Cross-period brain grain recognition method and system based on deep learning network introducing normalization compensation mechanism

Abstract

The invention discloses a cross-period brain grain recognition method and system based on a normalized compensation mechanism introduced deep learning network. The method comprises the following steps: firstly, collecting electroencephalogram data, and then preprocessing the electroencephalogram data; performing feature extraction on the preprocessed electroencephalogram data through a deep learning network based on introduction of a normalized compensation mechanism; and using a classifier to classify and identify the features output by the deep learning network based on introduction of the normalized compensation mechanism. According to the method, the instance normalization compensation module, the batch normalization compensation module and the layer normalization compensation module are introduced into the deep learning network, the feature distribution difference of the cross-period brain grain recognition task is effectively adjusted, and the problem that the performance is reduced when cross-period data are processed through a traditional method is solved. The compensation mechanisms effectively improve the robustness of data migration of the model in different time periods.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric recognition based on electroencephalogram (EEG) signals, and particularly relates to a cross-time brainprint recognition method and system based on a deep learning network introducing a normalization compensation mechanism. Background Art

[0002] In the field of information security, the reliability of key passwords is crucial. In high-security scenarios, biometric recognition has become the main means of authentication. Although traditional biometric features, such as fingerprints, irises, and faces, are widely used, there are still risks of being stolen, copied, or tampered with. Currently, there are mature technologies that can extract an individual's fingerprint and facial features from images and generate replicas through means such as 3D printing, thereby deceiving the system to pass authentication. In addition, with the development of deep learning technology, synthesis technology has been able to generate realistic facial images, further increasing the success rate of deceiving the system. Compared with traditional biometric features, brainprint recognition performs identity authentication by collecting an individual's EEG signals, and it has the following advantages: (1) Higher concealment: EEG signals originate from brain thinking activities, have unique neural path patterns, and the collection requires professional equipment, which makes it difficult for attackers to forge or copy through physiological imitation means; (2) Higher security: Compared with traditional biometric features, the collection of EEG signals requires the presence of a living body, thus greatly enhancing security. (3) Revocability: Even if the data is leaked, the EEG signals can be replaced in a timely manner, while once traditional biometric features are leaked, they will pose a long-term security threat. Therefore, brainprint recognition technology is particularly suitable for application scenarios with high security requirements.

[0003] The core of brainprint recognition lies in enabling the model to effectively capture the differences in EEG signals among different subjects, so as to achieve accurate identity recognition and authentication. In practical applications, cross-time brainprint recognition (i.e., the registration data and the recognition data come from different time periods or different days) is particularly important and meets the actual needs. Currently, in the cross-time brainprint recognition technology based on deep neural networks, although the commonly used normalization layer in the model can accelerate the training and convergence of the model, it may still lead to the loss of between-class and within-class difference information, thereby limiting the recognition performance of the model. On the other hand, existing research usually models cross-time brainprint recognition as a domain adaptation problem, that is, taking the data in the registration stage as the source domain and the data in the recognition stage as the target domain. Although methods such as adversarial learning or transfer learning can be used to train an identity classifier to achieve cross-time identity recognition, these methods often ignore the distribution differences of data in different time periods and may cause negative transfer, further reducing the model performance.

[0004] In view of the above problems, the present invention proposes a normalization compensation mechanism based on deep learning, and designs a cross-time brainprint recognition method in combination with this mechanism, aiming to effectively preserve and utilize the differences between subjects and improve the accuracy of cross-time identity recognition. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention proposes a cross-time brainprint recognition method and system based on a deep learning network introducing a normalization compensation mechanism.

[0006] In a first aspect, the present invention provides a cross-time brainprint recognition method based on a deep learning network introducing a normalization compensation mechanism, the method comprising:

[0007] Step S1, electroencephalogram data acquisition;

[0008] Step S2, preprocessing the electroencephalogram data;

[0009] Step S3, extracting features from the preprocessed electroencephalogram data through a deep learning network introducing a normalization compensation mechanism;

[0010] Step S4, using a classifier to classify and recognize the features output by the deep learning network introducing a normalization compensation mechanism;

[0011] The deep learning network introducing a normalization compensation mechanism includes a basic feature extraction module, an instance normalization compensation module, a batch normalization compensation module, and a layer normalization compensation module;

[0012] The basic feature extraction module is responsible for extracting primary features X containing spatio-temporal information from the preprocessed electroencephalogram data out ;

[0013] The instance normalization compensation module is responsible for processing each feature map instance in the primary features X out based on the compensation mechanism to obtain instance-corrected features X S ;

[0014] The batch normalization compensation module is responsible for further adjusting the overall information of all batches based on the instance-corrected features X S to obtain batch-corrected features X B ;

[0015] The layer normalization compensation module is responsible for finally optimizing and adjusting each sample in each batch based on the batch-corrected features X B to obtain sample-corrected features X c .

