Task-independent brainprint recognition method based on feature decorrelation decoupling

Abstract

The application discloses a task-independent brainprint recognition method based on feature decoupling. The existing method lacks the mining of inherent identity information of the brain, which leads to poor robustness of brainprint recognition in a cross-task scene, and is difficult to popularize in practical application. The application first adopts two branch networks to respectively coarsely decompose identity information and task-related information in electroencephalogram (EEG); secondly, considering the influence of the task state on the identity information, a decoupling method is adopted to make the identity information and the task-related information as independent as possible; finally, the brainprint features in the EEG are fully utilized for classification in an adversarial self-supervised manner. The method has good performance, can realize efficient task-independent brainprint recognition, and is a brainprint recognition method which can be robustly used in real life.

Description

Technical Field

[0001] This invention belongs to the field of EEG signal recognition in the field of biometric recognition, specifically involving a task-independent brainprint recognition method based on feature decorrelation decoupling. It utilizes feature decorrelation decoupling to obtain identity information independent of task information, and fully leverages brainprint features for identity recognition through adversarial self-supervision. Background Technology

[0002] Biometrics relies on personal characteristics and plays a crucial role in authentication systems. While physical biometric technologies such as facial recognition and fingerprint recognition are widely used in real life, the potential dangers of elaborate forgery or covert duplication remain unavoidable. In addition to physical biometrics, electroencephalography (EEG) recordings of brain activity have been proposed as a new cognitive biometric feature, meeting basic identification requirements and termed "brainprints." Furthermore, only a living individual can provide signals of brain activity, and these signals are not under the user's control. This means that the user's identity information cannot be intentionally leaked or stolen, making EEG-based biometrics suitable for applications with high security requirements.

[0003] In recent years, based on the type of task stimuli received by subjects, brainprint recognition can be broadly categorized into four main types: resting-state (RP) based brainprint recognition, visual evoked potential (VEP) based brainprint recognition, motor imagery (MI) based brainprint recognition, and event-related potential (ERP) based brainprint recognition. However, these brainprint recognition methods still have some limitations. External stimulus-based brainprint recognition requires subjects to have no corresponding physiological defects and be able to receive external stimuli during the experiment. Furthermore, these types of brainprint recognition are all task-specific, making them difficult to generalize and apply in real-world scenarios. Compared to existing methods, this paper proposes a task-independent brainprint recognition method based on feature decorrelation decoupling. This method decouples identity and task information from EEG data and fully utilizes brainprint features to achieve highly robust cross-task brainprint recognition. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a task-independent brainwave recognition method based on feature-based decorrelation and decoupling. This method primarily involves decoupling identity and task information from EEG data to extract independent identity information, and then fully utilizing this identity information for identity recognition through adversarial self-supervision.

[0005] In a first aspect, the present invention provides a task-independent brainwave recognition method based on feature decorrelation decoupling, specifically including the following steps:

[0006] Step 1: Preprocess the raw EEG data and construct the dataset;

[0007] Step 2: Construct an EEG feature extraction network to extract multi-scale spatiotemporal features of brain patterns;

[0008] 2-1 Each EEG sample was divided into low-frequency, high-frequency, and full-frequency sub-samples according to three frequency bands: low frequency, high frequency, and full frequency.

[0009] 2-2 The low-frequency, high-frequency, and full-frequency time-frequency features are extracted by passing the low-frequency, high-frequency, and full-frequency EEG patterns through two-layer one-dimensional time-domain convolution and one-layer one-dimensional frequency-domain convolution with three different kernel sizes.

[0010] 2-3 Through the extraction of the above-mentioned temporal and frequency domain features of brain patterns, each EEG sample yields... Brain stria time-frequency features, where n is the number of hidden layers and c represents the number of EEG channels;

[0011] 2-4 The brain pattern time-frequency features are concatenated along the frequency domain dimension to obtain... Then, the time-frequency features f are processed by a (c×3) two-dimensional convolution kernel. ts Perform channel spatial convolution to output the temporal-frequency spatial features of brain patterns.

[0012] Step 3: Construct a primary decoupled neural network model of brain patterns and tasks;

[0013] The brainprint and task primary decoupling neural network model includes a parallel brainprint primary decoupling neural network and a task primary decoupling neural network.

