Single-source-domain adaptive cross-time-period brain grain recognition method

By using multi-scale feature extraction and domain discriminator adversarial modules in EEG identity recognition, the problems of weak data generalization ability and insufficient EEG marker data are solved, and cross-time segments under single source domain conditions are realized EEG signal recognition improves the accuracy and stability of recognition.

CN118968560AActive Publication Date: 2024-11-15HANGZHOU DIANZI UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411144440.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-15
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

Traditional EEG identity recognition technology faces the problems of weak data generalization ability across time periods and insufficient EEG marker data, which leads to the difference in signal distribution of models in different time periods and between individuals that affects the accuracy of identification, and traditional deep learning networks ignore individual differences. and data distribution changes, it is difficult to adapt to complex EEG signals.

Method used

The single-source domain adaptation cross-time brain pattern recognition method is adopted, and the multi-scale feature extraction module, domain discriminator adversarial module, association domain adaptive module and linear classification module are combined with time convolution layer, spatial convolution layer, and depth separable volumes. Layering and attention mechanisms, data alignment and feature extraction, and a cross-time brain pattern recognition model is constructed to achieve adversarial training and aggregation of source and target domain features.

Benefits of technology

Under the single source domain conditions, it can identify EEG signals at different times, enhance the robustness and generalization ability of the model, extract domain invariant features, improve the accuracy and stability of identity prediction, and reduce the over-response of redundant information , improve the robustness of the characteristics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968560A_ABST
    Figure CN118968560A_ABST
Patent Text Reader

Abstract

The invention discloses a single-source domain adaptive cross-time-period brain grain recognition method, which comprises the following steps of: processing data of a subject through data alignment, and constructing a cross-time-period brain grain recognition model by using a domain identifier confrontation module and a correlation domain adaptive module; under the condition that electroencephalogram data of only one time period serves as a source domain, electroencephalogram signals of different time periods can still be recognized, and the problems that in traditional brain grain recognition, the generalization ability of cross-time-period data is poor, and electroencephalogram mark data is insufficient are solved; meanwhile, a data alignment mode adopted by the cross-time-period brain grain recognition method does not need a tested label, transformation is directly carried out on original electroencephalogram data, the original electroencephalogram data is made to adapt to the changes, domain invariant features are extracted, and therefore identity prediction of the single-source domain adaptive cross-time-period electroencephalogram signals is achieved. Besides, a self-attention mechanism is introduced in the cross-period brain grain recognition method to recognize key features, excessive response to redundant information is reduced, and therefore the robustness of the features is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-time period single-source domain adaptive brain pattern recognition, and specifically relates to a single-source domain adaptive cross-time period brain pattern recognition method. Background Art

[0002] As an advanced means for information security, biometric recognition technology has received more and more attention. However, traditional biometric recognition technology still has some challenges. Physiological features such as fingerprints, irises, and faces are easily stolen, copied, and tampered with. At the same time, behavioral characteristics are easily imitated and attacked in the surveillance environment of public places. Compared with existing traditional biometric recognition technologies, brainprint features based on EEG have the characteristics of high concealment, non-counterfeitability, continuous authentication, and must be alive. Thanks to the above characteristics, brainprint recognition technology can be applied to application fields with high security requirements such as military, finance, and medical security.

[0003] EEG identity recognition faces many technical and practical challenges in practical applications. The first is the weak generalization ability of cross-time data. Since the position of the EEG cap is different each time the data is collected, there is a channel position offset phenomenon, that is, the electrode position will change at different times and scenes, resulting in inconsistent EEG signals. In addition, EEG signals will change with the change of time, emotions, fatigue level and other factors, resulting in large differences in the signals collected by the same test subject at different time points, affecting the stability and accuracy of recognition. In this way, the time periods of training data and test data may not obey the same distribution, which in turn affects the accuracy of identity recognition. The second is the lack of EEG labeled data. The EEG labeled data across time periods is very limited, and the cost of obtaining a large amount of labeled data is too high and time-consuming. In order to cope with the limited labeled training data, many people try to directly apply the model trained on one source domain to another unlabeled target domain. Unfortunately, due to the existence of domain offset or data set bias, the distribution difference between the source domain and the target domain will lead to a decrease in the performance of the model on the target domain, and direct migration across domains is usually not effective. Domain adaptation technology uses various methods to reduce this distribution difference, so that the model trained on the source domain can be better generalized to the target domain.

