An online brainprint recognition method for cross-conversation scenarios

Abstract

This invention belongs to the field of brainprint recognition technology in brain-computer interfaces, and discloses an online brainprint recognition method for cross-session scenarios. The method includes: collecting EEG data from K users in a single session; dividing each trial into non-overlapping EEG segments of length T; performing sliding window processing on the EEG segments at S different dynamic window scales to generate S sets of datasets at different dynamic window scales; offline training M different random seed models using each set of dynamic window scale datasets to obtain M brainprint recognition models; in a second session, inputting the online received data of length T after sliding window processing at the corresponding scale into the corresponding brainprint recognition model for user identity prediction; averaging and then weighting the prediction results to obtain the user identity prediction result, thus achieving online brainprint recognition across sessions. This invention improves the accuracy, robustness, and recognition speed of cross-session EEG signal identity recognition.

Description

Technical Field

[0001] This invention belongs to the field of brainprint recognition technology in brain-computer interfaces, and more specifically, relates to an online brainprint recognition method for cross-conversation scenarios. Background Technology

[0002] The rapid development of brain-computer interface (BCI) technology and the surge in publicly available datasets and algorithms have propelled research progress but also brought increasingly serious privacy and security issues. Electroencephalogram (EEG) signals, as unique biometrics, contain a wealth of sensitive information, such as personal accounts, preferences, and physiological states—among which identity information is particularly crucial. The uniqueness of human neural pathway patterns and genetic characteristics makes each individual's EEG signals highly variable. Driven by specific cognitive tasks (such as visual stimuli or resting states), EEG signals exhibit stability and consistency; this "brainprint" provides a reliable basis for identity recognition.

[0003] The development of brainprint recognition technology is closely related to the advancement of EEG signal decoding technology. It has evolved from traditional machine learning to deep learning, and after more than a decade of research, it has achieved relatively mature results. Early brainprint recognition methods mainly relied on manually designed features, combined with machine learning classifiers for user identification. These methods included autoregressive coefficients (AR), power spectral density (PSD), and fuzzy entropy features (FuzzEn). However, these handcrafted feature-based methods not only heavily relied on expert prior knowledge but also consumed significant time and effort in feature extraction. These limitations resulted in poor generalization ability when dealing with increased data scale or signal variations (such as cross-session data).

[0004] In the field of deep learning for brainwave recognition, recent research has mainly focused on improving the accuracy of EEG signal classification and capturing complex feature relationships. Some studies have attempted to combine the powerful learning capabilities of neural networks with prior knowledge of traditional handcrafted features. While this has improved recognition performance to some extent, it still heavily relies on handcrafted features, limiting the model's adaptability and generalization. Other studies have focused on the dynamic relationships between EEG channels, proposing to construct feature matrices using dynamic functional connectivity indices and combining them with graph convolutional neural networks for classification, successfully capturing the spatial topological characteristics of EEG signals. However, these methods still lack robustness when processing cross-session data and struggle to adapt to signal variations under different session conditions.

[0005] Furthermore, most of the existing methods mentioned above focus on offline brainprint recognition within the same session. Compared to offline brainprint recognition, online brainprint recognition scenarios are significantly more challenging and have greater value for in-depth exploration in real-world applications. Therefore, conducting online brainprint recognition for EEG signals across different sessions and improving the accuracy, robustness, and recognition speed of cross-session EEG signal identity recognition is of great significance. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides an online brainprint recognition method for cross-conversation scenarios, the purpose of which is to improve the accuracy, robustness and recognition speed of EEG signal identity recognition in cross-conversation scenarios.

[0007] To achieve the above objectives, according to a first aspect of the present invention, an online brainprint recognition method for cross-conversation scenarios is provided, comprising:

[0008] Collect EEG data from K users during a single session. Divide each trial of the EEG data into non-overlapping EEG segments of length T. Then, apply the segmented EEG segments to S different dynamic window scales [T1, T2, ..., T]. S The process involves sliding window processing to generate S sets of datasets with different dynamic window scales; where K≥2, and T is the minimum trial length of all trials in the EEG data; T j ≤T,j=1,2,…,S;

[0009] Using datasets under each dynamic window scale, M different random seed models were trained offline to obtain M brain pattern recognition models.