[0016] Preferably, the basic feature extraction module includes a first convolutional neural network and the second convolutional neural network

[0017] The first convolutional neural network layer captures temporal information from the preprocessed EEG data to obtain a temporal feature map;

[0018] The second convolutional neural network layer captures spatial domain information from the temporal feature map to obtain the primary feature X containing spatio-temporal information out 。

[0019] Preferably, the instance normalization compensation module includes an instance internal correction unit, a power calculation unit, and a fusion unit;

[0020] The instance internal correction unit is responsible for performing normalization processing on each feature map instance inside the primary feature X out By adjusting the mean and variance, it reduces the deviation between the training and test distributions, ensuring the consistency of feature intensity;

[0021] The power calculation unit is responsible for further amplifying the discriminative key information by calculating the power of the primary feature, thereby compensating for the class-related features that may be lost during the normalization process, strengthening the uniqueness of the instance and the class discrimination ability, and obtaining the enhanced feature X power ;

[0022] The fusion unit is responsible for concatenating the output of the instance internal correction unit the output X of the power calculation unit power to obtain the instance-corrected feature X S 。

[0023] More preferably, the instance internal correction unit includes an instance centering correction operation and an instance scaling correction operation;

[0024] The instance centering correction operation refers to using trainable parameters W S1 、b S1 to filter X out before the traditional instance normalization operation to obtain the instance-centered corrected feature X in ;

[0025]

[0026] where, W S1 and b S1 represent learnable parameters respectively, represents the relu activation function, represents the dot product operation;

[0027] The instance scaling correction operation refers to performing an operation on the instance-centered corrected feature Xin After traditional instance normalization, an activation function is used to obtain weights, which restricts the feature intensity within the instance to obtain the features after instance scaling and correction.

[0028]

[0029] Among them, IN(·) represents the traditional instance normalization operation, δ(·) represents the sigmoid activation function, and W S2 and b S2 respectively represent learnable parameters, while γ in and β in respectively represent the learnable parameters of the affine transformation in the traditional instance normalization operation, and u1(·) and σ1(·) respectively represent the mean operation and variance operation in the traditional instance normalization operation.

[0030] More preferably, the implementation process of the power calculation unit is:

[0031]

[0032] Among them, represents the square operation, represents the average pooling operation, represents the logarithmic transformation operation.

[0033] The implementation process of the fusion unit is:

[0034]

[0035] Among them, CAT(·) represents the concatenation operation, represents the average pooling operation.

[0036] Preferably, the batch normalization compensation module includes a batch centering correction operation and a batch scaling correction operation;

[0037] The batch centering correction operation means that before the traditional batch normalization operation, the instance correction feature X s is filtered using learnable parameters to obtain the batch centering correction feature X bn ;

[0038]

[0039] Among them, W B1 and b B1 respectively represent learnable parameters, represents the relu activation function;

[0040] The batch scaling and correction operation refers to obtaining weights using an activation function after traditional batch normalization, restricting the feature intensity of the entire batch, and obtaining batch-corrected features X B ;

[0041]

[0042] where BN(·) represents the traditional batch normalization operation, δ(·) represents the sigmoid activation function, W B2 and b B2 respectively represent learnable parameters, while γ bn and β bn respectively represent the learnable parameters of the affine transformation in the traditional batch normalization operation, and u2(·) and σ2(·) respectively represent the mean operation and variance operation in the traditional batch normalization operation.

[0043] Preferably, the layer normalization compensation module includes a sample centering correction operation and a sample scaling correction operation;

[0044] The sample centering correction operation refers to filtering the batch-corrected features X B using trainable parameters before the traditional layer normalization operation to obtain sample-centered corrected features X ln

[0045]

[0046] where, W C1 and b C1 respectively represent learnable parameters, represents the relu activation function;

[0047] The sample scaling correction operation refers to obtaining weights using an activation function after the traditional layer normalization, restricting the feature intensity of each sample, and obtaining sample-corrected features X c ;

[0048]

[0049] where LN(·) represents the traditional layer normalization operation, δ(·) represents the sigmoid activation function, W C2 and b C2 respectively represent learnable parameters, while γ ln and β ln respectively represent the learnable parameters of the affine transformation in the traditional layer normalization operation, and u3(·) and σ3(·) respectively represent the mean operation and variance operation in the traditional layer normalization operation.

