Electroencephalogram multi-task analysis method and system based on dual-stream decoupling and prototype guidance

Abstract

This invention discloses a method and system for multi-task EEG analysis based on two-stream decoupling and prototype guidance, relating to the fields of brain-computer interface and biomedical signal processing technology. The method includes: standardizing initial EEG signal data, extracting prototype feature vectors, and converting them into high-dimensional semantic vectors to obtain embedded sequence tensors; constructing a two-stream collaborative backbone architecture based on task-shared backbone, individual-specific backbone, and feature decoupling mechanisms, and building independent task processing modules; training and optimizing the task processing modules based on a joint loss function to obtain trained task processing modules, and analyzing the embedded sequence tensors based on the trained task processing modules; calculating the individual fitness score of the initial EEG signals based on the prototype feature vectors, and modularly expanding the task processing modules based on the individual fitness scores. This invention alleviates the technical problems of catastrophic forgetting, structural rigidity, and cold-start difficulties in existing technologies.

Description

Technical Field

[0001] This invention relates to the fields of brain-computer interface and biomedical signal processing technology, and in particular to a method and system for EEG multitasking analysis based on dual-stream decoupling and prototype guidance. Background Technology

[0002] Electroencephalogram (EEG) signals, as non-invasive physiological signals reflecting the electrical activity of brain neurons, have extremely high application value in fields such as clinical diagnosis, brain-computer interfaces, cognitive function assessment, affective computing, and rehabilitation medicine. However, EEG signals themselves are highly non-stationary, exhibit significant individual variability, and are task-specific, making it extremely difficult to construct long-term stable EEG representation models that are cross-subject, cross-task, and robust.

[0003] In the field of EEG signal analysis, the mainstream technical approach is to build end-to-end classification models based on deep neural networks, such as Convolutional Neural Networks (CNNs) or Transformers. A typical workflow involves collecting EEG data from a large number of subjects, training a general feature extractor and classifier, and then directly applying it to new subjects; or employing a pre-training-fine-tuning paradigm, where when new subjects are introduced, their limited calibration data is used to update the model parameters using gradients to achieve domain adaptation. While these methods perform reasonably well in single experiments or on specific datasets, they suffer from the following significant drawbacks in real-world long-term, multi-task applications: (1) Catastrophic forgetting caused by the static fine-tuning paradigm: When dealing with new individuals or new tasks, existing technologies usually adopt a pre-training-fine-tuning paradigm, that is, updating the model parameters with full or partial gradients on the target domain data (Li et al., 2024). This global modification of the model parameters will destroy the feature distribution established by the model on old tasks or old individuals, causing the model to quickly forget old knowledge while learning new knowledge, i.e., catastrophic forgetting. In addition, existing methods have difficulty in achieving a balance between maintaining general knowledge and adapting to specific knowledge.

[0004] (2) Structural rigidity caused by fixed network topology: Existing neural network models have a fixed number of layers and neuron size before training, resulting in a fixed model capacity (BrainUICL, Zhou et al., 2025). EEG signals are highly non-stationary and exhibit individual variability. For simple tasks or typical individuals, a fixed large model may have parameter redundancy; while for complex tasks or abnormal individuals, a fixed model may lead to underfitting due to insufficient capacity.

[0005] (3) Cold-start difficulties caused by lack of physical priors: Mainstream models such as EEGNet (J. Neural Eng. et al., 2018) and DeepConvNet (Schirrmeister et al., 2017) usually treat EEG signals as pure digital matrices and rely entirely on data-driven methods to extract statistical features. This approach ignores the explicit neurophysiological physical properties of EEG signals, such as the energy entropy in the frequency domain, the nonlinear fractal dimension in the time domain, and the brain functional connectivity patterns in the spatial domain. The lack of physical prior guidance makes it impossible for the model to quickly evaluate and warm-start when faced with unlabeled or poorly labeled new individuals using general physical laws. It often requires a large amount of calibration data to converge, which greatly reduces the usability of the system. Summary of the Invention

