Cross-period brain grain recognition method and system based on bilateral attention deep neural network

Abstract

The invention discloses a cross-period brain grain recognition method and system based on a bilateral attention deep neural network. Performing primary feature extraction on the electroencephalogram data through a reference network; performing advanced semantic feature extraction on the primary features by using a bilateral attention deep neural network; the bilateral attention deep neural network comprises a spatial feature refining unit, a channel feature refining unit and a feature aggregation unit; and classifying and identifying the high-level semantic features by using a classifier. According to the invention, through the spatial feature refining unit and the channel feature refining unit, collaborative deep processing is carried out on the primary features extracted by the reference network, so that the extracted advanced semantic features are more representative, and the accuracy of cross-period brain grain recognition is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric feature recognition based on electroencephalogram signals, and specifically relates to a cross-time period brain pattern recognition method and system based on a bilateral attention deep neural network. Background Art

[0002] Brainprint recognition technology is an emerging identity recognition method that uses non-invasive electrodes to collect EEG signals generated by cerebral cortical activity and extract the individual's unique neural activity pattern to achieve identity recognition and verification. Compared with traditional biometric technologies (such as facial recognition and fingerprint recognition), brainprint recognition technology has significant advantages in security. Once traditional biometrics are cracked, they cannot be changed, which may bring continuous security risks. The identity data of brainprint recognition can be reset as needed, providing a dynamic protection method for responding to emerging security threats. In addition, because EEG signals rely on the real-time brain activity of the subject, they can effectively prevent counterfeit attacks from the root and ensure the authenticity and reliability of the identity authentication process. This technology is particularly suitable for continuous identity authentication in high-security or high-risk environments, and can meet application scenarios with high continuous monitoring requirements, providing unique and irreplaceable security guarantees for related industries.

[0003] Although brainprint recognition technology has made significant progress in the field of identity authentication, it still faces two major challenges. First, the signal-to-noise ratio of EEG signals is generally low, and in order to ensure high recognition accuracy, a long data collection time is often required. For example, EEG data collection in a resting state usually takes tens of seconds, which to some extent affects the user experience and system response efficiency. Second, the performance of EEG signals varies significantly at different times, which is affected by multiple factors such as electrode position, user fatigue, emotional fluctuations, and environmental noise. This difference usually requires training or recalibration in a specific period of time, further limiting the feasibility and stability of brainprint recognition technology in practical applications. At present, the main methods to solve the problem of cross-period recognition rely on domain adaptation and adversarial learning techniques. However, these methods themselves have two key limitations. First, although adversarial learning can reduce feature transfer between domains, it often ignores the differences in category distribution between different domains, making it difficult to ensure stable feature representation learning, thereby affecting the recognition effect of the model. Second, domain adaptation technology may cause negative transfer, that is, performance degradation in the new data domain, thereby reducing the overall accuracy of the model. In addition, these methods usually require retraining the model when new data is introduced, which leads to decreased efficiency of the system and is not conducive to high efficiency and flexibility in practical applications.

[0004] In order to overcome the challenges of brainprint recognition technology in cross-time applications, the present invention proposes a cross-time brainprint recognition method and system based on bilateral attention deep neural network. The bilateral attention deep neural network plug-in consists of two parallel units: a spatial feature refining unit and a channel feature refining unit. Among them, the spatial feature refining unit dynamically adjusts the importance of features by statistically analyzing feature distribution, and combines the channel attention mechanism to evaluate the information content of different features in the layer normalization process using a trainable scaling factor. Then, the self-attention weights generated by the point convolution layer are used to reconstruct the spatial information to increase the amount of information and enhance the discrimination of key features. The channel feature refining unit optimizes the expression of channel features by grouping channels and adaptively merging sub-features within the group, reducing channel redundancy and improving the accuracy and representation ability of overall features. The bilateral attention deep neural network plug-in of the present invention is plug-and-play, can be applied to existing benchmark networks, and can directly enhance the features extracted from the original EEG signals. Without retraining, efficient identity recognition can be achieved through short-duration EEG signal input under cross-time conditions. This method not only improves the robustness of brainprint recognition, but also expands its feasibility and convenience in practical applications. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a cross-time brain pattern recognition method and system based on a bilateral attention deep neural network.