[0010] In session two, the EEG data of length T received online are distributed across S different dynamic window scales [T1, T2, ..., T]. S The system performs sliding window processing, inputting EEG data at different dynamic window scales into M brainprint recognition models at the corresponding dynamic window scales for user identity prediction. The average of the M prediction results at the same dynamic window scale is used as the prediction result at the current dynamic window scale. The prediction results at S different dynamic window scales are weighted and summed, and then normalized to obtain the user identity prediction result, thus realizing online brainprint recognition across sessions.

[0011] Furthermore, when weighting and summing the prediction results under S different dynamic window scales, the weighting coefficients used are:

[0012]

[0013] In the formula, ω j For the j-th dynamic window scale T jThe weighting coefficient of the prediction result is L, which is the unit online transmission length of the EEG data packet in Session 2.

[0014] Furthermore, the brainprint recognition model includes:

[0015] The convolution module is used to perform multi-scale convolution operations on each data sample under the current dynamic window scale in order to extract EEG features at different time scales.

[0016] The encoding module is used to encode the EEG features at different time scales using a Transformer-based self-attention mechanism to obtain the encoded features.

[0017] A classification head is used to predict user identity based on the encoded features.

[0018] Furthermore, in the offline training of M different randomly seeded models, the loss function L used is... total for:

[0019] L total =L ce +L im

[0020]

[0021] Among them, L ce For cross-entropy loss, L im To maximize the information loss, N is the number of trials for the current batch of training samples. To determine the true label of trial i in category c, Let i be the predicted probability of class c for trial i. This represents the predicted probability distribution of category c in the current batch.

[0022] Furthermore, during offline training, an exponential learning rate decay strategy is used to optimize the training process, and after the training reaches a preset number of training rounds, adversarial examples are introduced to further train the model.

[0023] Furthermore, before training M different random seed models offline using the datasets under each dynamic window scale, the process further includes: performing data augmentation and bandpass filtering on each data sample in the dataset; wherein, the data augmentation includes performing trial synthesis, trial aliasing, time reversal, or adding Gaussian noise on the data samples;

[0024] In session two, the EEG data of length T received online are distributed across S different dynamic window scales [T1, T2, ..., T]. S Before performing sliding window processing, the method further includes: bandpass filtering of the EEG data of length T.

[0025] Furthermore, when each trial of the EEG data is divided into non-overlapping segments according to a length T, if the remaining data length after a certain trial is less than T, a data segment is extracted from the continuous data preceding the remaining data of that trial, and the extracted data segment is spliced ​​with the remaining data to form an EEG segment of length T.

[0026] According to a second aspect of the present invention, an online brainprint recognition system for cross-session scenarios is provided, comprising a computer-readable storage medium and a processor;

[0027] The computer-readable storage medium is used to store executable instructions;

[0028] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the online brainprint recognition method for cross-session scenarios as described in any of the first aspects.

[0029] According to a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the online brainprint recognition method for cross-session scenarios as described in any of the first aspects.

[0030] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when run on a computer, causes the computer to perform the online brainprint recognition method for cross-session scenarios as described in any one of the first aspects.

[0031] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0032] (1) To achieve online brainprint recognition across conversational scenarios, this invention uses non-overlapping slices of EEG signals collected offline in the first conversation as data samples of length T (the minimum trial length of all trials in the EEG data). Each data sample of length T is processed by sliding windowing at S different dynamic window scales, generating S sets of datasets with different dynamic window scales. M random seed models are trained for each dataset at each dynamic window scale. During online prediction, based on the M trained random seed models at each scale, a multi-scale dynamic window weighted ensemble strategy is adopted to fully utilize the information of the current online trial. By combining different dynamic window scales, effective features in the EEG signals can be captured more comprehensively, thereby improving the accuracy and robustness of the prediction results. Furthermore, in the second conversational scenario, only the EEG data packet of length T for the current trial needs to be received to quickly predict the user's identity, without waiting for all data packets of the trial to be fully transmitted before analysis. Thus, this invention can achieve fast, accurate, and robust identity recognition with the shortest possible data length T.

[0033] (2) Preferably, when weighting and summing the prediction results under different dynamic window scales, the weighting coefficients used can effectively represent the proportion of effective information contained in the prediction results under each dynamic window scale. By assigning greater weights to dynamic window scales with higher information content, the effect of multi-scale integration is further optimized, thereby improving the accuracy of the final prediction results.