[0006] To address the aforementioned technical problems in existing technologies, embodiments of the present invention provide a method and system for EEG multitasking analysis based on dual-stream decoupling and prototype guidance. The technical solution is as follows: On the one hand, a method for EEG multitasking analysis based on dual-stream decoupling and prototype guidance is provided. The method includes: standardizing initial EEG signal data to obtain a standard input tensor, and extracting the prototype feature vector of the standard input tensor; converting the standard input tensor into a high-dimensional semantic vector based on a deep convolutional neural network with shared weights to obtain an embedded sequence tensor; constructing a dual-stream collaborative backbone architecture based on a task-shared backbone, an individual-specific backbone, and a feature decoupling mechanism; constructing independent task processing modules based on task-level modular expansion, a multilayer perceptron-based task classifier, and similarity-based dynamic parameter hot-start of the dual-stream collaborative backbone architecture; training and optimizing the task processing modules based on a joint loss function to obtain trained task processing modules, and analyzing the embedded sequence tensor based on the trained task processing modules; calculating the individual fitness score of the initial EEG signal based on the prototype feature vector, and modularly expanding the task processing modules based on the individual fitness score.

[0007] Optionally, extracting the prototype feature vector of the standard input tensor includes: performing a Fourier transform on the standard input tensor to obtain multiple frequency band signals; calculating the differential entropy of each frequency band signal to obtain the frequency domain features of the standard input tensor; calculating the fractal dimension of the standard input tensor to obtain the time domain features of the standard input tensor; calculating the rational asymmetry index of the symmetrical electrode pairs of the left and right hemispheres in the standard input tensor to obtain the spatial domain features of the standard input tensor; and concatenating and normalizing the frequency domain features, the time domain features, and the spatial domain features to obtain the prototype feature vector of the standard input tensor.

[0008] Optionally, a deep convolutional neural network based on shared weights transforms the standard input tensor into a high-dimensional semantic vector to obtain an embedded sequence tensor, including: truncating the standard input tensor into segments based on local windows divided along the time axis to generate multiple sub-segments; performing feature encoding on each sub-segment using a deep one-dimensional convolutional neural network based on shared weights to obtain a feature map group; and performing global average pooling on the time dimension of the feature map group and rearranging it along the time step dimension to obtain the embedded sequence tensor.

[0009] Optionally, based on a task-sharing backbone, an individual-specific backbone, and a feature decoupling mechanism, a dual-stream collaborative backbone architecture is constructed, including: initializing two structurally identical but parameter-space-independent deep neural network modules, serving as the task-sharing backbone and the individual-specific backbone, respectively; wherein, the task-sharing backbone is used to learn the universal neurophysiological patterns across subjects in the current task and construct a task-sharing subspace; the individual-specific backbone is used to capture the inherent EEG feature shifts of the target subject and construct an individual-specific subspace; a feature independence measurement module is constructed at the output of the task-sharing backbone and the individual-specific backbone, and a Hilbert-Schmidt independence criterion constraint index is introduced into the feature space of the feature independence measurement module to construct a decoupling loss term; wherein, the decoupling loss term includes:

[0010] In the formula, For the decoupling loss term, These are the output feature matrices of the task-shared backbone and the individual-specific backbone, respectively, with HSIC representing the Hilbert-Schmidt independence criterion constraint index. The output splicing scheme of the dual-stream collaborative backbone architecture includes:

[0011] In the formula, The joint features output by the dual-stream collaborative backbone architecture are defined as follows: D is the embedding dimension, and N is the number of feature vectors in the embedding sequence tensor.

[0012] Optionally, based on task-level modular expansion, a multilayer perceptron-based task classifier, and similarity-based dynamic parameter hot-start of the dual-stream collaborative backbone architecture, an independent task processing module is constructed, including: determining the task domain to which the current input signal belongs based on the task routing module, and constructing an independent task processing module based on the task domain; determining whether the distance between the new task and all historical tasks exceeds a preset threshold; if so, training the task-sharing backbone of the dual-stream collaborative backbone architecture based on a contrastive predictive coding framework; if not, initializing the new task based on the task-sharing backbone parameters of the nearest historical task; constructing a task classifier based on a multilayer perceptron and a Softmax normalization layer, and constructing a joint loss function; wherein, the joint loss function includes:

[0013] In the formula, Let the joint loss function be... Let λ be the cross-entropy loss function, and λ be the balancing hyperparameter.

[0014] Optionally, calculating the individual fitness score of the initial EEG signal based on the prototype feature vector includes: establishing a dynamically growing pool of specific experts based on the individual-specific backbone parameters of learned historical individuals; replicating the model parameters that performed best in the previous time step for the current task to construct an initial guidance model; fine-tuning the initial guidance model by comparative predictive coding based on the initial EEG signal to obtain a fine-tuned guidance model; predicting the initial EEG signal based on the fine-tuned guidance model to obtain a prediction probability distribution, and constructing a high-confidence pseudo-label dataset based on a preset confidence threshold; calculating the sample confidence mean based on the high-confidence pseudo-label dataset; retrieving the nearest individual from the specific expert pool based on the prototype feature vector; calculating the cosine similarity between the prototype feature center of the nearest individual and the prototype feature vector to obtain a prototype similarity component; and weighting and summing the sample confidence mean and the prototype similarity component based on a balanced hyperparameter to obtain the individual fitness score of the initial EEG signal.