[0050] In a second aspect, the present invention provides a cross-time brain pattern identity recognition device, including:

[0051] A data acquisition module, responsible for acquiring EEG data;

[0052] A data preprocessing module, responsible for preprocessing the acquired EEG data;

[0053] An identification module, responsible for inputting the preprocessed EEG data into a deep learning network and a classifier based on the introduced normalization compensation mechanism, and obtaining the cross - time brainprint identity recognition result.

[0054] Thirdly, the present invention provides a computer - readable storage medium, on which a computer program is stored. When the computer program is executed in a computer, the computer is made to operate according to the cross - time brainprint recognition method of the present invention, so as to achieve efficient and accurate cross - time identity recognition.

[0055] Fourthly, the present invention provides a computing device, including a memory and a processor. An executable code is stored in the memory. When the processor executes the executable code, cross - time brainprint recognition is performed according to the method of the present invention. Specifically, by loading and executing the program in the memory, the processor can implement steps such as data acquisition, preprocessing, feature extraction, normalization compensation, and identity recognition, so as to efficiently complete the brainprint recognition task.

[0056] Through these devices and apparatuses, the present invention provides an efficient and accurate cross - time brainprint recognition solution, bringing innovative technical support to fields such as biometric recognition and identity verification, and promoting the technological progress and application development of related fields.

[0057] The present invention has the following characteristics and beneficial effects:

[0058] 1. Normalization compensation mechanism: By introducing instance normalization compensation, batch normalization compensation, and layer normalization compensation modules, the present invention effectively adjusts the feature distribution differences for the cross - time brainprint recognition task, overcoming the problem of performance degradation of traditional methods when dealing with cross - time data. These compensation mechanisms effectively improve the robustness of the model in data migration between different time periods.

[0059] 2. Enhanced cross - time recognition ability: Through multi - level compensation of normalization operations, the present invention balances the statistical characteristic differences of data in different time periods, enabling the model to more accurately perform identity recognition in cross - time scenarios, and improving the accuracy and stability of cross - time recognition.

[0060] 3. Adaptive extraction of spatial and temporal features: Through the basic feature extraction module and combining with depth convolution operations, the present invention extracts the key features of EEG signals from two dimensions of time series and space respectively. This multi - dimensional feature extraction method

[0061] Ensures the integrity and consistency of brain pattern information in different time periods.

[0062] 4. Efficient processing of short - time data: The present invention can use electroencephalogram (EEG) signal data within a short time period for cross - time identity recognition, making the method more efficient in practical applications and meeting the requirements of rapid authentication. By precisely processing short - time data, the response speed and practicality of the system are improved.

[0063] 5. High security and confidentiality: Due to the individual uniqueness and difficulty of forgery of EEG signals, the cross - time brain pattern recognition method provided by the present invention has high confidentiality and security, is applicable to identity authentication scenarios with high security requirements, and provides a new biometric technology solution.

[0064] 6. Wide application prospects: The cross - time brain pattern recognition method of the present invention has wide application prospects and can be widely applied in the fields of biometric recognition, identity verification, security authentication, intelligent device control, etc., having important social value and market potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] To more intuitively show the technical solution of the present invention, the following briefly introduces the drawings required in the embodiments of the present invention. It should be noted that the drawings in the following description are only a part of the embodiments of the present invention, and those skilled in the relevant art can deduce other possible drawings based on these drawings without creative labor.

[0066] Figure 1 Shows the workflow of the cross - time brain pattern recognition method of the present invention.

[0067] The figure shows the detailed operation processes of steps such as data acquisition, pre - processing, feature extraction, normalization compensation, and identity recognition, aiming to help understand the mutual relationship of each module and its role in cross - time brain pattern recognition. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] Without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0069] This embodiment provides a cross - temporal brainprint recognition method, which introduces a deep learning network based on a normalization compensation mechanism. It consists of the following modules: a basic feature extraction module, an instance normalization compensation module, a batch normalization compensation module, and a layer normalization compensation module. Among them, the basic feature extraction module is responsible for extracting primary feature representations containing spatio - temporal information from the original EEG data; the instance normalization compensation module processes each feature map instance in the primary feature representation based on the compensation mechanism; the batch normalization compensation module further adjusts the overall information of each batch of data input into the model; finally, the layer normalization compensation module performs final optimization and adjustment for each sample in each batch, thereby enhancing the generalization ability and recognition performance of the model. The normalization compensation mechanism proposed by the present invention can significantly enhance the feature representations extracted from EEG signals. In a cross - temporal scenario, this method can achieve high - precision identity recognition through short - term EEG signal input without cumbersome pre - processing or additional secondary training. The present invention not only significantly improves the accuracy and robustness of cross - temporal brainprint recognition, but also further expands its applicability and convenience in practical applications.