[0034] (3) Preferably, the deep neural network architecture (brainprint recognition model) in this invention combines multi-scale temporal convolution to capture key information features at different time scales in EEG signals, and further enhances the modeling ability of temporally dependent information through the self-attention mechanism of the Transformer encoder, ultimately generating encoded features with enhanced expressive power. This network can effectively extract abstract features highly related to identity, and can further improve feature representation ability and recognition accuracy.

[0035] (4) Preferably, during offline training, cross-entropy loss L is used. ce To minimize empirical error, while introducing information to maximize loss L im Encouraging the model to generate diverse outputs improves its generalization ability and robustness on cross-session data.

[0036] (5) Furthermore, during offline training, an exponential learning rate decay strategy is adopted to optimize the training process, and after the training reaches the preset number of training rounds, adversarial examples are introduced for training to improve the robustness and generalization of the model in cross-session scenarios.

[0037] (6) As a preferred option, before offline training, a variety of data augmentation strategies, including trial synthesis, trial aliasing, time reversal, and Gaussian noise, are introduced into the training samples to enable the model to have a stronger generalization ability for data from different sessions.

[0038] (7) Further, when each trial in the EEG data is segmented without overlap according to length T, if the remaining data length after a certain trial is less than T, the necessary data segments are extracted from the continuous data before the remaining data of that trial and spliced ​​with the remaining data to form a complete EEG segment of length T, so as to make full use of the training data of Session 1 and ensure the consistency of data length.

[0039] In summary, this invention trains a highly adaptive brainprint recognition model using only data from a single session. During the offline training phase, data augmentation and adversarial training significantly improve the model's generalization ability and robustness. In the cross-session online prediction of EEG signal streams, a multi-scale dynamic window weighted ensemble strategy is adopted to achieve rapid, accurate, and robust identification of the subject's identity based on online EEG signals. Attached Figure Description

[0040] Figure 1 This is a scene diagram of online brainprint recognition in a cross-session scenario according to an embodiment of the present invention.

[0041] Figure 2 This is a diagram of the neural network architecture for brain pattern recognition based on multi-scale convolution and Transformer coding layers in an embodiment of the present invention.

[0042] Figure 3 This is a framework diagram of the multi-scale dynamic weighted integrated prediction strategy for online prediction in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0044] Example 1

[0045] like Figures 1-3 As shown, this embodiment of the invention provides an online brainprint recognition method for cross-conversation scenarios, including:

[0046] (1) Dataset Construction: For K (K≥2) users in the same batch, collect their EEG data during a single conversation. Process the collected EEG data, including:

[0047] Each trial in the EEG data is segmented non-overlappingly according to a length T, resulting in an initial dataset consisting of data samples of length T after segmentation. Each data sample in the initial dataset is labeled with the user's ID. If the remaining data length after segmentation for a certain trial is less than T, a necessary data segment is extracted from the preceding continuous data and concatenated with the remaining data to form a complete EEG segment of length T. This fully utilizes the training data from Session 1 and ensures data length consistency; where T is the minimum trial length among all trials in the EEG data.

[0048] Each data sample of length T is processed across S different dynamic window scales [T1, T2, ..., T]. S ](T j Sliding window processing is performed under the condition ≤T,j=1,2,…,S) to generate S sets of datasets under different dynamic window scales; and each set of datasets is divided into training set and validation set.

[0049] After augmenting the training set, it is combined with the original training set to form a new training set. Then, bandpass filtering is applied to all data in both the training and validation sets to remove noise, resulting in the updated dataset.

[0050] (2) Offline training: The brainprint recognition model is trained offline based on the updated dataset.

[0051] Specifically, for each set of datasets at the dynamic window scale (the dataset corresponding to each time window length), different random seeds are used for offline training to build M brainprint recognition models, with the model architecture being exactly the same for all scales.

[0052] As a preferred implementation, each brainprint recognition model includes a convolutional module, an encoding module, and a classification head. The convolutional module performs convolutions on each data sample at different scales under the current dynamic window scale to obtain EEG features at different scales. The encoding module uses a Transformer-based self-attention mechanism to encode EEG features at different scales, captures the long and short temporal dependencies of EEG features for global modeling, and finally generates encoded features with enhanced expressive power. The classification head includes multiple fully connected layers, and the encoded features are sequentially passed through multiple fully connected layers to belong to the predicted user identity information.

[0053] As a preferred implementation, an exponentially decaying learning rate strategy is used to train the brainprint recognition model. After a certain number of training epochs, adversarial training (AWP) is introduced to further improve the model's generalization ability. The loss function combines cross-entropy loss and information maximization loss, ensuring both the minimization of empirical error and the diversification of model output. Simultaneously, an early stopping strategy on the validation set is applied to avoid overfitting.