[0015] Optionally, the task processing module is modularly extended based on the individual fitness score, including: if the individual fitness score is greater than or equal to a reuse threshold, identifying the nearest multiple historical expert models in a preset historical model library and reusing the parameter-weighted ensemble inference of the nearest multiple historical expert models; if the individual fitness score is less than the reuse threshold but greater than or equal to an extension threshold, performing few-sample fine-tuning on the model; if the individual fitness score is less than the extension threshold, retrieving the nearest historical expert model from the preset historical model library as the initial model for full-parameter training, and using the trained model as the extended model.

[0016] On the other hand, a multi-task EEG analysis system based on dual-stream decoupling and prototype guidance is also provided to implement the multi-task EEG analysis method based on dual-stream decoupling and prototype guidance provided in the embodiments of the present invention. The system includes: a preprocessing unit, a feature embedding unit, a first construction unit, a second construction unit, a training and analysis unit, and a scoring and expansion unit. The preprocessing unit is used to standardize the initial EEG signal data to obtain a standard input tensor and extract the prototype feature vector of the standard input tensor. The feature embedding unit is used to convert the standard input tensor into a high-dimensional semantic vector based on a deep convolutional neural network with shared weights to obtain an embedded sequence tensor. The first construction unit is used to establish a task-sharing backbone and individual... A body-specific backbone and feature decoupling mechanism are used to construct a dual-stream collaborative backbone architecture; the second construction unit is used to construct an independent task processing module based on task-level modular expansion, a multilayer perceptron-based task classifier, and a similarity-based dynamic parameter hot-start of the dual-stream collaborative backbone architecture; the training and analysis unit is used to train and optimize the task processing module based on a joint loss function to obtain the trained task processing module, and to analyze the embedded sequence tensor based on the trained task processing module; the scoring and expansion unit is used to calculate the individual fitness score of the initial EEG signal based on the prototype feature vector, and to modularly expand the task processing module based on the individual fitness score.

[0017] On the other hand, an electronic device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0018] On the other hand, a computer-readable storage medium is also provided, wherein program code is stored in the computer-readable storage medium, and the program code can be called by a processor to execute the method provided in the embodiments of the present invention.

[0019] This invention provides a method and system for EEG multitasking analysis based on dual-stream decoupling and prototype guidance. By constructing a shared and specific dual-stream decoupling architecture, it freezes general features and updates specific branches in a targeted manner, effectively overcoming memory interference and achieving compatibility between new and old knowledge. It introduces a dynamic structure expansion mechanism based on distribution drift, breaking the fixed capacity limitation and adaptively growing network nodes according to signal complexity to accurately capture dynamic evolution features. It integrates an explicit physical prototype guidance strategy and uses physical prior guidance parameters for hot start, significantly reducing dependence on labeled data, improving adaptation efficiency and biological interpretability across individual scenarios, and alleviating problems in existing technologies such as catastrophic forgetting, rigid network structure unable to adapt to dynamic evolution, and cold start difficulties caused by the lack of physical priors. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of an EEG multitasking analysis method based on dual-stream decoupling and prototype guidance provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a dual-stream collaborative backbone architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the processing flow of the dual-stream collaborative backbone architecture provided in this embodiment of the invention when facing continuously arriving heterogeneous task streams. Figure 4 This is a schematic diagram of an EEG multitasking analysis system based on dual-stream decoupling and prototype guidance provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0023] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] Figure 1 This is a flowchart of a multi-task EEG analysis method based on dual-stream decoupling and prototype guidance, provided by an embodiment of the present invention. Figure 1 As shown, the method specifically includes the following steps: Step S102: Standardize the initial EEG signal data to obtain a standard input tensor, and extract the prototype feature vector of the standard input tensor.

[0026] Step S104: The deep convolutional neural network based on shared weights transforms the standard input tensor into a high-dimensional semantic vector to obtain the embedded sequence tensor.

[0027] Step S106: Based on the task-sharing backbone, individual-specific backbone, and feature decoupling mechanism, construct a dual-stream collaborative backbone architecture.