[0006] In a first aspect, the present invention provides a method for cross-period brain pattern recognition based on a bilateral attention deep neural network, characterized in that the method comprises the following steps: Step S1, EEG data collection; Step S2, preprocessing the EEG data; Step S3, performing primary feature extraction on the preprocessed EEG data through a reference network; Step S4, using a bilateral attention deep neural network to extract high-level semantic features from the primary features; The bilateral attention deep neural network includes a spatial feature refining unit, a channel feature refining unit, and a feature aggregation unit; The spatial feature refining unit statistically analyzes the feature distribution of the primary features, dynamically adjusts the importance of the features in combination with the channel attention mechanism, reconstructs the spatial features, and obtains the spatial optimized features; The channel feature refining unit optimizes the expression of channel features by grouping the channels represented by the primary features and adaptively merging the features within the group, thereby reducing channel redundancy and obtaining channel optimized features; The feature aggregation unit combines the spatial optimization features, the channel optimization features, and the primary features to obtain high-level semantic features; Step S5: Use a classifier to classify and identify the high-level semantic features output by the bilateral attention deep neural network.

[0007] Preferably, the reference network in step S3 adopts EEGNet.

[0008] Preferably, in step S4, the spatial feature refining unit in the bilateral attention deep neural network includes a weight redistribution module and a spatial feature reconstruction module; The implementation process of the weight redistribution module is as follows: Using layer normalization network Primary features extracted from the baseline network Process and obtain the standardized features , and then redistribute each feature through the Sigmoid activation function The weights of each feature are dynamically adjusted using the dot multiplication operation The importance of weight redistribution is obtained ; The spatial feature reconstruction module reconstructs the features after weight redistribution Reconstruction is performed to obtain spatial optimization features.

[0009] Preferably, in step S4, the spatial feature reconstruction module includes a channel attention layer CA, a first convolutional layer, a second convolutional layer, and a weight calculation module; The specific implementation process of the channel attention layer CA is: (1) in, represents the global pooling layer, and Represent two different fully connected layer coefficients, represents the relu activation function, represents the sigmoid activation function, Represents the weight of each channel; Then, the weights of each channel are used to adaptively extract features related to identity recognition, which can be expressed as: (2) (3) in Indicates strong identity characteristics, Indicates weak identity features; The first convolution layer and the second convolution layer project the features output by the channel attention layer CA into a new feature subspace through a point convolution operation; The weight calculation module performs weight calculation on the features processed by the first convolution layer and the second convolution layer.

[0010] Preferably, in step S4, the channel feature refining unit in the bilateral attention deep neural network includes a grouping module and a channel feature reconstruction module; The grouping module extracts the primary features from the baseline network Processing is performed to obtain channel groupings of primary feature representations; The channel feature reconstruction module reconstructs the channel grouping represented by the primary features to obtain channel optimization features.

[0011] Preferably, in step S4, the specific implementation process of the grouping module is: Primary features extracted from the baseline network Divide into G groups, and then divide each group into two sub-features along the channel dimension ; Use the first channel attention layer to the first sub-feature Perform channel attention operation to obtain the first channel sub-feature of the kth group ; Use the second channel attention layer to the second channel sub-feature Perform channel attention operation to obtain the second sub-feature of the kth group .

[0012] Preferably, in step S4, the specific implementation process of the channel feature reconstruction module is: The features obtained by processing the grouping module and Segmentation is performed and a new feature combination is obtained through mutual exchange operations and ; Applying a depthwise separable convolutional layer And point-wise convolutional layers For features and Extract high-level representation information, specifically expressed as: (4) The global average pooling operation is used to collect global spatial information and count channel descriptors, which can be expressed as: (5) By superimposing the global channel descriptor, the channel importance descriptors of the two sub-features are constrained to be within the range of 1, and the channel importance vector is obtained. and ; In the channel importance vector and Under the guidance of , the characteristics after channel optimization are obtained, which can be specifically expressed as: (6) Finally, the feature aggregation unit is used to combine the spatial optimization features and channel optimization features , and the original primary features , generating the final global identity feature: (7).

[0013] In a second aspect, the present invention provides a cross-time period brainprint recognition system, comprising: Data acquisition module, responsible for collecting EEG data; Data preprocessing module, responsible for preprocessing EEG data; The primary feature extraction module is responsible for extracting primary features from the preprocessed EEG data; The high-level semantic feature extraction module is responsible for extracting high-level semantic features from primary features using a bilateral attention deep neural network; The recognition module is responsible for using the classifier to realize cross-time brain pattern recognition based on high-level semantic features.