[0070] The specific implementation steps of the above - mentioned method are as follows:

[0071] Step S1, EEG data acquisition: Electrodes are placed on the scalp through a dedicated device to non - invasively collect EEG signals of multiple subjects at different times, ensuring the diversity and representativeness of the data.

[0072] Step S2, EEG data pre - processing and identity annotation: The collected EEG data is pre - processed, and the identity label of each subject to which each data segment belongs is annotated. Then, the data is divided into registration data and data to be recognized according to the order of the acquisition time period. The registration data corresponds to known identity labels and is used as a training set for training the model; the data to be recognized is used for testing and has an identity label to be predicted.

[0073] 2 - 1: Filter and down - sample the EEG data collected in step 1 to reduce noise interference and improve the signal quality.

[0074] 2 - 2: The EEG data obtained in step 2 - 1 is segmented into multiple time segments, ensuring that the length of each segment is T. Each segment is labeled with the identity label of a certain subject, forming a training set and a test set. The training set is used for training the model, and the test set is used to evaluate the recognition performance of the model.

[0075] Step S3, construct a deep learning network based on the introduced normalization compensation mechanism, and extract features from the pre - processed EEG data;

[0076] 3-1 Construct a basic feature extraction module to extract the primary feature representation from the samples in step 2-2. The basic feature extraction module consists of two convolutional neural network layers, namely the first convolutional neural network and the second convolutional neural network Among them, the first convolutional neural network layer is designed to capture temporal information, and the second convolutional neural network layer is designed to capture spatial information. Through the stacking of these two convolutional layers, the basic feature extraction module can effectively extract the temporal features and spatial features in the original EEG signals, forming the primary feature representation X out , which is to be processed by the subsequent normalization compensation module.

[0077] 3-2 Construct an instance normalization compensation module to process the primary feature representation X out . Through the primary feature representation X out , it is processed by the instance normalization compensation module, which includes two parallel processing units. The first unit is the intra-instance correction unit, specifically including an intra-instance centering correction operation and an intra-instance scaling correction operation. The intra-instance centering correction operation means that before the traditional instance normalization operation, X out is filtered using trainable parameters. The intra-instance scaling correction operation means that after the traditional instance normalization, an activation function is used to obtain weights to limit the feature intensity within the instance. The second unit is the power calculation unit, which calculates the power of each instance through a square operation and applies a logarithmic transformation to further amplify the features with larger differences. Then, an average pooling operation is used to reduce the dimension of the power features, thereby retaining key information and enhancing the discriminability of the features. Then, the features after being processed by these two parallel units are concatenated to obtain the instance-corrected feature X S .

[0078] 3-3 Construct a batch normalization compensation module to process the instance-corrected feature X S . After passing through the instance-corrected feature X s , an overall processing is performed by the batch normalization compensation module. Specifically, this module includes a batch centering correction operation and a batch scaling correction operation. First, the batch centering correction operation means that before the traditional batch normalization operation, X s is filtered using trainable parameters, and then through the batch scaling correction operation, an activation function is used to obtain weights after the traditional batch normalization to limit the feature intensity of the entire batch. After being processed by the batch normalization compensation module, the batch-corrected feature X B is obtained.

[0079] 3-4 Construct a layer normalization compensation module to process the batch-corrected feature X B . After passing through the batch-corrected feature XB After that, each sample in this batch is processed by the layer normalization compensation module. Specifically, this module includes a sample centering correction operation and a sample scaling correction operation. First, the sample centering correction operation means that before the traditional layer normalization operation, the parameter X is filtered using trainable parameters b and then, through the sample scaling correction operation, weights are obtained using an activation function after the traditional layer normalization to limit the feature intensity of each sample. After being processed by the layer normalization compensation module, the sample corrected feature X c is obtained.

[0080] Step S4: Train the network to achieve cross-time identity recognition: Use the sample corrected feature X c finally extracted through the above steps to train an identity classifier, and continuously adjust the network parameters through an optimization algorithm, so that the model can effectively learn identity features from the training data, thereby improving the accuracy and generalization ability of the model in the cross-time brainprint recognition task.