[0028] Step S108: Based on task-level modular extension, a task classifier based on a multilayer perceptron, and a similarity-based dynamic parameter hot start for the dual-stream collaborative backbone architecture, an independent task processing module is constructed.

[0029] Step S110: Train and optimize the task processing module based on the joint loss function to obtain the trained task processing module, and analyze the embedded sequence tensor based on the trained task processing module.

[0030] Step S112: Calculate the individual fitness score of the initial EEG signal based on the prototype feature vector, and expand the task processing module modularly based on the individual fitness score.

[0031] Specifically, steps S102 and S104 aim to acquire high-quality EEG signals, convert raw EEG data from different sources, with varying sampling rates and channel configurations into a unified tensor format acceptable to the model, and extract two different data forms: one for temporal input to the deep network and the other for prototype features for fitness calculation.

[0032] Specifically, in step S102, the initial EEG signal data is standardized, including: Raw, multi-channel continuous EEG signals are acquired using an EEG acquisition device, with each channel recording the electrical activity of different brain regions. Let the initial EEG signal data be... ,in Indicates the number of electrode channels. This indicates the number of sampling time points. After resampling to a uniform frequency, Butterworth bandpass filtering (0.5Hz-50Hz), and power frequency notch filtering, Z-score normalization is performed on each channel to obtain the standard input tensor. .

[0033] Specifically, in step S102, the prototype feature vector of the standard input tensor is extracted, including: Step S1021: Perform a Fourier transform on the standard input tensor to obtain multiple frequency band signals.

[0034] Step S1022: Calculate the differential entropy of each frequency band signal to obtain the frequency domain characteristics of the standard input tensor.

[0035] Step S1023: Calculate the fractal dimension of the standard input tensor to obtain the temporal characteristics of the standard input tensor.

[0036] Step S1024: Calculate the rational asymmetry index of the symmetrical electrode pairs of the left and right hemispheres in the standard input tensor to obtain the spatial characteristics of the standard input tensor.

[0037] Step S1025 involves concatenating and normalizing the frequency domain features, time domain features, and spatial domain features to obtain the prototype feature vector of the standard input tensor.

[0038] Specifically, in order to assess the fitness level of the new individual in subsequent steps, the present invention... Explicit statistical features are extracted from the data. The following three-domain features are extracted for each channel: (1) Frequency domain feature extraction: capturing neural rhythm energy: For each EEG channel, the signal is decomposed into short-time Fourier transform (STFT) values. (0.5-4Hz) (4-8Hz) (8-13Hz) (13-30Hz) Five typical frequency bands (30-50Hz). Unlike traditional power spectral density (PSD), this invention calculates the differential entropy (DE) for each frequency band. EEG signals approximately follow a Gaussian distribution within a short time window; DE is equivalent to taking the logarithm of the power spectrum, which stretches the energy values ​​of the long-tailed distribution to a linear space, making it more consistent with the input assumptions for subsequent similarity calculations. The calculation formula is as follows:

[0039] in Indicates the channel index. Indicates the frequency band index. This represents the variance of the signal in that frequency band.

[0040] (2) Time-domain feature extraction: capturing signal nonlinearity complexity: To capture the essential characteristics of EEG signals as non-stationary chaotic signals, this invention abandons linear statistical moments such as mean and variance, which are susceptible to baseline drift, and instead calculates the Higuchi fractal dimension (HFD) of the signal. HFD quantitatively describes the coarsness and self-similarity of the EEG signal trajectory. The Higuchi algorithm is used to calculate the time series complexity of each channel. Set the maximum time interval. Calculate the curve length With scale The logarithmic slope is used as a complexity feature of the channel.

[0041] (3) Spatial feature extraction: capturing individual-specific brain functional connectivity patterns: This invention incorporates spatial topological features, selecting symmetrical electrode pairs between the left and right hemispheres (e.g., left frontal F3 and right frontal F4, left parietal P3 and right parietal P4), and calculates the rational asymmetry index (RASM):

[0042] This feature explicitly encodes the spatial coordination structure of the brain, enabling the feature vector to contain anatomically significant topological information.

[0043] The features from the three dimensions mentioned above are concatenated and normalized to generate the final prototype feature vector:

[0044] For each channel, 5 DE features, 1 HFD feature, and whole-brain features are spliced ​​together. For R-ASM features, since the entropy value and the fractal dimension have different dimensions, Min-Max normalization is used to map all features to... Intervals eliminate dimensional heterogeneity.