[0014] The primary decoupling neural network for brain patterns includes an EEG feature extraction network G. f Identity detector C f The EEG feature extraction network G f For identity information extraction, an EEG feature extraction network constructed in step 2 is used; the identity discriminator C f For identity recognition, it includes a parallel main classifier C and an auxiliary classifier Cs; both the main classifier C and the auxiliary classifier Cs use fully connected layers;

[0015] The EEG feature extraction network G f The loss function is as follows:

[0016]

[0017] in Let n represent the loss function and m represent the number of samples. Indicates input data x i The probability that a subject m belongs to the subject m;

[0018] The primary decoupling neural network for the task includes an EEG feature extraction network G. t Task classifier C t The EEG feature extraction network Gt For task information extraction, an EEG feature extraction network constructed in step 2 is used; the task classifier C t A fully connected layer is used for task discrimination.

[0019] Step 4: Train the brainprint and task-based primary decoupling neural network model;

[0020] 4-1 Constructing the EEG Feature Extraction Network G f and G t Constraints: Identity information S = G f (x) and task information A = G t (x) Discorrelation; specifically:

[0021] (1) The EEG feature extraction network G f and G t The original identity feature matrix S and the original task feature matrix A are concatenated to obtain a new matrix. m S m is the number of samples. Q The feature dimension is used; the matrix Q is mapped to a high-dimensional reproducing kernel Hilbert space (RKHS), and the Hilbert-Smitter independence criterion (HSIC) is used to determine the independence of any two vectors q in matrix Q. i ,q j Dependencies, where i,j∈m Q And calculate the Frobenius norm of the cross-covariance operator in RKHS;

[0022] Kernel functions obtain independent representations by mapping the original data to RKHS, as follows:

[0023]

[0024] in It is a measurable symmetric positive definite kernel function. For mapping functions, It is Hilbert space, α i It is an eigenvalue;

[0025] (2) After obtaining the kernel function, the vector q is detected by the Hilbert-Smitter Independence Criterion (HSIC). i ,q j Independence;

[0026] For random variable q i ,q j and kernel function The definition of HSIC is:

[0027]

[0028] in It is in the regenerating kernel Hilbert space with respect to the kernel function as and The covariance operator, ||·|| F It is a Frobenius norm and exists. q i Independent of q j ;

[0029] (3) For kernel functions Optimize

[0030] Because the kernel function is calculated in Hilbert space The complexity is relatively high, so the stochastic Fourier feature method is used to approximate the kernel function. Fourier transform sampling is used to obtain a dimensionality-reduced function that approximates the original kernel function, capturing two vectors q. i ,q j The nonlinear correlation between them; specifically:

[0031] RFF uses the following formula (5) to convert vector q i ,q j Mapping to a low-dimensional Euclidean space, the inner product after mapping is the estimate of the kernel function; using RFF features to achieve linear calculations removes nonlinear correlations, thereby achieving statistical independence of the features;

[0032] Random Fourier function space Represented as:

[0033]

[0034] Where ω is sampled from the standard normal distribution N(0,1), φ is sampled from the uniform distribution U(0,2π); x represents the vector q. i ,q j According to formula (5), matrix Q is converted into RFF features to approximate the kernel functions of identity features and task features.

[0035] (4) Independence test

[0036] Assuming the existence of measurable spaces Ω1 and Ω2, and In the regenerated kernel Hilbert space of Ω1 and Ω2, correspondingly, and It is also measurable; therefore, space arrive There exists a unique cross-covariance operator Σ XY , so that:

[0037]

[0038] in Cov(·) is the covariance matrix;

[0039] From formula (6), we can see that Σ XY The calculation is extended to the calculation of the covariance matrix in Euclidean space, where f(X) and g(Y) represent nonlinear kernel functions; since If Σ XY If the Hilbert-Schmidt norm is zero, then X and Y can be considered independent; because kernel methods are difficult to compute, RFF can provide a function space. To achieve this goal, the cross-covariance matrix Σ XY It can be represented as:

[0040]

[0041] in u and v are elements of a random Fourier space;

