect the EEG signals of the listener to obtain EEG signals of N trials; preprocess the EEG signals of the N trials to obtain the preprocessed EEG signals of the N trials; perform time-frequency transformation on the preprocessed EEG signals of the N trials to obtain the time-frequency energy spectrum of the EEG signals of the N trials; perform frame processing on the time-frequency energy spectrum of the EEG signals of the N trials to obtain an original decision window set, perform random sampling and average superposition on the original decision window set to obtain a typical sample set containing M samples, wherein M and N are positive integers greater than 1; and The training module is used for training using a typical sample set: a sample is selected from the typical sample set and input into the auditory orientation attention decoding model to obtain the classification probability of spatial orientation attention, the model loss is calculated using the cross entropy function in combination with the true label of the classification, and the parameters of the convolution module and the fully connected module are updated using the gradient back propagation algorithm. The above process of selecting a sample from the typical sample set and inputting into the auditory orientation attention decoding model is repeated continuously until the auditory orientation attention decoding model converges to obtain a trained auditory orientation attention decoding model.

6. The auditory position attention decoding device according to claim 5, characterized in that: The EEG signal processing module is specifically used to perform wavelet transform on the preprocessed EEG signal, take the absolute value and then perform logarithmic transform to obtain the time-frequency energy spectrum of the EEG signal.

7. The auditory position attention decoding device according to claim 5, characterized in that: The EEG signal processing module is specifically used to segment the obtained time-frequency energy spectrum in the time dimension to obtain an original decision window set.

8. The auditory position attention decoding device according to claim 5, characterized in that: The auditory orientation attention decoding module is specifically used for the average pooling layer to average the spatiotemporal features in the time-frequency energy spectrum in the time dimension to obtain the reduced-dimensional spatial features.

Images