[0081] Through the above specific implementation steps, the present invention realizes an efficient cross-time brainprint recognition method, which can significantly improve the accuracy and robustness of brainprint recognition in actual application scenarios.

[0082] Specifically, a cross-time brainprint recognition method based on a deep learning network introducing a normalization compensation mechanism is shown in the appendix Figure 1 and includes:

[0083] Step S1: Collect electroencephalogram (EEG) data;

[0084] Exemplarily, an EEG cap is used to connect electrodes to the corresponding brain regions of the subject to collect EEG data. The EEG signals in each period are obtained in a non-invasive manner.

[0085] Step S2: Preprocess the EEG data;

[0086] Preprocess the EEG data collected in step 1 to remove the noise and interference in the original signal. First, perform band-pass filtering on the signal from 1 to 75 Hz through a Butterworth filter to remove the power frequency interference generated by the EEG acquisition device and the myoelectric interference of the subject. Then, downsample the EEG data to 200 Hz to ensure data quality and computational efficiency. Next, segment the EEG signal according to a predetermined time window size T (set to 4 seconds in the present invention) to obtain multiple segments, and label the corresponding subject identity label for each segment. The data is divided into a training set and a test set according to the order of the acquisition periods, with the data collected first as the training set and the data collected later as the test set.

[0087] Step S3: Extract features from the preprocessed EEG data through a deep learning network introducing a normalization compensation mechanism;

[0088] The deep learning network based on the introduced normalization compensation mechanism includes: a basic feature extraction module, an instance normalization compensation module, a batch normalization compensation module, and a layer normalization compensation module, and each module undertakes different levels of feature adjustment tasks;

[0089] The basic feature extraction module is responsible for extracting the primary feature X containing spatio-temporal information from the preprocessed EEG data out , covering spatio-temporal information, and using means such as convolutional layers to capture the feature representations in the frequency domain, time domain, and spatial domain from the original signal, providing rich input features for the subsequent processing modules. Exemplarily, the basic feature extraction module includes a first convolutional neural network arranged in a stack and a second convolutional neural network

[0090] The first convolutional neural network layer captures the timing information from the preprocessed EEG data to obtain a timing feature map;

[0091] The second convolutional neural network layer captures the spatial domain information from the timing feature map to obtain the primary feature X containing spatio-temporal information out ;

[0092] Denote the input preprocessed EEG data where C represents the number of electrodes and L represents the input duration;

[0093] The size of the convolutional kernel in the first convolutional neural network is determined by the input sample duration L and the proportionality coefficient α, and is specifically expressed as:

[0094] K=(1,α·L)

[0095] Specifically, the proportionality coefficient α is set to 0.5, that is, the length of the convolutional kernel is half of the input duration.

[0096] The convolutional kernel of the second convolutional neural network is set as:

[0097] S=(C,1)

[0098] where C represents the number of electrodes.

[0099] Therefore, the primary feature X processed by the basic feature extraction module out can be expressed as:

[0100]

[0101] The instance normalization compensation module is responsible for based on the compensation mechanism for the primary feature Xout Process each feature map instance in it; it includes an instance internal correction unit, a power calculation unit, and a fusion unit;

[0102] The instance internal correction unit is responsible for correcting the primary feature X out inside each feature map instance, reducing the bias between the training and test distributions, and ensuring the consistency of feature intensity; it includes an instance centering correction operation and an instance scaling correction operation;

[0103] The instance centering correction operation means that before the traditional instance normalization operation, use the trainable parameters W S1 , b S1 to filter X out to obtain the feature X in after instance centering correction;

[0104]

[0105] Among them, W S1 and b S1 respectively represent learnable parameters, represents the relu activation function, represents the dot product operation;

[0106] The instance scaling correction operation means that after the instance centering correction of the feature X in , use the activation function to obtain the weight after the traditional instance normalization, limit the feature intensity inside the instance, and obtain the feature

[0107]

[0108] after instance scaling correction. Among them, IN(·) represents the traditional instance normalization operation, δ(·) represents the sigmoid activation function, W S2 and b S2 respectively represent learnable parameters, while γ in and β in respectively represent the learnable parameters of the affine transformation in the traditional instance normalization operation, u1(·) and σ1(·) respectively represent the mean operation and variance operation in the traditional instance normalization operation;

[0109] The power calculation unit is responsible for calculating the power of each instance of the primary feature through a square operation, further amplifying the discriminative key information (i.e., features with large differences), and then using average pooling operation to reduce the dimension of the power feature, so as to retain the key information, improve the discriminability of the feature, make up for the class-related features that may be lost in the normalization process, strengthen the uniqueness of the instance and the ability to distinguish classes, and obtain the enhanced feature X power ;

[0110]

[0111] Among them, represents a square operation, represents an average pooling operation, where the pooling stride step = 4, represents a logarithmic transformation operation;

[0112] This process not only strengthens the expression of key differences in the feature map but also retains the uniqueness of each instance.