[0045] Specifically, step S104 further includes the following steps: Step S1041: The standard input tensor is segmented based on local windows divided along the time axis to generate multiple sub-segments; Step S1042: A deep one-dimensional convolutional neural network based on shared weights encodes the features of each sub-segment to obtain a feature map group. Step S1043: Perform global average pooling on the time dimension of the feature map group and rearrange it along the time step dimension to obtain the embedded sequence tensor.

[0046] Specifically, in order to transform continuous EEG time series into discrete sequences that can be efficiently processed by deep attention networks, while preserving the local frequency domain texture and transient waveform features of the signal, this invention utilizes a deep convolutional neural network with shared weights as a feature extractor to extract standard input tensors. Transform it into a high-dimensional semantic vector.

[0047] (1) Fixed-length time domain slices.

[0048] To reduce computational complexity and capture local context, the system sets the number of blocks to be [number missing]. , timeline Evenly divided into The first non-overlapping local window. For the first... time step Extracting sub-segments from the original signal :

[0049] in, The time length of a single block. This segmentation deconstructs a long time-series signal into a series of short time-series physical segments, with each segment fully preserving the synchronous activity information of all channels within the same time window.

[0050] (2) Feature embedding based on multi-layer convolution.

[0051] To extract robust features from noisy EEG signals, this invention does not employ simple linear projection, but instead constructs a deep one-dimensional convolutional neural network with shared weights as an embedding module (CNN module). For each sub-segment... The features are independently input into the CNN module for feature encoding. The first layer uses a large convolutional kernel to capture low-frequency slow-wave features and overall trends in the EEG signal. Subsequent layers are connected with multiple small-kernel convolutions, combined with a non-linear activation function (GELU) and batch normalization, to extract high-frequency transient features (such as spindle waves and K-complex waves) and detailed textures. Max pooling is inserted between layers to reduce feature dimensionality and enhance translation invariance.

[0052] (3) Global pooling and sequence recombination.

[0053] After CNN processing, each sub-segment The feature maps are mapped to a set of feature maps. To generate the final token vector, the temporal dimension of the feature maps is globally averaged and pooled, compressing them into a fixed-dimensional feature vector. ( (For the embedding dimension). The feature vectors are rearranged along the time step dimension to form the final embedding sequence tensor. :

[0054] Each vector It represents the depth feature representation of a specific time interval in the original signal.

[0055] Specifically, step S106 further includes the following steps: Step S1061: Initialize two deep neural network modules with the same structure but independent parameter spaces, which serve as a task-shared backbone and an individual-specific backbone, respectively. The task-shared backbone is used to learn the general neurophysiological patterns across subjects in the current task and construct the task-shared subspace; the individual-specific backbone is used to capture the inherent EEG feature shifts of the target subject and construct the individual-specific subspace. Step S1062: Construct a feature independence measurement module at the output end of the task-sharing backbone and the individual-specific backbone, and introduce the Hilbert-Schmidt independence criterion constraint index into the feature space of the feature independence measurement module to construct a decoupling loss term.

[0056] Figure 2 This is a schematic diagram of a dual-stream collaborative backbone architecture provided according to an embodiment of the present invention. Figure 2 As shown, specifically, to address the significant distribution differences in EEG signals among different subjects, i.e., the domain shift problem, this invention constructs a dual-stream collaborative backbone architecture based on statistical independence. This architecture comprises two parallel and structurally identical backbone networks: a task-shared backbone and an individual-specific backbone. A feature decoupling mechanism is introduced to force the task-shared backbone and the individual-specific backbone to remain orthogonal in the feature space, ensuring that they capture complementary semantic information respectively. This step embeds the temporal features generated in step 1. Mapped to a deep semantic space.

[0057] (1) Construction of a dual-flow collaborative backbone network: Initialize two deep neural network modules with identical structures but independent parameter spaces, denoted as the task-shared backbone. and individual-specific backbone Both receive the same embedded sequence tensor. , and output the feature representations respectively.

[0058] The aim is to learn common neurophysiological patterns across subjects in the current task and construct a task-shared subspace. During training, the parameters of this backbone network tend to capture common frequency band energy and waveform structures shared by the group, such as the standard morphology of K-complexes in sleep stages and the overall energy variation trend of ERD / ERS in motor imagery—features that do not change with individuals. The network parameters are updated during the pre-training phase and can be selectively frozen or fine-tuned during the adaptation phase for specific new individuals, serving as stable memories for the system.