[0042] Theoretically, detecting two vectors q i ,q j The independence between them will make vector q i ,q j As X respectively i Y i Substituting into formulas (6)-(7), it is necessary to determine the relationship between u(q) and u(q). i ) and v(q j The cross-covariance operator Σ ST Whether it tends to 0, the elements u and v in the random Fourier space are represented as:

[0043]

[0044]

[0045] in Indicates from The number of sampling functions;

[0046] We construct a cross-covariance matrix and minimize the Frobenius norm of the cross-covariance matrix to achieve the goal of decoherence. The loss function is defined as follows:

[0047]

[0048] The hyperparameter λ represents the sigmoid ramp-up function, which increments from 0 based on the number of training epochs, and takes the form:

[0049]

[0050] Where t∈[0,epoch];

[0051] 4-2 Utilizing adversarial self-supervised networks to fully mine the identity information output by the primary decoupled neural network of brainprint;

[0052] 4-2-1 EEG Feature Extraction Network G f The output identity feature S is input into the adversarial self-supervised network H to obtain a mask representation; each dimension of the mask representation is treated as a discrete random variable, and an approximate K-hot vector is obtained by sampling each dimension; β is the approximate sampling of the κ-hot vector through the Gumbel-Softmax technique, defined as follows:

[0053]

[0054] Where κ∈(0,1), N is the dimension obtained through the adversarial self-supervised network. The mask closest to 1 in the sampled κN is regarded as the important feature, and the others are regarded as the secondary features.

[0055] Gumbel-Softmax definition: For a predefined τ>0, i∈{1,...,N}, p∈{1,...,κ}, we have:

[0056]

[0057] in It is a probability vector. It is a sample that follows a Gumbel distribution, and π i ≥0, i∈1,…,N,∑ i π i =1;

[0058] 4-2-2 Using the EEG Feature Extraction Network G f The output identity feature S is multiplied by the mask β and 1-β to obtain the important dimension features and the secondary dimension features, and then the main classifier C and the auxiliary classifier C are used. S Train an adversarial self-supervised network H to fully utilize discriminative identity features;

[0059] Therefore, formula (2) Redefining

[0060] Loss function of main classifier C:

[0061]

[0062] Auxiliary classifier C S Loss function:

[0063]

[0064] in Dot product;

[0065] First, the optimized feature extraction network G is trained by minimizing the aforementioned loss. f G t Main classifier C, Auxiliary classifier C S Secondly, by minimizing and maximizing The adversarial self-supervised network H is trained and optimized, i.e.:

[0066]

[0067]

[0068] Step 5: Verify and test the trained brainprint and task-based primary decoupled neural network model;

[0069] Step 6: Using the trained, verified, and tested brainprint and task-based primary decoupled neural network model, realize brainprint recognition of EEG.

[0070] As a preferred option, step 1 specifically involves:

[0071] 1-1 The EEG data was downsampled to 250Hz, and the raw EEG data was filtered from 0 to 75Hz using a Butterworth filter;

[0072] 1-2 Perform a short-time Fourier transform on the EEG data x processed in step 1-1 to extract time-frequency features;

[0073] 1-3 The time-frequency features obtained in step 1-2 are extracted using time windows, and then labeled with the subject's label and the corresponding task label.

[0074] 1-4 The EEG sample data obtained after processing in steps 1-3 will be used as a test set for one task {X}. t ,Y t ,Y tl The remaining samples are divided into training sets proportionally. and the validation set {X v ,Y v ,Y vl}, where X, Y, Y l K represents the sample, identity label, task label, and number of tasks, respectively.

[0075] As a preferred option, steps 1-2 are specifically as follows:

[0076] Using a time-finite window function h(t), assuming the non-stationary signal x is stationary within a time window, by shifting the window function h(t) along the time axis, a set of local "spectrums" of the signal x is obtained through segment-by-segment analysis. The short-time Fourier transform of the signal x(τ) is defined as:

[0077]

[0078] Where STFT(t,f) represents the short-time Fourier transform of signal x(τ) at time t and frequency f, and h(τ-t) is the window function.

[0079] Preferably, the low frequency in step 2-1 is 1-12Hz, and the high frequency is 12-30Hz.