[0113] The fusion unit is responsible for concatenating the output of the instance internal correction unit the output X of the power calculation unit power to obtain the instance correction feature X S ;

[0114] Before concatenation, the output of the instance internal correction unit needs to go through an average pooling operation with a consistent stride to keep the output lengths of the two units the same, which can be specifically expressed as:

[0115]

[0116] Among them, CAT(·) represents the concatenation operation, represents an average pooling operation, where the pooling stride step = 4.

[0117] The batch normalization compensation module performs overall normalization on the features of the entire batch. By introducing a center correction and scaling correction mechanism, it precisely adjusts the feature distribution, avoiding the problem that traditional batch normalization ignores the differences between classes due to rough processing, thereby better retaining class-related information. Exemplarily, the batch normalization compensation module includes a batch centering correction operation and a batch scaling correction operation;

[0118] The batch centering correction operation means that before the traditional batch normalization operation, the instance correction feature X s is filtered using trainable parameters to obtain the batch centering correction feature X bn ;

[0119]

[0120] Among them, W B1 and b B1 respectively represent learnable parameters, represents the relu activation function;

[0121] The batch scaling correction operation means that after the traditional batch normalization, weights are obtained using an activation function to limit the feature intensity of the entire batch, obtaining the batch correction feature X B ;

[0122]

[0123] Among them, BN(·) represents the traditional batch normalization operation, δ(·) represents the sigmoid activation function, and W B2 and b B2 represent learnable parameters respectively, while γ bn and β bn represent the learnable parameters of the affine transformation in the traditional batch normalization operation respectively, and u2(·) and σ2(·) represent the mean operation and variance operation in the traditional batch normalization operation respectively;

[0124] The layer normalization compensation module performs final optimization and adjustment on each sample in each batch. By correcting the feature distribution within a single sample, it ensures the balance between different feature dimensions, avoids the over-dominance of a single-dimensional feature, and thus enhances the generalization ability and recognition performance of the model. Exemplarily, the layer normalization compensation module includes a sample centering correction operation and a sample scaling correction operation;

[0125] The sample centering correction operation means that before the traditional layer normalization operation, the batch-corrected feature X B is filtered using trainable parameters to obtain the sample centering correction feature X ln ;

[0126]

[0127] Among them, W C1 and b C1 represent learnable parameters respectively, represents the relu activation function;

[0128] The sample scaling correction operation means that after the traditional layer normalization, weights are obtained using the activation function to limit the feature intensity of each sample, and the sample correction feature X c is obtained;

[0129]

[0130] Among them, LN(·) represents the traditional layer normalization operation, δ(·) represents the sigmoid activation function, and W C2 and b C2 represent learnable parameters respectively, while γ ln and β ln represent the learnable parameters of the affine transformation in the traditional layer normalization operation respectively, and u3(·) and σ3(·) represent the mean operation and variance operation in the traditional layer normalization operation respectively;

[0131] The instance normalization compensation module, batch normalization compensation module, and layer normalization compensation module are mainly responsible for overall adjustment of the data, aiming to overcome the problem of ignoring inter-class differences due to rough processing in the traditional batch normalization method. In the traditional method, the normalization operation based on the statistics of the entire batch of data may lead to the loss of local statistical characteristics, thereby reducing the sensitivity of the model to class differences. To this end, the batch normalization compensation module introduces a calibration mechanism, which performs center calibration and scaling calibration before and after normalization respectively, to more precisely retain the class discriminant information. Finally, the features processed by the batch normalization compensation module are sent to the layer normalization compensation module for further optimization. The layer normalization compensation module normalizes the feature intensity within each sample, ensuring a more balanced distribution among different feature dimensions and avoiding the over-dominant effect of a single-dimensional feature. As the network structure closest to the classifier, the layer normalization compensation module can directly access the key information closely related to identity recognition during the training process. This mechanism not only effectively retains the class-related features but also further captures the statistical characteristics closely related to identity discrimination, thus significantly improving the discriminative ability and generalization performance of the model. Through the hierarchical processing of the above modules, cross-temporal brainprint identity recognition is finally achieved by the classifier.