[0004] In addition, for complex EEG signals, traditional deep learning networks usually only use a fixed convolution kernel for processing. This method has certain limitations because it ignores the individual differences between different subjects and the changes in data distribution in different time periods. This method that does not consider the differences in data distribution may make it difficult for the model to adapt to changing signals in practical applications, thereby affecting the accuracy of recognition and classification. Therefore, how to design more flexible and adaptive convolution kernels to process and capture these complex signal changes is a major challenge in current research. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a single-source domain adaptive cross-time period brain pattern recognition method to learn stable domain invariant features, and predicts the identity information of the subject from a plurality of pre-set target object identities through newly collected EEG data of the subject.

[0006] The present invention proposes a single-source domain adaptive cross-time period brainprint recognition method, comprising the following steps: Step 1: Collect EEG data of multiple subjects and add labels as the source domain; then collect EEG data of different time periods from the source domain as the target domain; the target domain does not require labels.

[0007] Step 2: Construct a cross-time brainprint recognition model. The cross-time brainprint recognition model includes a multi-scale feature extraction module, a domain discriminator adversarial module, an associated domain adaptation module, and a linear classification module; the feature extraction module includes a temporal convolution layer, a spatial convolution layer, a depth-separable convolution layer, and an attention mechanism module connected in sequence; extract features through the temporal convolution layer, the spatial convolution layer, and the depth-separable convolution layer, obtain the initial features and input them into the attention mechanism module to obtain the final output features. The final output features are input into the domain discriminator adversarial module, the associated domain adaptation module, and the linear classification module respectively; the domain discriminator adversarial module is used to perform adversarial training on the final output features of the source domain and the target domain; the associated domain adaptation module is used to cluster source domain samples of the same category in the feature space.

[0008] Step 3: Use the data set obtained in step 1 to train the cross-time brain pattern recognition model constructed in step 2.

[0009] Step 4: Collect the EEG data of the subject and input it into the trained cross-time brain pattern recognition model to identify the subject's identity.

[0010] Preferably, in step 1, the method of data alignment is as follows: Get n EEG samples X of the subject i The mean matrix R of the covariance matrix is ​​expressed as: in, For EEG sample X i transpose of ; i=1,2,...,n.

[0011] Get aligned EEG samples according to the mean matrix R : Preferably, in the step 3, the expression of the loss function L for training the cross-period brain pattern recognition model is: Among them, L cls is the cross entropy loss; L domain is the domain discriminator loss; L association is the association domain adaptation loss.

[0012] Preferably, the cross entropy loss L cls , domain discriminator loss L domain and the cross entropy loss L cls The expressions are: in, and are the domain discriminator losses for the source domain and the target domain respectively; L visit and L walker are visitor loss and pedestrian loss respectively; α is the weight of pedestrian loss, α=0.6; is the cross entropy loss; y is the true subject label; is the predicted subject label; N is the number of subjects.

[0013] Domain Discriminator Loss and The expression is: Among them, n s and n t are the number of samples in the source domain and the number of samples in the target domain respectively; D s and D t are the source domain of the training set and the target domain of the test set respectively; G d is the domain identifier; G f is the feature extraction module; d i For the input sample x i The corresponding domain label.

[0014] Visitor loss L visit and the Pacers lost L walker The expression is: in, ; ; A i and B j are the output features of the source domain samples and the target domain samples after the feature extraction network, respectively; class(·) is the subject category label corresponding to the output feature; From the output feature A i Transfer to output feature B j The transition probability of From the output feature A i Transfer to output feature B j Then return the output feature A i The probability of a two-step round trip.