[0054] (3) Online Prediction: For the aforementioned K users, in Session 2, when the length of the received EEG data packets accumulates to T, user identity prediction is performed using the data packets of the current length T. Specifically, the received EEG data packets of length T are bandpass filtered and then copied S times. The S copies of data correspond to S different dynamic window scales [T1, T2, ..., T]. S ](T jSliding window processing is performed under the condition ≤T,j=1,2,…,S) to obtain S groups of online data under different dynamic window scales; the online data under the same dynamic window scale are input into M brainprint recognition models under the corresponding dynamic window scale for user identity prediction, and the M prediction results are averaged and integrated as the prediction result under the current dynamic window scale; then the prediction results under the S groups of dynamic window scales are weighted and summed, and the output of the neural network is normalized by softmax to obtain the final user identity prediction result, realizing online brainprint recognition across conversation scenarios.

[0055] As a preferred implementation, when weighting and summing the prediction results under the S groups of dynamic window scales, the weighting coefficients used are the proportion of effective information in the prediction results under that group of dynamic window scales:

[0056]

[0057] In the formula, ω j T represents the weighting coefficient of the prediction results under the j-th dynamic window scale. j Let L be the dynamic window scale of the j-th group, and L be the unit online transmission length of the EEG data packet in Session 2.

[0058] The above method will be further illustrated with specific examples below:

[0059] In this embodiment of the invention, for the same batch of K users, EEG data from different conversations (conversation one and conversation two) were collected over two days, such as... Figure 1 As shown, in Session 1 and Session 2, different colors represent EEG data from different users. Data from Session 1 was used for offline model training, while data from Session 2 was collected online for testing. The length of the data collected from each subject (user) was randomized, but the following constraints were imposed on the data length in both sessions: the minimum acquisition length for all trials was the same, denoted as T, where T represents the signal length containing sufficient information, avoiding data that was too short or too long. For the online data stream (EEG data collected in Session 2), the data was transmitted in the form of data packets of length L (input to the offline-trained model).

[0060] (a) Data sliding window:

[0061] First, for all trials of EEG data from K users collected during the session, each trial is segmented without overlap using a time window of length T. If the remaining data after segmentation for a certain trial is less than T, a necessary data segment is extracted from the continuous data preceding the remaining data of that trial and concatenated with the remaining data to form a complete EEG segment of length T. The segmented trial data is denoted as... Where k represents the k-th user, and k takes values ​​from 1, 2, ..., K, and i represents the i-th trial data for that user. The data for each trial has a shape of C×T, where C is the number of channels in the EEG signal, and T is the length of the trial data after segmentation. For each trial data... Its label y i It is set to the corresponding user's ID k.

[0062] For the trial length T, select three suitable dynamic window lengths [T1, T2, T3](T j ≤T,j=1,2,3), usually the larger common factor or integer of T is taken as T. j Preferably, it also ensures T j The dynamic window length should be an integer multiple of the length L of a single data packet to fully utilize the information in each packet. For example, if T = 3s and L = 0.2s, the dynamic window length can be 1s, 1.6s, 2s, etc. Next, the data from Session 1 is copied into three copies and randomly divided into training and validation sets in an 8:2 ratio. Then, based on the dynamic window length T... j Sliding window processing is applied to the training and validation sets to generate three sets of training data with different sliding window step sizes. and validation set data

[0063] (II) Data Augmentation and Filtering:

[0064] For the training set data of each dynamic window The data were randomly divided into four groups, and four data augmentation techniques were applied to each group: trial synthesis, trial aliasing, time reversal, and Gaussian noise. The definitions of these techniques are as follows:

[0065] Trial synthesis: EEG data from two trials (trial a and trial b) from the same user and If a random coefficient p is randomly generated within the range of [0.3, 0.7], then the synthesized new sample EEG data X new This can be expressed as follows:

[0066] X new =p*X a +(1-p)*X b

[0067] Trial aliasing: two trial EEG data from the same user and By swapping the first and second halves of its time dimension to form a new trial, the formula is as follows:

[0068]

[0069] Among them, X a (:,1:T j / / 2) indicates X a The sample is along the first half of the time dimension, X a (:,T j / / 2:T j ) represents X a The sample along the latter half of the time dimension, X b That is also true.