[0132] Step S4 of the further setting of the present invention is to train the network model. To optimize the performance of the network, the present invention adopts the backpropagation algorithm, and continuously iteratively updates the network parameters until the model reaches the preset performance standard on the training data. Specifically, the finally obtained sample corrected feature X c , is trained by the classifier by minimizing the cross-entropy, which is specifically expressed as:

[0133]

[0134] where, represents the i-th sample in this batch, and its corresponding identity label is y (i) , E g (·) represents the fully connected layer, which is responsible for flattening and inputting it into the classifier .

[0135] The cross-time brainprint recognition method proposed by the present invention has been tested on two cross-time datasets (including 30 and 54 subjects respectively), verifying its effectiveness in cross-time identity recognition and authentication scenarios. The experiment uses the EEG dataset collected based on the SSVEP protocol and compares it with the classical methods in the field of brain-computer interface and other recently proposed methods to evaluate the performance in cross-time recognition and cross-time identity authentication. Especially in cross-time identity authentication, there are two cases: known attacker (i.e., the attacker's data is included in the training set) and unknown attacker (i.e., the attacker's data is not included in the training set). Through comparative analysis, the recognition accuracy, F1 score, and equal error rate (EER) in the authentication scenario of each method are evaluated. The experimental results and comparison data are shown in Tables 1 and 2.

[0136] Table 1 Cross-time Identity Recognition Results

[0137]

[0138]

[0139] Table 2 Cross-time Identity Authentication Results (Equal Error Rate EER%)

[0140]

[0141] This embodiment also provides a cross-time brainprint identity recognition device based on the above method, including:

[0142] A data acquisition module, responsible for acquiring EEG data;

[0143] A data preprocessing module, responsible for preprocessing the acquired EEG data;

[0144] A recognition module, responsible for inputting the preprocessed EEG data into a deep learning network and a classifier based on the introduction of a normalization compensation mechanism to obtain the cross-time brainprint identity recognition result.

[0145] This embodiment also provides an electronic device. Specifically, the electronic device includes a memory and a processor. An executable code is stored in the memory, and when the processor executes the executable code, the method described in any one of the above embodiments is implemented.

[0146] Among them, the memory may include a high-speed random access memory (RAM, Random Access Memory), and may also include a non-volatile memory (Non-volatile Memory), such as at least one disk memory. Through at least one communication interface (which can be wired or wireless), the communication connection between this system network element and at least one other network element can be realized, and the Internet, wide area network, local area network, metropolitan area network, etc. can be used.

[0147] The bus can be an ISA bus, a PCI bus, an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0148] Among them, the memory is used to store a program. After receiving an execution instruction, the processor executes the program. The method executed by the device defined by the flow process disclosed in any embodiment of the foregoing embodiments of the present invention can be applied to or implemented by the processor.

[0149] 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 the integrated logic circuit in the hardware of the processor or the instructions in the form of software. The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can 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 gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present invention can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art 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. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0150] The computer program product of the readable storage medium provided by the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For the specific implementation, reference can be made to the foregoing method embodiments and will not be elaborated here.

[0151] If the above-mentioned 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, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0152] Finally, it should be noted that the above-mentioned embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or easily conceive of changes, or perform equivalent replacements for some of the technical features; and these modifications, changes, or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for cross-period brain pattern recognition based on a deep learning network with a normalized compensation mechanism, the method comprising: Step S1, EEG data collection; Step S2, preprocessing the EEG data; Step S3, extracting features from the preprocessed EEG data by introducing a deep learning network based on a normalization compensation mechanism; Step S4: using a classifier to classify and identify the features output by the deep learning network based on the introduction of the normalization compensation mechanism; It is characterized in that the deep learning network based on the introduction of normalization compensation mechanism includes a basic feature extraction module, an instance normalization compensation module, a batch normalization compensation module, and a layer normalization compensation module; The basic feature extraction module is responsible for extracting primary features X containing spatiotemporal information from the preprocessed EEG data. out ; The instance normalization compensation module is responsible for compensating the primary feature X based on the compensation mechanism. out Each feature map instance in is processed to obtain the instance correction feature X S ; The batch normalization compensation module is responsible for correcting the feature X based on the instance S The overall information of all batches is further adjusted to obtain the batch correction feature X B ; The layer normalization compensation module is responsible for batch correction feature X B Perform the final optimization adjustment on each sample in each batch to obtain the sample correction feature X c .