[0015] Transition probability and the two-step round trip probability The expression is: Among them, M ij The output feature A i and B j The dot product of Output feature B j Return to output feature A i probability.

[0016] Preferably, in the step three, the cross-period brainprint recognition model is trained by inputting labeled source domain samples and unlabeled target domain samples into the cross-period brainprint recognition model.

[0017] Preferably, in the step 2, the temporal convolution layer uses three convolution kernels of different lengths to extract temporal features, the heights of the three convolution kernels are all 1, and the lengths are respectively the number of sample time points × 0.5, the number of sample time points × 0.25, and the number of sample time points × 0.125.

[0018] Preferably, in the step 2, the attention mechanism module includes a global pooling layer and two fully connected layers connected in sequence.

[0019] Preferably, in the step 2, a BatchNorm layer, an ELU layer, an AvgPool layer and a Dropout layer are added after the temporal convolution layer, the spatial convolution layer and the depth-separable convolution layer.

[0020] In a second aspect, the present invention provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; the processor executes the aforementioned single-source domain adaptive cross-time period brainprint recognition method.

[0021] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned single-source domain adaptive cross-time period brainprint recognition method.

[0022] The present invention has the following beneficial effects: 1. The present invention processes the data of the subjects through data alignment, and uses the domain discriminator adversarial module and the associated domain adaptive module to construct a cross-time brainprint recognition model. Compared with the traditional brainprint recognition which requires EEG data of multiple time periods as the source domain, the present invention can still recognize EEG signals of different time periods when only EEG data of one time period is used as the source domain, thus solving the problems of weak generalization ability of cross-time data and insufficient EEG marker data in traditional brainprint recognition.

[0023] 2. The data alignment method adopted by the present invention does not require subject labels, but directly transforms the original EEG data, which can enhance the robustness of the model, enable it to adapt to these changes and extract domain-invariant features, thereby realizing identity prediction of EEG signals across time periods in a single-source domain.

[0024] 3. The self-attention mechanism introduced in the present invention can identify key features in time series data, help the model dynamically adjust the weights of different convolution kernels to adapt to different domain data with distribution differences, and through the compression incentive mechanism, help the model identify and amplify the channel information and time domain information features related to the task, reduce excessive response to redundant information, and thus improve the robustness of the features.

[0025] 4. The present invention introduces a domain discriminator adversarial module, so that the source domain features and the target domain features are mapped to the same feature space through the feature extractor. The domain discriminator attempts to distinguish the features of the source domain and the target domain, while the feature extractor attempts to deceive the domain discriminator so that it cannot distinguish between the two types of features. Through this adversarial learning method, the feature distributions of the source domain and the target domain are as similar as possible, thereby reducing the distribution difference between them.

[0026] 5. The present invention calculates the association matrix between source domain and target domain samples through an association domain adaptation module, and minimizes the KL divergence between the distribution of source domain samples and the target domain category distribution, so that source domain samples of the same category are more clustered together and match the category distribution of target domain samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall flow chart of the present invention.

[0028] Figure 2 Schematic diagram of the overall architecture of the cross-time brain pattern recognition model in the present invention.

[0029] Figure 3 Schematic diagram of the feature extraction module in the present invention.

[0030] Figure 4 Schematic diagram of the attention mechanism module in the present invention.

[0031] Figure 5Schematic diagram of the domain identifier adversarial module in the present invention.

[0032] Figure 6 Schematic diagram of the association domain adaptation module in the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings.