[0070] Time Reversal: X-ray of EEG Data a =[x a (:,1),x a (:,2),...,x a (:,T′-1),x a [:,T′)], where x a (:,T′) represents the signal amplitude vector of all channels of the EEG data at time T′. The time-reversed data augmentation is defined as follows:

[0071] X ′ a =[x a (:,T′),x a (:,T′-1),...,x a (:,2),x a (:,1),]

[0072] Gaussian noise: for EEG data X a The data after Gaussian noise enhancement is as follows:

[0073] X ′ a =X a +rand*std(X a ) / C noise

[0074] Where rand is a random number that follows a Gaussian distribution, std(X) a The original signal (EEG data X) a The standard deviation of C noise This is a hyperparameter, typically set to 2.

[0075] The enhanced data is mixed with the original training data to form the training set. The training set data and the validation set data are then subjected to a 4-40Hz bandpass filter to remove low-frequency and high-frequency noise from the EEG signals, resulting in the updated dataset.

[0076] (III) Offline Model Training

[0077] For three time windows of different scales, M random seed models are trained for each time window to obtain the corresponding brainprint recognition model. The overall framework diagram of the model is as follows. Figure 2 As shown. During model training, an exponential learning rate decay strategy is adopted. The initial learning rate lr0 is set to 1e-3, then the decayed learning rate lr... t It can be represented as:

[0078] lr t =lr0*γ t

[0079] Where t is the current iteration number, and γ is the decay factor (0 < γ < 1), which controls the decay rate of the learning rate.

[0080] After a certain number of training rounds, adversarial examples (AWPs) are introduced for adversarial training. The core idea is to apply small adversarial perturbations to the weights during model training, allowing the model to learn in a more challenging environment, thereby improving its generalization ability. In this embodiment, the Adam optimizer is used for parameter optimization, and the loss function is the cross-entropy loss L. ce To minimize empirical error, while introducing information to maximize loss L im Encourage models to generate diverse outputs. At the same time, use a validation set early stopping strategy to avoid model overfitting.

[0081] In this embodiment of the invention, each loss term and the total loss function L total The formula is as follows:

[0082]

[0083] L total =L ce +L im

[0084] In the formula, N represents the number of trials in the current batch, and K represents the number of users; This represents the true label of trial i in category c. This represents the predicted probability of trial i in category c; This represents the predicted probability distribution of category c in the current batch.

[0085] (iv) Online prediction:

[0086] For online test data streams (EEG data packets acquired online during Session 2), Information Transfer Rate (ITR) is a comprehensive indicator measuring the speed and accuracy of an algorithm. Its calculation formula is:

[0087]

[0088] Where T is the average trial duration, which is the data length T after segmenting the EEG data as mentioned above. K is the number of classification categories, which is the number of subjects mentioned above. P is the recognition accuracy, and the unit of ITR is bits / min.

[0089] To improve the Information Transmission Rate (ITR) during online prediction, if the current trial is too long, testing can be performed using only the currently received data of length T, ensuring that the model can make predictions quickly with limited data. When the length of the received EEG data packets reaches T, the received EEG data of length T is processed by a 4-40Hz bandpass filter to maintain consistency with the training data. Then, the bandpass-filtered EEG data of length T is copied three times, and the current trial is divided into three groups of data with different dynamic window scales using a sliding window. These groups are then input into M pre-trained random seed models corresponding to different dynamic window scales for prediction.

[0090] The prediction results of the trained random seed models are averaged and integrated. This involves summing the probability vectors of the M trained random seed models for the current trial and taking the average, which is then used as the prediction result for the current dynamic window scale. For the prediction results of models at different dynamic window scales, the weighting coefficients are used to weight the prediction results of the corresponding dynamic window scales. After a nonlinear mapping, this summation is used as the final prediction result p. c To achieve cross-session online brainprint recognition, such as Figure 3 As shown. Wherein, the weighting coefficient ω j The calculation formula is as follows:

[0091]

[0092] The final predicted probability p c The calculation formula is:

[0093]

[0094] Among them, f i For the current dynamic window size T j The i-th trained random seed model has a probability p c After softmax standardization, the index corresponding to the maximum predicted probability is taken as the predicted user ID.

[0095] This invention introduces adversarial training and various data augmentation strategies, which significantly improves the model's generalization ability and robustness across session data. By combining multi-scale convolution and Transformer encoder, key identity-related features are extracted from the data. For online data streams, a multi-scale dynamic window weighted integration strategy is adopted to ensure fast and accurate identity recognition with the shortest possible data length, providing an efficient and practical online brainprint recognition solution.