[0014] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

[0015] In a fourth aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

[0016] Through these devices and means, the present invention realizes an efficient and accurate cross-time brainprint recognition solution, bringing innovative technical support to the fields of biometric recognition and identity authentication.

[0017] The beneficial effects of the present invention include at least: The present invention uses the spatial feature refining unit and the channel feature refining unit in the bilateral attention deep neural network to deeply process the primary features extracted by the reference network. The spatial feature refining unit statistically analyzes the feature distribution and combines the channel attention mechanism to dynamically adjust the feature importance, reconstruct the spatial features, and enhance the discrimination of key features; the channel feature refining unit reduces channel redundancy, optimizes channel feature expression, and improves overall feature accuracy by grouping and adaptively merging features within the group. The synergy of the two makes the extracted high-level semantic features more representative, thereby significantly improving the accuracy of cross-time brain pattern recognition.

[0018] By introducing grouping operations and channel attention mechanisms, the present invention can process features more carefully, quantify the importance of different channels, and achieve adaptive extraction of identity recognition related features. Both strong and weak identity features can be reasonably paid attention to and utilized, ensuring that the model can focus on key information in different time periods and under different conditions, further enhancing the ability to adapt to cross-time data changes and improving recognition accuracy.

[0019] The present invention uses the trainable parameters in the standardization operation as a calibration operation to effectively balance the differences in the statistical characteristics of EEG signals between different time periods. EEG signals are easily affected by many factors, and data from different time periods show large differences. This operation of the present invention can make the model better adapt to these changes, reduce recognition errors caused by data differences, and improve the stability and reliability of the model across time periods.

[0020] The bilateral attention deep neural network plug-in proposed in the present invention is plug-and-play and is applicable to existing benchmark networks (such as EEGNet used in the present invention). It does not require large-scale changes to the original network architecture and can directly enhance the features extracted from the original EEG signals, thereby reducing the cost and difficulty of system development and optimization. It has good compatibility and practicality and is easy to integrate and apply in existing brain pattern recognition systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 It is a schematic diagram of the bilateral attention deep neural network structure provided by an embodiment of the present invention.

[0023] Figure 2 It is a schematic diagram of the structure of the spatial feature refining unit provided in an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of the structure of the channel feature refining unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0026] See also Figure 1, this embodiment provides a paradigm adaptive decoupling cross-time brain pattern recognition method, which uses the existing benchmark network as a feature extractor to extract primary feature representation from the original EEG data. Then, the extracted primary EEG feature representation enters the bilateral attention plug-in, through two parallel units in the plug-in: the spatial feature refining unit and the channel feature refining unit. On the one hand, the spatial feature refining unit dynamically adjusts the importance of the feature by statistically analyzing the feature distribution, and combines the channel attention mechanism to use a trainable scaling factor to evaluate the information content of different features in the layer normalization process. Then, the self-attention weights generated by the point convolution operation are used to reconstruct the spatial features to obtain the spatial optimization features. On the other hand, the channel feature refining unit optimizes the expression of the channel features by grouping the channels represented by the primary features and adaptively merging the sub-features in the group, reducing channel redundancy, improving the accuracy and representation ability of the overall features, and obtaining the channel optimization features. Finally, the spatial optimization features, channel optimization features, and primary features are merged by the feature aggregation unit, and finally the cross-time identity recognition is realized by the classifier. The specific implementation steps include the following:

[0027] Step S1: EEG data collection: Electrodes are placed on the scalp using dedicated equipment to collect EEG signals of multiple subjects in a non-invasive manner at different times using the same paradigm.

[0028] Step S2: EEG data is preprocessed and labeled with the identity label of the subject. Then, it is divided into registration data and data to be identified according to the order of collection time. Among them, the registration data has the known identity label of the subject and is used as the training set to train the model, while the data to be identified is the test set with predicted identity label.

[0029] Step S2-1: The original EEG signal contains noise frequency. In order to remove the power frequency interference caused by the EEG acquisition equipment and the myoelectric interference of the subject, the EEG data collected in step S1 is downsampled to 200 Hz, and the original EEG data is filtered at 1 to 75 Hz using a Butterworth filter to reduce noise and improve signal quality.

[0030] Step S2-2: Segment the EEG features obtained in step S2-1 to obtain multiple segments with a sample length of T (which can be 4s in this embodiment), where the i-th sample of the s-th subject is represented as: ,in Indicates the number of electrodes and labels the corresponding segments with the subject’s labels: The training set and the test set are divided according to the order of the collection time periods. The training set and the data of the later collection time period are used as the test set of the identity to be predicted.