[0080] Preferably, the kernel sizes of the two one-dimensional frequency domain convolution layers in steps 2-3 are [(6×1),(7×1)],[(7×1),(11×1)],[(15×1),(15×1)], respectively, and the kernel sizes of the one-dimensional frequency domain convolution layer are (12×1),(18×1),(30×1) respectively.

[0081] As a preferred option, gradient backpropagation is used to optimize the loss function (16)-(17).

[0082] In a second aspect, the present invention provides a task-independent brainwave recognition device, including...

[0083] The data acquisition module is used to collect EEG data;

[0084] The data preprocessing module preprocesses the EEG data;

[0085] The brainprint recognition module uses a well-trained, verified, and tested brainprint and task-based primary decoupled neural network model.

[0086] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0087] Fourthly, the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the above-described method.

[0088] The beneficial effects of this invention are as follows: Firstly, for EEG data under different task stimuli, this invention coarsely decomposes the identity information and task-related information in the EEG. Further considering the non-stationarity of EEG, it decouples brainwave features and task features through decorrelation. Finally, through self-supervised adversarial processing, it makes full use of all identity-related features to ensure the model's stability across time and tasks. Attached Figure Description

[0089] Figure 1 This is a flowchart of the brainprint recognition model proposed in this invention;

[0090] Figure 2 This is a diagram of the feature-decorrelated and decoupled task-independent brain texture recognition model architecture proposed in this invention; Detailed Implementation

[0091] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description is provided in conjunction with the technical solutions and accompanying drawings of this invention:

[0092] This invention relates to a task-independent brainwave recognition method based on feature decorrelation decoupling, the flowchart of which is shown below. Figure 1 As shown, the original EEG is first preprocessed using Fast Fourier Transform (FFT). Then, two branch networks are used to coarsely decompose the identity information and task-related information in the EEG. Subsequently, considering the influence of task state on identity information, a decorrelation method is employed to make the identity information and task-related information as independent as possible. Finally, identity-related brainprint features in the EEG are fully utilized for classification through adversarial self-supervision.

[0093] Step 1: Preprocess raw EEG data

[0094] 1) The noise frequency contained in the raw EEG signal is usually below 0.5Hz or above 50Hz. In order to remove the power frequency interference caused by the EEG acquisition equipment and the electromyography interference of the subject, the EEG data is downsampled to 250Hz, and the raw EEG data is filtered from 0 to 75Hz using a Butterworth filter.

[0095] 2) Perform a short-time Fourier transform on the signal x output from operation 1) to extract time-frequency features. A time-finite window function h(t) is used. Assuming the non-stationary signal x is stationary within a time window, the signal x is analyzed segment by segment by shifting the window function h(t) along the time axis to obtain a set of local "spectrums". The specific window size in this scheme is 0.5s. The short-time Fourier transform of signal x(τ) is defined as:

[0096]

[0097] Let h(τ-t) represent the short-time Fourier transform of signal x(τ) at time t and frequency f, where h(τ-t) is a window function.

[0098] 3) Use a 15-second time window to capture the EEG data obtained in 2), and label the corresponding EEG sample data with the subject's label and the corresponding task label;

[0099] 4) Use the EEG sample data obtained after processing in step 3) as a test set for one task {X}.t ,Y t ,Y tl The remaining samples are divided into training sets proportionally. and the validation set {X v ,Y v ,Y vl}, where X, Y, Y l K represents the sample, identity label, task label, and number of tasks, respectively. EEG sample Where c represents the number of EEG channels, s represents the frequency domain dimension, and t represents the time domain dimension. Specifically, this scheme selects nine channels: Fz, F7, F8, C3, C4, P7, P8, O1, and O2, ranging from 1 to 30 Hz, with a sampling rate of 250 Hz, i.e., c = 9, s = 30, and t = 30.

[0100] Step 2: Construct a neural network model for extracting multi-scale time-frequency brain texture features;

[0101] 1) Each EEG sample was divided into three sub-samples according to three frequency bands: 1-12Hz, 12-30Hz, and the full frequency band. One-dimensional temporal convolution kernels of size (1×21), (1×5), and (1×11) were used to extract the temporal features of brain patterns.