[0034] like Figure 1 As shown, a single-source domain adaptive cross-time brainprint recognition method includes the following steps: Step 1: Get the dataset The dataset uses the collaborative brain-computer interface cross-period dataset based on rapid serial visual presentation (RSVP) from Tsinghua University. In this dataset, a total of 14 subjects participated in two identical experiments, with an average interval of about 23 days. The 14 subjects were divided into 14 groups in pairs. Each trial contained 100 images (10 seconds, 10Hz rate), including four target images. The images displayed in the first 1 second and the last 1 second of a trial were non-target images to avoid the target image appearing during the onset or offset of the steady-state visual evoked potential (SSVEP) induced by RSVP. The interval between the two target images was at least 500 milliseconds to reduce the impact of attention flicker. The subjects were asked to press the key immediately after detecting the target. The key-pressing task was used to make the subjects focus on target detection. Since there is a time delay between the target image and the key press, during the experiment, the key press within 500ms after the target image was considered a correct response to the target image. In the experiment, the subjects needed to find four targets from 100 images and make four key presses. If the subject missed some targets, the system would display the missed targets at the end of the trial. The experiment consisted of two experiments on different days, in which the stimulation paradigm was exactly the same. EEG data of the two subjects were recorded simultaneously by two Neuroscan Synamps2 systems. A 64-electrode EEG cap based on the 10-20 system was used to record 62 channels of EEG data from the two subjects (M1 and M2 were not used). The reference electrode was located at the vertex. The impedance of the electrode was kept below 10K, and the sampling rate was 1,000 Hz. The dataset was preprocessed, which included filtering, re-referencing, and data segmentation. A 50 Hz notch filter was used to remove common power line noise. The passband of the amplifier was set to 0.15 Hz to 200 Hz. All event triggers were transmitted through the parallel port and marked on the EEG data. All EEG data corresponding to the target image were selected, using all channels, the filtering frequency band was 1 Hz to 40 Hz, and 0.5 s of data were selected as a sample, with 500 time steps for each sample. The samples obtained from the two experiments were used as the training set and the test set, respectively. The training set and the test set were used as the source domain and the target domain, respectively.

[0035] Step 2: Data Alignment Each subject has n preprocessed EEG samples X i , calculate the covariance matrix of each EEG sample point, and obtain the mean matrix R of n covariance matrices. The expression of the mean matrix R is: in, For EEG sample X i transpose of ; i=1,2,...,n.

[0036] Get aligned EEG samples according to the mean matrix R : Aligned EEG samples With the original EEG sample X i The dimensions of are the same. EEG samples after alignment of any subject The mean matrix of the covariance matrix They are all unit matrices, and the overall distribution is more consistent.

[0037] Where I is the identity matrix.

[0038] Step 3: Build a cross-period brain pattern recognition model like Figure 2 , 3 As shown in Figure 4, the cross-time brain pattern recognition model is a multi-scale attention feature fusion network, which includes a multi-scale feature extraction module, a domain discriminator adversarial module, an associated domain adaptation module and a linear classification module; the feature extraction module is used to extract EEG features, and the feature extraction module includes a temporal convolution layer, a spatial convolution layer, a depth-separable convolution layer and an attention mechanism module based on a compression excitation mechanism; the temporal convolution layer uses three convolution kernels of different lengths to extract temporal features; the depth-separable convolution layer inputs the output initial features into the attention mechanism module to obtain the corresponding feature weights; the attention mechanism module includes a global pooling layer and two fully connected layers connected in sequence. The global pooling layer globally pools the features input into the attention mechanism module and compresses the input features into a feature vector whose length is the number of channels to obtain a larger receptive field; the two fully connected layers obtain a feature weight matrix composed of multiple channels; the initial features are multiplied and flattened by the feature weight matrix to obtain the final output features. The final output features are respectively input into the domain discriminator adversarial module, the associated domain adaptation module and the linear classification module. The linear classification module uses the Kan classifier to classify the subject's identity according to the final features.

[0039] like Figure 5As shown in the figure, the domain discriminator adversarial module inputs the final output features of the source domain and the final output features of the target domain respectively for adversarial training. In the domain adversarial training stage, gradient reversal is used to achieve domain confusion, so that the feature extractor cannot distinguish that the source domain and target domain samples are from two different domains, thereby extracting stable domain invariant features.