[0031] Step S3: Perform primary feature extraction on the samples in step S2-2 through a benchmark network; use the classic EEGNet algorithm in brain-computer interface as the benchmark network.

[0032] Step S4: Use bilateral attention deep neural network to extract high-level semantic features from primary features; The bilateral attention deep neural network includes a spatial feature refining unit, a channel feature refining unit, and a feature aggregation unit; The spatial feature refining unit includes a weight redistribution module and a spatial feature reconstruction module. Figure 2 .

[0033] The weight redistribution module includes a layer normalization operation and a point multiplication operation; the layer normalization operation receives the primary features extracted from the reference network , by normalizing the network through layers Then, the feature distribution is statistically analyzed through the trainable scaling factor in the layer normalization network to obtain the standardized features. ; (1) in, represents the primary features extracted from the baseline network, The number of channels, height and width of the features, respectively. Presentation layer standardization layer, and Respectively from The mean and standard deviation of represents a very small constant used for numerical stability, and represents the trainable affine transformation parameters in the normalization operation, Used to scale the standardized features. Used to translate the standardized features.

[0034] Then redistribute the features through the Sigmoid activation function The weights are then used to dynamically adjust the features using the dot product operation The importance of weight redistribution is obtained ; Specifically expressed as: (2) The spatial feature reconstruction module receives the features processed by the weight redistribution module. , and reconstructed, which includes the channel attention layer CA, the first convolution layer, the second convolution layer, and the weight calculation module; The specific implementation process of the channel attention layer CA is as follows: (3) (4) in, represents the global pooling layer, and Represent two different fully connected layer coefficients, represents the relu activation function, Represents the sigmoid activation function; represents the weight of the kth channel; k=1,2,…,C, C represents the total number of channels; Then, the weights of each channel are used to adaptively extract features related to identity recognition, which can be expressed as: (5) (6) in Indicates strong identity characteristics, Indicates weak identity features; The first convolutional layer and the second convolutional layer are both composed of The convolution kernel is composed of and projected into the new feature subspace through the point convolution layer operation; The weight calculation module performs weight calculation on the features processed by the first convolution layer and the second convolution layer, which can be expressed as: (7) in, Represents the processing result of the first convolutional layer, Represents the processing result of the second convolutional layer.

[0035] The channel feature refining unit includes a grouping module and a channel feature reconstruction module. Figure 3 .

[0036] The grouping module extracts the primary features from the baseline network The processing includes a grouping operation and two channel attention operations; the grouping operation is responsible for dividing the primary features extracted by the baseline network into G groups along the channel dimension, G ≥ 1, which can be expressed as: (8) In each group, Then it is divided into two sub-features along the channel dimension and expressed as: (9) Channel attention operation, for each sub-feature in the grouping operation, obtains the corresponding attention weight of each channel in different ways, expressed as the importance of the channel, and uses the first channel attention layer to Perform channel attention operation, obtain the channel importance coefficient through the relationship between channels, and obtain the first channel sub-feature of the kth group ; Use the second channel attention layer to the second channel sub-feature Perform channel attention operation, use the spatial relationship in the feature to obtain the importance coefficient, and get the second sub-feature of the kth group ; Among them, the first channel attention layer consists of a global pooling layer , Sigmoid activation layer, RELU activation layer, fusion layer, after the first channel attention operation, the first channel sub-feature of the kth group is obtained ; The second channel attention layer consists of a group normalization layer , Sigmoid activation layer, fusion layer, after the second channel attention layer processing, the second channel sub-feature of the kth group is obtained ; (10) (11) in, represents the global pooling layer, , and represents three different fully connected layer coefficients, represents the relu activation function, represents the sigmoid activation function, represents the group normalization operation, Represents the bias coefficient.