[0096] Example 2

[0097] This invention provides an online brainprint recognition system for cross-session scenarios, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method in Embodiment 1 above.

[0098] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor performs various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0099] The relevant technical solutions are the same as above, and will not be repeated here.

[0100] Example 3

[0101] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method in Embodiment 1 above.

[0102] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0103] The relevant technical solutions are the same as above, and will not be repeated here.

[0104] Example 4

[0105] This application provides a computer program product, including a computer program that, when run on a computer, causes the computer to perform the steps of the method in Embodiment 1 above.

[0106] The relevant technical solutions are the same as above, and will not be repeated here.

[0107] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An online brainprint recognition method for cross-conversation scenarios, characterized in that, include: Collect EEG data from K users during a single conversation, and categorize each trial in the EEG data according to its length. Perform non-overlapping segmentation of brain electronic fragments, and then divide the segmented brain electronic fragments into individual segments. Different dynamic window scales Perform sliding window processing below, and generate the corresponding window. Data sets with different dynamic window scales; where K≥2, The minimum trial length for all trials in the EEG data; ; Using datasets at each dynamic window scale, M different random seed models are trained offline to obtain M brainprint recognition models. Each brainprint recognition model includes: a convolution module, used to perform multi-scale convolution operations on each data sample at the current dynamic window scale to extract EEG features at different time scales; an encoding module, used to encode the EEG features at different time scales using a Transformer-based self-attention mechanism to obtain encoded features; and a classification head, used to predict user identity based on the encoded features. In session two, the length received online is... The EEG data were respectively in the Different dynamic window scales The process involves sliding window processing, inputting EEG data at different dynamic window scales into M brainprint recognition models at the corresponding dynamic window scale for user identity prediction, and averaging the M prediction results at the same dynamic window scale as the prediction result for the current dynamic window scale. After weighted summation of prediction results under different dynamic window scales, and then normalization, the user identity prediction result is obtained, realizing online brainprint recognition across sessions.

2. The online brainprint recognition method for cross-session scenarios according to claim 1, characterized in that, The When summing the prediction results under different dynamic window scales using weighted methods, the weighting coefficients used are: In the formula, For the first A dynamic window scale The weighting coefficients of the prediction results below, The unit online transmission length of the EEG data packet in Session 2.

3. The online brainprint recognition method for cross-session scenarios according to claim 1, characterized in that, The loss function used in offline training of M different randomly seeded models for: in, For cross-entropy loss, To minimize information loss, This represents the number of trials for the current batch of training samples. To determine the true label of trial i in category c, Let i be the predicted probability of class c for trial i. This represents the predicted probability distribution of category c in the current batch.

4. The online brainprint recognition method for cross-session scenarios according to claim 3, characterized in that, During offline training, an exponential learning rate decay strategy is used to optimize the training process, and after the training reaches the preset number of training rounds, adversarial examples are introduced to further train the model.

5. The online brainprint recognition method for cross-session scenarios according to claim 1, characterized in that, Before training M different random seed models offline using datasets at each dynamic window scale, the process further includes: performing data augmentation and bandpass filtering on each data sample in the dataset; wherein, the data augmentation includes performing trial synthesis, trial aliasing, time reversal, or adding Gaussian noise on the data samples; In session two, the length received online is... The EEG data were respectively in the Different dynamic window scales Before performing the sliding window process, the method further includes: processing the length of the window. Bandpass filtering was performed on the EEG data.

6. The online brainprint recognition method for cross-session scenarios according to claim 1, characterized in that, Each trial in the EEG data is arranged according to length When performing non-overlapping segmentation, if the remaining data length after a certain trial segmentation is insufficient... Then, a data segment is extracted from the continuous data preceding the remaining data in that trial, and the extracted data segment is concatenated with the remaining data to form a data segment of length [length missing]. Brain electronic fragments.

7. An online brainprint recognition system for cross-conversation scenarios, characterized in that, Includes computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the online brainprint recognition method for cross-session scenarios as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the online brainprint recognition method for cross-session scenarios as described in any one of claims 1-6.

9. A computer program product, characterized in that, Includes a computer program that, when run on a computer, causes the computer to perform the online brainprint recognition method for cross-session scenarios as described in any one of claims 1-6.

Images