[0040] like Figure 6 As shown in the figure, the association domain adaptation module calculates the association matrix between the source domain and the target domain samples, inputs the labeled source domain output features and the unlabeled target domain final output features, and calculates the conditional probability distribution P from the source domain to the target domain by performing matrix multiplication on the source domain and target domain features. st And the conditional probability distribution P from the target domain to the source domain ts , generate the correlation matrix P between source domain samples sts And the category distribution P of the target domain samples t By minimizing the association matrix between source domain samples and target domain category distribution P t The KL divergence (‌Kullback-Leibler Divergence)‌ between them is used to achieve the aggregation of source domain samples of the same category in the feature space.

[0041] Step 4: Training the model 4-1. Input the labeled source domain samples and the unlabeled target domain samples into the temporal convolution layer, spatial convolution layer and depth-separable convolution layer to obtain the initial features. After the initial features are input into the attention mechanism module, the weight matrix is ​​obtained; after the weight matrix is ​​weighted on the initial features, the final features are obtained.

[0042] 4-2. Input the final feature map into the domain discriminator adversarial module, the associated domain adaptation module and the linear classification module to obtain the domain discriminator loss L domain , correlation domain adaptive loss L association and the cross entropy loss L cls , whose expressions are: in, and are the domain discriminator losses for the source domain and the target domain respectively; L visit and L walker are visitor loss and pedestrian loss respectively; α is the weight of pedestrian loss, α=0.6; is the cross entropy loss; y is the true subject label; is the predicted subject label; N is the number of subjects.

[0043] Domain Discriminator Loss and The expression is: Among them, n s and n t are the number of samples in the source domain and the number of samples in the target domain respectively; D s and D t are the source domain of the training set and the target domain of the test set respectively; G d is the domain identifier; G f is the feature extraction module; d i For the input sample x i The corresponding domain label.

[0044] Visitor loss L visit Ensure that the category distribution of the target domain samples is as uniform as possible by minimizing the target domain category distribution (predicted by the source domain samples) and the uniform distribution p visit The KL divergence between the two is realized; the walker loss L walker It aims to enable the model to align features between different domains by minimizing the KL divergence between the predicted category distribution and the target category distribution. Visitor loss L visit and the Pacers lost L walker The expression is:

[0045] in, ; ; A i and B j are the output features of the source domain samples and the target domain samples after the feature extraction network; class(·) is the subject category label corresponding to the output feature; From the output feature A i Transfer to output feature B j The transition probability of From the output feature A i Transfer to output feature B j Then return the output feature A i The probability of a two-step round trip.

[0046] Transition probability and the two-step round trip probability The expression is: Among them, M ij The output feature Ai and B j The dot product of Output feature B j Return to output feature A i probability.

[0047] 4-3. Construct the loss function L of the cross-time brain pattern recognition model: The loss function L is minimized through cyclic iteration, and the final cross-time brain pattern recognition model is obtained.

[0048] Step 5: Evaluate the model In order to compare with the existing technology, the top-1 classification accuracy is used as the performance metric. All experimental results are the average of 5 tests with different random seeds. The temporal convolution layer, spatial convolution layer and depth-separable convolution layer are all 2D convolution layers. The convolution kernel size of the temporal convolution layer is (1,250), the convolution kernel size of the spatial convolution layer is (62,1), and the depth-separable convolution kernel size is (1,1). BatchNorm layer, ELU layer, AvgPool layer and Dropout layer are inserted after each convolution layer, and the Dropout coefficient is 0.25. The model is trained using the Adam optimizer with a learning rate of 5e-4, a batch size of 8, and a total of 200 epochs.