[0037] The channel feature reconstruction module receives the features processed by the grouping module. and , and reconstruct it, which includes segmentation operation, depth convolution layer, point-by-point convolution layer, weight normalization operation, splicing layer. The specific implementation process is: Features after processing the first channel attention layer and Segment along the channel and obtain new feature combinations through mutual exchange operations and ; (12) (13) (14) in Represents a split operation; Applying a depthwise separable convolutional layer And point-wise convolutional layers For features and Extract high-level representation information, where The convolution kernel size is expressed as , The convolution kernel size is expressed as Due to the sparse convolutional connections, The calculation complexity of the parameter kernel is reduced, but the information flow between channels is reduced to a certain extent. This makes up for this and promotes the flow of information between feature channels, which can be specifically expressed as:

[0038] (15) The global average pooling operation is used to collect global spatial information and count channel descriptors, which can be expressed as: (16) in, Then, by superimposing the global channel descriptor, the descriptors of the channel importance of the two sub-features are constrained to be within the range of 1, which is specifically expressed as:

[0039] (17) Where e is the base of the natural logarithm, through the exponential operation , Implement normalized aspect calculation. This calculation method helps to standardize the features so that they are within a specific range, thereby optimizing the numerical distribution and expression effect of the features.

[0040] In the channel importance vector and Under the guidance of , the characteristics after channel optimization are obtained, which can be specifically expressed as: (18) Finally, the feature aggregation unit is used to combine the spatial optimization features and channel optimization features , and the original primary features , generating the final global identity feature: (19) Step S5: using a classifier to classify and identify the final feature representation obtained by the feature aggregation unit; Classifier The loss function used during training is: (20) in, represents the flattening operation, is the identity label corresponding to the sample, Represents a classifier.

[0041] To verify the effectiveness of the method of the present invention, two cross-period data sets (containing 30 and 54 subjects respectively) were used: based on the EEG data set collected by the SSVEP protocol, the classic method EEGNet in the field of brain-computer interface was used as the benchmark network, and the performance of cross-period identity recognition and cross-period identity authentication scenarios were compared with other existing methods under the cross-period recognition framework. Among them, cross-period identity authentication includes two situations: known attackers (i.e., their data is included in the training set) and unknown attackers (i.e., their data is not included in the training set). The obtained recognition accuracy and F1 score (%) and the equal error rate EER (%) in the authentication scenario are shown in Tables 1 and 2 below:

[0042] Table 1 Cross-period identity recognition results

[0043] Table 2 Cross-period identity authentication results (EER %)

[0044] The embodiment of the present invention also provides a cross-time period brainprint recognition system, including: Data acquisition module, responsible for collecting EEG data; Data preprocessing module, responsible for preprocessing EEG data; The primary feature extraction module is responsible for extracting primary features from the preprocessed EEG data; The high-level semantic feature extraction module is responsible for extracting high-level semantic features from primary features using a bilateral attention deep neural network; The recognition module is responsible for using the classifier to realize cross-time brain pattern recognition based on high-level semantic features.

[0045] An embodiment of the present invention further provides an electronic device. Specifically, the electronic device includes a memory and a processor. The memory stores executable code. When the processor executes the executable code, the method described in any one of the implementation modes is implemented.

[0046] The memory may include high-speed random access memory (RAM), and may also include non-volatile memory (Non-volatile Memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0047] The bus may be an ISA bus, a PCI bus or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc.

[0048] Among them, the memory is used to store the program, and the processor executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor or implemented by the processor.

[0049] The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in a decoding processor. The software module may be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0050] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0051] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0052] In summary, the present invention can use a shorter period of data input for cross-period identity recognition. For example, in the actual embodiment, only a 4s data segment is needed to achieve effective recognition. Compared with the traditional method that requires a longer time for data collection (such as tens of seconds in a resting state), the data collection time is greatly shortened, the system response efficiency is improved, the user experience is improved, and it is more in line with the efficiency requirements in actual application scenarios. EEG signals have the characteristic of being difficult to be stolen, so that the cross-period brainprint recognition technology based on the present invention has high confidentiality. In the process of identity authentication and identification, it can effectively prevent identity information from being stolen or forged, and provide users with more secure and reliable identity recognition services. It has unique advantages in high-security demand scenarios (such as financial transactions, confidential information access, etc.), and is expected to become a method of identity authentication and identification with a higher security level. The cross-period brainprint recognition method of the present invention can be widely used in multiple fields such as biometric identification, identity authentication, and security authentication, providing innovative technical support for these fields, meeting the needs of different industries for the accuracy, security and efficiency of identity recognition technology, and promoting the development and progress of related industry technologies. It has broad market application prospects and social value.