[0102] 2) Then, using two layers of one-dimensional frequency domain convolution kernels [(6×1),(7×1)],[(7×1),(11×1)],[(15×1),(15×1)] respectively, brain pattern frequency domain features of low frequency, high frequency and full frequency EEG are extracted respectively;

[0103] 3) Through the above time-domain and frequency-domain feature extraction, each sample obtains Time-frequency features, where n is the number of hidden layers;

[0104] 4) The brain pattern time-frequency features are spliced ​​together according to the frequency domain dimension to obtain... Then, the time-frequency features f are processed by a (9×3) two-dimensional convolution kernel. ts Perform channel spatial convolution to output the temporal-frequency spatial features of brain patterns.

[0105] Step 3: Construct a primary decoupled neural network model of brain patterns and tasks;

[0106] 1) Construct an EEG feature extraction network G as described in step 2. f It is used for identity information extraction, and an identity discriminator C is constructed using a fully connected layer. f Used for identity recognition; Identity discriminator C f Includes the main classifier C and the auxiliary classifier Cs

[0107]

[0108] in Let n represent the loss function and m represent the number of samples. Indicates input data x i The probability that a subject m belongs to the subject m;

[0109] 2) Construct an EEG feature extraction network G as described in step 2. t It is used for task information extraction, and a task classifier C is built using a fully connected layer. t Used for task identification;

[0110] Step 4: Construct the EEG feature extraction network G during training. f and G t Constraints: Identity information S = G f (x) and task information A = G t (x) Discorrelate;

[0111] 4) The EEG feature extraction network G f and G t The original identity feature matrix S and the original task feature matrix A are concatenated to obtain a new matrix Q. Then, matrix Q is mapped to a high-dimensional reproducing kernel Hilbert space (RKHS). The Hilbert-Smitter independence criterion (HSIC) is used to determine the independence of any two vectors q in matrix Q. i ,q j The dependencies are determined, and the Frobenius norm of the cross-covariance operator in RKHS is calculated. The kernel function obtains independent representations by mapping the original data to RKHS, as follows:

[0112]

[0113] in It is a measurable symmetric positive definite kernel function. For mapping functions, It is Hilbert space, α i These are eigenvalues. Function The original data is mapped to a high-dimensional space, and the inner product of the mapping function in the high-dimensional space yields the kernel function.

[0114] 5) After obtaining the kernel function, the HSCI detection vector q is used. i ,q j Independence. For random variable q i ,q j and kernel function The definition of HSIC is:

[0115]

[0116] in It is in the regenerating kernel Hilbert space with respect to the kernel function as and The covariance operator, ||·|| F It is a Frobenius norm. And it exists. q i Independent of q j .

[0117] 6) For kernel functions Optimize

[0118] Because the kernel function is calculated in Hilbert space Therefore, this invention proposes to approximate the kernel function using the stochastic Fourier feature method. It utilizes Fourier transform sampling to obtain a dimensionality-reduced function that approximates the original kernel function, capturing two vectors q. i ,q j The nonlinear correlation between them. Specifically:

[0119] RFF uses the following formula (5) to convert vector q i ,q j Mapping to a low-dimensional Euclidean space, the inner product after mapping is an estimate of the kernel function. Using RFF features to perform linear computation removes nonlinear correlations, thus achieving statistical independence of the features.

[0120] Random Fourier function space Represented as:

[0121]

[0122] Where ω is sampled from a standard normal distribution, and φ is sampled from a uniform distribution.

[0123] According to formula (5), matrix Q is converted into RFF features to approximate the kernel functions of identity features and task features.

[0124] 4) Independence testing

[0125] Assuming the existence of measurable spaces Ω1 and Ω2, and In the regenerated kernel Hilbert space of Ω1 and Ω2, correspondingly, and It is also measurable. Therefore, space... arrive There exists a unique cross-covariance operator Σ XY , so that:

[0126]

[0127] in Cov(·) is the covariance matrix.