2. The method according to claim 1, characterized in that: The basic feature extraction module includes a first convolutional neural network And the second convolutional neural network The first convolutional neural network layer Capturing the time series information of the preprocessed EEG data to obtain a time series feature graph; The second convolutional neural network layer Capture the spatial domain information of the time series feature graph and obtain the primary feature X containing time and space information out .

3. The method according to claim 1, characterized in that: The instance normalization compensation module includes an instance internal correction unit, a power calculation unit, and a fusion unit; The internal correction unit of the instance is responsible for the primary feature X out Each feature map instance in is normalized internally to reduce the deviation between the training and test distributions by adjusting the mean and variance to ensure the consistency of feature strength; The power calculation unit is responsible for further amplifying the key information with discriminativeness by calculating the power of the primary features, thereby compensating for the category-related features that may be lost during the normalization process, strengthening the uniqueness and category distinction ability of the instance, and obtaining the enhanced feature X power ; The fusion unit is responsible for converting the output of the instance internal correction unit The output of the power calculation unit X power Concatenate to get instance correction feature X S .

4. The method according to claim 3, characterized in that: The instance internal correction unit includes an instance centering correction operation and an instance scaling correction operation; The instance centering correction operation refers to using a trainable parameter W before the traditional instance normalization operation. S1 、b S1 X out Filter to obtain the instance-centered corrected feature X in ; Among them, W S1 and b S1 They represent the learnable parameters, represents the relu activation function, Represents the dot product operation; The instance scaling correction operation refers to the correction of the instance center feature X in After the traditional instance normalization, the activation function is used to obtain the weight, which limits the feature strength inside the instance and obtains the instance scaled and corrected feature. Where IN(·) represents the traditional instance normalization operation, δ(·) represents the sigmoid activation function, and W S2 and b S2 denote the learnable parameters, and γ in and β in They represent the learnable parameters of the affine transformation in the traditional instance normalization operation, u1(·) and σ1(·) represent the mean operation and variance operation in the traditional instance normalization operation, respectively.

5. The method according to claim 3, characterized in that: The implementation process of the power calculation unit is: in, represents the square operation, represents the average pooling operation, represents the logarithmic transformation operation; The implementation process of the fusion unit is: Among them, CAT(·) represents the concatenation operation, Represents an average pooling operation.

6. The method according to claim 1, characterized in that: The batch normalization compensation module includes a batch centering correction operation and a batch scaling correction operation; The batch centering correction operation refers to using trainable parameters to correct the instance feature X before the traditional batch normalization operation. s Filter to obtain batch centered correction feature X bn ; Among them, W B1 and b B1 They represent the learnable parameters, Represents the relu activation function; The batch scaling correction operation refers to using the activation function to obtain weights after the traditional batch normalization, limiting the feature strength of the entire batch, and obtaining the batch correction feature X B ; Among them, BN(·) represents the traditional batch normalization operation, δ(·) represents the sigmoid activation function, and W B2 and bB2 represent learnable parameters, and γ bn and β bn They represent the learnable parameters of the affine transformation in the traditional batch normalization operation, u2(·) and σ2(·) represent the mean operation and variance operation in the traditional batch normalization operation, respectively.

7. The method according to claim 1, characterized in that: The layer normalization compensation module includes a sample centering correction operation and a sample scaling correction operation; The sample centering correction operation refers to using trainable parameters to correct the batch feature X before the traditional layer normalization operation. B Filter to obtain the sample center correction feature X ln Among them, W C1 and b C1 They represent the learnable parameters, Represents the relu activation function; The sample scaling correction operation refers to using the activation function to obtain the weight after the traditional layer normalization, limiting the feature strength of each sample, and obtaining the sample correction feature X c ; Where LN(·) represents the traditional layer normalization operation, δ(·) represents the sigmoid activation function, and W C2 and b C2 denote the learnable parameters, and γ ln and β ln They represent the learnable parameters of the affine transformation in the traditional layer normalization operation, u3(·) and σ3(·) represent the mean operation and variance operation in the traditional layer normalization operation, respectively.

8. A cross-time brainprint identity recognition device 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 the collected EEG data; The recognition module is responsible for inputting the pre-processed EEG data into the deep learning network and classifier based on the introduction of the normalization compensation mechanism to obtain the cross-time brainprint identity recognition results.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to operate according to the method described in any one of claims 1 to 7, thereby achieving cross-time period identity recognition.

10. A computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it performs cross-time brainprint recognition according to the method according to any one of claims 1 to 7.

Images