[0049] The models used for comparative experiments with the present invention are divided into the following three categories: (1) traditional deep learning models, which refer to deep learning methods that do not contain domain adaptation structures; (2) single-source domain adaptation models, which have only one source domain; (3) multi-source domain adaptation models, which align the data distributions of different source domains and target domains respectively. The specific comparative models are introduced below. The EEGNet model first uses a one-dimensional convolution operation to extract the temporal features of EEG, and then uses deep separable convolution to extract EEG frequency-space features. This model is a traditional deep learning model and has been applied to brain pattern recognition. The CNNRNN model uses a two-dimensional convolutional neural network CNN and a bidirectional recurrent neural network BiLSTM structure to extract EEG frequency-space features and time-dependent temporal features respectively. This model is a traditional deep learning model and has been proposed for brain pattern recognition. BrainNet uses a one-dimensional convolution operation to extract EEG time-series features, and then uses triplet loss (TripletLoss) to make samples of the same type closer to each other and samples of different types farther away from each other. MEERNet is a multi-source domain adaptation model that considers both domain-invariant features and domain-specific features, and is proposed for cross-time and cross-subject EEG emotion recognition. MTDANN is a multi-source domain adaptation method for EEG signals, which consists of EEGNet and a domain discrimination network, and aims to learn domain-invariant representations between source and target domains.

[0050] Since there is no single-source domain adaptation model in the traditional model, in order to better compare the traditional model with the model of the present invention, the method of the present invention is used to transform the traditional model into a single-source domain adaptation model for comparison. The comparative experimental results are shown in Table 1 below. On the cross-period data set of the collaborative brain-computer interface based on rapid serial visual presentation (RSVP), the accuracy of the present invention in the test set (the second period) reached 84.37%. It has a great advantage over traditional deep learning models and multi-source domain adaptation models; in the single-source domain adaptation model, the average accuracy of the present invention is improved by 13.86% compared with EEGNet, and by 3.42% compared with BrainNet. It can be seen that the recognition result of the present invention is more accurate.

[0051] Table 1 Test results of different models We conducted multiple ablation experiments on the overall model to evaluate the impact of each module on the model performance. By removing or modifying different modules and observing the changes in model accuracy, we can understand the role of each module in the overall model. The results of the ablation experiment are shown in Table 2.

[0052] Table 2 Ablation experiment Model Accuracy Overall Model 84.37% Domain Discriminator Adversarial Module 34.15% No-correlation domain adaptation module 77.34% No attention mechanism module 83.65% The accuracy of the overall model is 84.37%, and the model performance is better than the existing single-source domain adaptation method and multi-source domain adaptation method, which proves the effectiveness of the transferable feature learning of the present invention.

[0053] After removing the domain discriminator adversarial module, the model accuracy dropped significantly to 34.15%, which means that domain discriminator adversarial learning can learn rich stable and invariant features, narrow the distribution differences between the source domain and the target domain through adversarial training, help the model better adapt to the distribution differences of different domains, and enhance the generalization ability of the model.

[0054] After removing the association domain adaptation module, the model accuracy dropped to 77.34%, which means that the association domain adaptation module can enhance the feature association between the source domain and the target domain, thereby helping the model to better adapt to the data distribution of the target domain, more effectively transfer the knowledge of the source domain to the target domain, and improve the performance of the model on the target domain.

[0055] After removing the attention mechanism module, the model accuracy dropped to 83.65%, which means that the introduction of the attention mechanism provides the model with more fine-grained information processing capabilities, which can effectively identify and highlight important features related to the task.

Claims

1. A single-source domain adaptive cross-time brainprint recognition method, characterized by: The following steps are involved: Step 1: Collect EEG data of multiple subjects and add labels as the source domain; then collect EEG data of different time periods from the source domain as the target domain; Step 2: Construct a cross-period brain pattern recognition model; The cross-time brainprint recognition model includes a multi-scale feature extraction module, a domain discriminator adversarial module, a correlation domain adaptation module and a linear classification module; the feature extraction module includes a temporal convolution layer, a spatial convolution layer, a depth-separable convolution layer and an attention mechanism module connected in sequence; features are extracted through the temporal convolution layer, the spatial convolution layer and the depth-separable convolution layer, the initial features are obtained and input into the attention mechanism module to obtain the final output features; The final output features are input into the domain discriminator adversarial module, the associated domain adaptation module and the linear classification module respectively; the domain discriminator adversarial module is used to perform adversarial training on the final output features of the source domain and the target domain; The associated domain adaptation module is used to cluster source domain samples of the same category in the feature space; Step 3: Use the data set obtained in step 1 to train the cross-period brain pattern recognition model constructed in step 2; Step 4: Collect the EEG data of the subject and input it into the trained cross-time brain pattern recognition model to identify the subject's identity.