[0128] From formula (6), we can see that Σ XY The calculation is extended to the calculation of the covariance matrix in Euclidean space, where f(X) and g(Y) represent nonlinear kernel functions. Since If Σ XY If the Hilbert-Schmidt norm is zero, then X and Y can be considered independent. Since kernel methods are computationally difficult, RFF can provide a function space. To achieve this goal, the cross-covariance matrix Σ XY It can be represented as:

[0129]

[0130] in

[0131] Theoretically, detecting two vectors q i ,q j The independence between them (represented by X and Y above) requires determining the relationship between u(q). i ) and v(q j The cross-covariance operator Σ ST Whether it tends to 0, u and v are elements of a random Fourier space, represented as:

[0132]

[0133]

[0134] in Indicates from The number of sampling functions;

[0135] We construct a cross-covariance matrix and minimize the Frobenius norm of the cross-covariance matrix to achieve the goal of decoherence. The loss function is defined as follows:

[0136]

[0137] The hyperparameter λ represents the sigmoid ramp-up, a function that increments from 0 according to the epoch, and has the following form:

[0138]

[0139] Step 5: Construct an adversarial self-supervision module to make full use of identity information;

[0140] 1)G fThe output identity feature S is input into the self-supervised network H to obtain a mask representation. Each dimension of the representation is treated as a discrete random variable, and sampling is performed on each dimension to obtain an approximate K-hot vector. β is the approximate sampling of the κ-hot vector using the Gumbel-Softmax technique, defined as follows:

[0141]

[0142] Where κ∈(0,1), N is the dimension obtained through the self-supervised network. The mask closest to 1 in the sampled κN is considered as an important feature, and the others are considered as secondary features.

[0143] Gumbel-Softmax definition: For a predefined τ>0, i∈{1,...,N}, p∈{1,...,κ}, we have:

[0144]

[0145] in It is a probability vector. It is a sample that follows a Gumbel distribution, and π i ≥0, i∈1,…,N,∑ i π i =1; set τ=0.8.

[0146] 2) G f The output identity feature S is multiplied by the mask β and 1-β to obtain the important dimension features and the secondary dimension features, and then a main classifier C and an auxiliary classifier C are used. S Formula (2) It can be redefined as

[0147] Main classifier loss function

[0148] Auxiliary classifier loss function

[0149] First, the optimized feature extraction network C is trained by minimizing the aforementioned loss. f G t Main classifier C, Auxiliary classifier C S Secondly, by minimizing and maximizing The adversarial self-supervised network H is trained and optimized, i.e.:

[0150]

[0151]

[0152] The proposed adversarial self-supervised module allows secondary dimensions to also play a role, because it is beneficial for optimizing the auxiliary classifier C. S Let the secondary dimension classify the labels to minimize the loss. Self-supervised networks learn β to select good dimensions to maximize loss. Secondary dimensions that contribute less can be identified. The classifier and the mask network are in conflict. By optimizing G... f and G t To minimize and The disadvantaged dimension is forced to carry more identity features and is independent of task features. Finally, by repeatedly removing low-level representations to make them new high-level representations, the learned representations tend to be cleaner identity features.

[0153] Step 6: Train the network model;

[0154] Using the training set obtained in step 1.4, the loss function is optimized by gradient backpropagation of the model constructed in steps 2 to 5. The best model is saved using the validation set obtained in step 1.4 for testing.

[0155] The SGD optimizer was used with a learning rate of 0.025 and a batch size of 64.

[0156] Step 7: Validate the effectiveness of this approach on a multi-task identity recognition dataset, which includes 30 participants (N=30). A comparative experiment was conducted with existing methods, and the results are shown in Table 1. The validation results demonstrate that the proposed model can effectively extract brainwave features under different cognitive tasks, is not limited by cognitive tasks, and exhibits strong robustness.

[0157] Table 1 shows the accuracy and equal error rate of the model on the multi-task identity recognition dataset.

[0158]