[0053] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A cross-time brain pattern recognition method based on bilateral attention deep neural network, characterized in that: The method comprises the following steps: Step S1, EEG data collection; Step S2, preprocessing the EEG data; Step S3, performing primary feature extraction on the preprocessed EEG data through a reference network; Step S4, using a bilateral attention deep neural network to extract high-level semantic features from the primary features; The bilateral attention deep neural network includes a spatial feature refining unit, a channel feature refining unit, and a feature aggregation unit; The spatial feature refining unit statistically analyzes the feature distribution of the primary features, dynamically adjusts the importance of the features in combination with the channel attention mechanism, reconstructs the spatial features, and obtains the spatial optimized features; The channel feature refining unit optimizes the expression of channel features by grouping the channels represented by the primary features and adaptively merging the features within the group, thereby reducing channel redundancy and obtaining channel optimized features; The feature aggregation unit combines the spatial optimization features, the channel optimization features, and the primary features to obtain high-level semantic features; Step S5: Use a classifier to classify and identify the high-level semantic features output by the bilateral attention deep neural network.

2. The method according to claim 1, characterized in that: The reference network in step S3 adopts EEGNet.

3. The method according to claim 1, characterized in that: In step S4, the spatial feature refining unit in the bilateral attention deep neural network includes a weight redistribution module and a spatial feature reconstruction module; The implementation process of the weight redistribution module is as follows: Using layer normalization network Primary features extracted from the baseline network Process and obtain the standardized features , and then redistribute each feature through the Sigmoid activation function The weights of each feature are dynamically adjusted using the dot multiplication operation The importance of weight redistribution is obtained ; The spatial feature reconstruction module reconstructs the features after weight redistribution Reconstruction is performed to obtain spatial optimization features.

4. The method according to claim 3, characterized in that: In step S4, the spatial feature reconstruction module includes a channel attention layer CA, a first convolutional layer, a second convolutional layer, and a weight calculation module; The specific implementation process of the channel attention layer CA is: (1) in, represents the global pooling layer, and Represent two different fully connected layer coefficients, represents the relu activation function, represents the sigmoid activation function, Represents the weight of each channel; Then, the weights of each channel are used to adaptively extract features related to identity recognition, which can be expressed as: (2) (3) in Indicates strong identity characteristics, Indicates weak identity features; The first convolution layer and the second convolution layer project the features output by the channel attention layer CA into a new feature subspace through a point convolution operation; The weight calculation module performs weight calculation on the features processed by the first convolution layer and the second convolution layer.

5. The method according to claim 1, characterized in that: In step S4, the channel feature refining unit in the bilateral attention deep neural network includes a grouping module and a channel feature reconstruction module; The grouping module extracts the primary features from the baseline network Processing is performed to obtain channel groupings of primary feature representations; The channel feature reconstruction module reconstructs the channel grouping represented by the primary features to obtain channel optimization features.

6. The method according to claim 5, characterized in that: In step S4, the specific implementation process of the grouping module is: Primary features extracted from the baseline network Divide into G groups, and then divide each group into two sub-features along the channel dimension ; Use the first channel attention layer to the first sub-feature Perform channel attention operation to obtain the first channel sub-feature of the kth group ; Use the second channel attention layer to the second channel sub-feature Perform channel attention operation to obtain the second sub-feature of the kth group .

7. The method according to claim 6, characterized in that: In step S4, the specific implementation process of the channel feature reconstruction module is: The features obtained by processing the grouping module and Segmentation is performed and a new feature combination is obtained through mutual exchange operations and ; Applying a depthwise separable convolutional layer And point-wise convolutional layers For features and Extract high-level representation information, specifically expressed as: (4) The global average pooling operation is used to collect global spatial information and count channel descriptors, which is specifically expressed as: (5) By superimposing the global channel descriptor, the channel importance descriptors of the two sub-features are constrained to be within the range of 1, and the channel importance vector is obtained. and ; In the channel importance vector and Under the guidance of , the characteristics after channel optimization are obtained, which is specifically expressed as: (6) Finally, the feature aggregation unit is used to combine the spatial optimization features and channel optimization features , and the original primary features , generating the final global identity feature: (7)。 8. A cross-time brain pattern recognition system implementing the method described in any one of claims 1 to 7, characterized in that include: Data acquisition module, responsible for collecting EEG data; Data preprocessing module, responsible for preprocessing EEG data; The primary feature extraction module is responsible for extracting primary features from the preprocessed EEG data; The high-level semantic feature extraction module is responsible for extracting high-level semantic features from primary features using a bilateral attention deep neural network; The recognition module is responsible for using the classifier to realize cross-time brain pattern recognition based on high-level semantic features.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 7.

10. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 7 is implemented.

Images