2. According to claim 1, a single-source domain adaptive cross-time brain pattern recognition method is characterized by: In the step 1, the method of data alignment is as follows: Get n EEG samples X of the subject i The mean matrix R of the covariance matrix is ​​expressed as: in, For EEG sample X i transpose of ; i=1,2,...,n; Get aligned EEG samples according to the mean matrix R : 。 3. According to claim 1, a single-source domain adaptive cross-time brain pattern recognition method is characterized by: In the step 3, the expression of the loss function L for training the cross-period brain pattern recognition model is: Among them, L cls is the cross entropy loss; L domain is the domain discriminator loss; L association is the association domain adaptation loss.

4. According to claim 3, a single-source domain adaptive cross-time brain pattern recognition method is characterized by: The cross entropy loss L cls , domain discriminator loss L domain and the cross entropy loss L cls The expressions are: in, and are the domain discriminator losses for the source domain and the target domain respectively; L visit and L walker are visitor loss and pedestrian loss respectively; α is the weight of pedestrian loss, α=0.6; is the cross entropy loss; y is the true subject label; is the predicted subject label; N is the number of subjects; Domain Discriminator Loss and The expression is: Among them, n s and n t are the number of samples in the source domain and the number of samples in the target domain respectively; D s and D t are the source domain of the training set and the target domain of the test set respectively; G d is the domain identifier; G f is the feature extraction module; d i For the input sample x i The corresponding domain label; Visitor loss L visit and the Pacers lost L walker The expression is: in, ; ; A i and B j are the output features of the source domain samples and the target domain samples after the feature extraction network, respectively; class(·) is the subject category label corresponding to the output feature; From the output feature A i Transfer to output feature B j The transition probability of is the two-step round trip probability; Transition probability and the two-step round trip probability The expression is: Among them, M ij The output feature A i and B j The dot product of Output feature B j Return to output feature A i probability.

5. According to claim 1, a single-source domain adaptive cross-time brain pattern recognition method is characterized by: In the step three, the cross-period brainprint recognition model is trained, and the input to the cross-period brainprint recognition model is labeled source domain samples and unlabeled target domain samples.

6. According to claim 1, a single-source domain adaptive cross-time brain pattern recognition method is characterized by: In the step 2, the temporal convolution layer uses three convolution kernels of different lengths to extract temporal features. The heights of the three convolution kernels are all 1, and the lengths are respectively the number of sample time points × 0.5, the number of sample time points × 0.25, and the number of sample time points × 0.

125.

7. The single-source domain adaptive cross-time period brainprint recognition method according to claim 1 is characterized by: In the step 2, the attention mechanism module includes a global pooling layer and two fully connected layers connected in sequence.

8. The single-source domain adaptive cross-time period brainprint recognition method according to claim 1 is characterized by: In the step 2, a BatchNorm layer, an ELU layer, an AvgPool layer, and a Dropout layer are added after the temporal convolution layer, the spatial convolution layer, and the depth-separable convolution layer.

9. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the single-source domain adaptive cross-time period brainprint recognition method as described in claim 1.

10. A readable storage medium, characterized in that: A computer program is stored; when the computer program is executed by a processor, it is used to implement a single-source domain adaptive cross-time period brainprint recognition method as shown in claim 1.

Citation Information

Patent Citations

  • Cross-period brain grain recognition method based on tensor frequency-space attention domain adaptation network

    CN115969392A

  • Cross-period electroencephalogram signal identity recognition method based on Fourier amplitude feature generalization

    CN117272271A

  • Normal form self-adaptive decoupling cross-time-period brain grain recognition method and normal form self-adaptive decoupling cross-time-period brain grain recognition system

    CN117992729A