Claims

1. A task-independent brainprint recognition method based on feature decorrelation decoupling, characterized in that... The method includes the following steps: Step 1: Preprocess the raw EEG data and construct the dataset; Step 2: Construct an EEG feature extraction network to extract multi-scale spatiotemporal features of brain patterns; 2-1 Each EEG sample was divided into low-frequency, high-frequency, and full-frequency sub-samples according to three frequency bands: low frequency, high frequency, and full frequency. 2-2 The low-frequency, high-frequency, and full-frequency time-frequency features are extracted by passing the low-frequency, high-frequency, and full-frequency EEG patterns through two-layer one-dimensional time-domain convolution and one-layer one-dimensional frequency-domain convolution with three different kernel sizes. 2-3 Through the extraction of the above-mentioned temporal and frequency domain features of brain patterns, each EEG sample yields... Brain stria time-frequency features, where n is the number of hidden layers and c represents the number of EEG channels; 2-4 The brain pattern time-frequency features are concatenated along the frequency domain dimension to obtain... Then through Two-dimensional convolution kernels for time-frequency features Perform channel spatial convolution to output the temporal-frequency spatial features of brain patterns. ; Step 3: Construct a primary decoupled neural network model of brain patterns and tasks; The brainprint and task primary decoupling neural network model includes a parallel brainprint primary decoupling neural network and a task primary decoupling neural network. The primary decoupling neural network for brain patterns includes an EEG feature extraction network. Identity verification device The EEG feature extraction network For identity information extraction, an EEG feature extraction network constructed in step 2 is used; the identity discriminator For identity recognition, it includes a parallel main classifier C and an auxiliary classifier Cs; both the main classifier C and the auxiliary classifier Cs use fully connected layers; The EEG feature extraction network The loss function is as follows: Equation (2) in Let n represent the loss function and m represent the number of samples. Indicates input data The probability that a subject m belongs to the subject m; The primary decoupling neural network for the task includes an EEG feature extraction network. Task classifier The EEG feature extraction network For task information extraction, an EEG feature extraction network constructed in step 2 is used; the task classifier A fully connected layer is used for task discrimination. Step 4: Train the brainprint and task-based primary decoupling neural network model; 4-1 Constructing an EEG Feature Extraction Network and Constraints: Identity information S = With task information A= Remove the relevant information; specifically: 1) EEG feature extraction network and The original identity feature matrix S and the original task feature matrix A are concatenated to obtain a new matrix. , For the sample size, The feature dimension is used; the matrix Q is mapped to the high-dimensional reproducing kernel Hilbert space RKHS, and the Hilbert-Smitter independence criterion (HSIC) is used to determine the independence of any two vectors in matrix Q. The dependency relationships, among which And calculate the Frobenius norm of the cross-covariance operator in RKHS; Kernel functions obtain independent representations by mapping the original data to RKHS, as follows: Equation (3) in It is a measurable symmetric positive definite kernel function. For mapping functions, It is Hilbert space. It is an eigenvalue; 2) After obtaining the kernel function, the vector is detected using the Hilbert-Smitter Independence Criterion (HSIC). Independence; For random variables and kernel function 2. The definition of HSIC is: Equation (4) in It is in the regenerating kernel Hilbert space with respect to the kernel function as 1 and The cross-covariance operator of 2, It is a Frobenius norm and exists. , Independent of ; 3) For kernel functions Optimize Because the kernel function is calculated in Hilbert space Due to its high complexity, a stochastic Fourier feature method is used to approximate the kernel function. Fourier transform sampling is used to obtain a dimensionality-reduced function that approximates the original kernel function, capturing two vectors. The nonlinear correlation between them; specifically: RFF uses the following formula (5) to convert the vector Mapping to a low-dimensional Euclidean space, the inner product after mapping is the estimate of the kernel function; using RFF features to achieve linear calculations removes nonlinear correlations, thereby achieving statistical independence of the features; Random Fourier function space Represented as: Equation (5) in It is sampled from the standard normal distribution , It is sampled from a uniform distribution , Representative vector ; According to formula (5), matrix Q is converted into RFF features to approximate the kernel functions of identity features and task features. ; 4) Independence testing Assuming the existence of a measurable space , Indicates in The regenerating nucleus Hilbert space, correspondingly, It is also measurable; therefore, space arrive There exists a unique cross-covariance operator. , so that: Equation (6) in , , It is the covariance matrix; From formula (6), it can be seen that... The calculation is extended to the calculation of the covariance matrix in Euclidean space. because ,if The Hilbert-Schmidt norm of X and Y is zero, indicating that X and Y are independent. Because kernel methods are computationally difficult, RFF provides a function space. To achieve this goal, the cross-covariance matrix is ​​needed. Represented as: Equation (7) in u and v are elements of the random Fourier space; In theory, detecting two vectors The independence between them will make vectors As respectively , Substituting into formulas (6)-(7), it is necessary to determine the relationship between them. and cross-covariance operator Whether it tends to 0, the elements u and v in the random Fourier space are represented as: Equation (8) Equation (9) in , Indicates from The number of sampling functions; We construct a cross-covariance matrix and minimize the Frobenius norm of the cross-covariance matrix to achieve the goal of decoherence. The loss function is defined as follows: Equation (10) hyperparameters This represents the sigmoid ramp-up, a function that increments from 0 based on the number of training epochs, and has the following form: Equation (11) in ; 4-2 Utilizing adversarial self-supervised networks to fully mine the identity information output by the primary decoupled neural network of brainprint; 4-2-1 EEG Feature Extraction Network Output identity features Input adversarial self-supervised network Obtain the mask representation; treat each dimension of the mask representation as a discrete random variable, and sample each dimension to obtain an approximation. vector; Approximate sampling using the Gumbel-Softmax technique A vector is defined as follows: Equation (12) in N is the dimension obtained through an adversarial self-supervised network, sampled from the results. The mask closest to 1 is considered an important feature, and the others are considered minor features; Gumbel-Softmax definition: For a predefined... ,have: Equation (13) in It is a probability vector. It is a sample that follows a Gumbel distribution, and ; 4-2-2 Brainwave Feature Extraction Network Output identity features Multiplying by the mask β and 1-β yields the important and secondary dimension features, which are then used by the main classifier C and the auxiliary classifier. ; Therefore, formula (2) Redefining Loss function of main classifier C: Equation (14) Auxiliary classifier Loss function: Equation (15) in Dot product; First, the feature extraction network is trained and optimized by minimizing the aforementioned loss. , Main classifier Auxiliary classifier Secondly, by minimizing and maximizing Optimize through training ,Right now: Equation (16) Equation (17); Step 5: Verify and test the trained brainprint and task-based primary decoupled neural network model; Step 6: Using the trained, verified, and tested brainprint and task-based primary decoupled neural network model, realize brainprint recognition of EEG.

2. The method according to claim 1, characterized in that... Step 1 is as follows: 1-1 The EEG data was downsampled to 250Hz, and the raw EEG data was filtered from 0 to 75Hz using a Butterworth filter; 1-2 Perform a short-time Fourier transform on the EEG data x processed in step 1-1 to extract time-frequency features; 1-3 The time-frequency features obtained in step 1-2 are extracted using time windows, and then labeled with the subject's label and the corresponding task label. 1-4 The EEG sample data obtained after processing in steps 1-3 will be used as a test set for one task. The remaining samples are divided into training sets proportionally. and verification set ,in These represent the sample, identity label, task label, and number of tasks, respectively.

3. The method according to claim 2, characterized in that... Steps 1-2 are as follows: Using a time-finite window function h(t), assuming the non-stationary signal x is stationary within a time window, by shifting the window function h(t) along the time axis, a set of local "spectrums" of the signal x is obtained through segment-by-segment analysis. The short-time Fourier transform is defined as: Equation (1) in Indicates signal In time ,frequency Short-time Fourier transform on For window functions.

4. The method according to claim 1, characterized in that... The low frequency mentioned in step 2-1 is 1-12Hz, and the high frequency is 12-30Hz.

5. The method according to claim 1, characterized in that... The kernel sizes of the two one-dimensional frequency domain convolutions in steps 2-3 are respectively .

6. The method according to claim 1 or 5, characterized in that... The sizes of the one-dimensional frequency domain convolution kernels in steps 2-3 are respectively .

7. The method according to claim 1, characterized in that... The loss function (16)-(17) is optimized by gradient backpropagation.

8. A task-independent brainwave recognition device for implementing the method of any one of claims 1-7, characterized in that... include The data acquisition module is used to collect EEG data; The data preprocessing module preprocesses the EEG data; The brainprint recognition module uses a well-trained, verified, and tested brainprint and task-based primary decoupled neural network model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that... When the computer program is executed in a computer, it causes the computer to perform the method according to any one of claims 1-7.

10. A computing device, comprising a memory and a processor, characterized in that... The memory stores executable code, and when the processor executes the executable code, it implements the method according to any one of claims 1-7.

Images