[0059] The module aims to capture unique EEG characteristic shifts specific to individual subjects, constructing an individual-specific subspace, such as latency delays in P300 potentials, specific frequency band peak shifts, or amplitude scaling factors caused by scalp impedance. This module maintains high plasticity, allowing for rapid adaptation or expansion when new individuals are introduced, serving as a dynamic extension of the system.

[0060] (2) Feature decoupling based on statistical independence: To prevent the two backbone networks from learning redundant features, this invention introduces orthogonal decoupling constraints, constructing a feature independence measurement module at the output of the two backbone networks. Specifically, let... The feature matrix output by the shared backbone is used for tasks. The feature matrix is ​​the individual-specific backbone output. In the feature space, the Hilbert-Schmidt Independence Criterion (HSIC) constraint index is introduced to measure... and The non-linear dependency between them. Construct a decoupling loss term. :

[0061] By minimizing this metric, the output distributions of the two backbone networks are forced to become statistically independent.

[0062] (3) Backbone network internal architecture: Both the task-sharing backbone and the individual-specific backbone employ a Transformer encoder architecture to capture the long-term temporal dependencies of EEG signals. Each backbone network contains... The encoder layers are stacked, and each layer consists of the following sub-modules: Multi-head self-attention mechanism (MSA): Input features are first passed through... Multiple parallel attention heads are used to calculate the temporal correlations within the sequence:

[0063] in, , , These are the query, key, and value matrices, respectively. This mechanism allows the model to dynamically focus on key time segments within the signal.

[0064] Feedforward Neural Network (FFN): Consists of two fully connected layers and the GELU activation function, used to perform non-linear transformations and dimensionality adjustments on features.

[0065] Residual connections and layer normalization: Residual connections and layer normalization are introduced after each sublayer of MSA and FFN to prevent gradient vanishing and accelerate convergence.

[0066] (4) Feature fusion: To fully preserve the commonalities and unique characteristics after decoupling and avoid information aliasing caused by direct addition, this invention adopts a splicing scheme:

[0067] Among them, the dimension of the joint features is expanded to twice that of the single backbone dimension ( This constructs an augmented feature space that includes general patterns and individual biases. The task classifier receives... As input, the nonlinear combination weights of the two types of features are automatically learned through a multi-layer fully connected network, and finally mapped to the task category space to output the prediction result.

[0068] Specifically, step S108 further includes the following steps: The task routing module determines the task domain to which the current input signal belongs, and an independent task processing module is built based on the task domain; Determine whether the distance between the new task and all historical tasks exceeds a preset threshold; If so, then a task-sharing backbone of a two-stream collaborative backbone architecture is trained based on a contrastive predictive coding framework; If not, the new task is initialized based on the task-shared backbone parameters of the nearest historical task; A task classifier is constructed based on a multilayer perceptron and a Softmax normalization layer, and a joint loss function is also constructed.

[0069] Specifically, step S108 aims to address how a brain-computer interface system can overcome catastrophic forgetting while leveraging general EEG knowledge to achieve rapid cold start for new tasks when faced with a continuous stream of heterogeneous tasks. The system employs a combination of modular expansion and dynamic parameter warm start, constructing an independent parameter space for each new task and automatically selecting an initialization path based on the similarity between tasks. This allows for the rapid construction of a feature extractor adapted to the feature distribution of the new task while retaining the capabilities of the old task.

[0070] Figure 3 This is a schematic diagram illustrating the processing flow of a dual-stream collaborative backbone architecture provided by an embodiment of the present invention when facing continuously arriving heterogeneous task streams. For example... Figure 3 As shown: (1) Task-level modular extension: First, the task domain to which the current input signal belongs is determined by the task routing module, and the processing module is constructed accordingly. The system receives the task identifier at the current moment. The source of this identifier includes two methods: (1) Explicit instructions: Task mode switching instructions directly issued by an external host computer or user interface, such as the user selecting "sleep monitoring mode".

[0071] (2) Implicit inference: When there is no explicit instruction, the input signal is processed by a lightweight global discriminator. The spectral characteristics, such as power spectral density distribution, are classified to automatically infer which task domain the current signal belongs to.

[0072] For each newly identified task identifier Create a separate task processing module This module is defined as:

[0073] in, It is a task-specific shared backbone used to extract common features of all subjects under that task. It is a task-specific expert pool, containing a set of specific networks for addressing individual differences. . It is a task-specific classifier used as a decision layer to output the final classification result.

[0074] (2) Similarity-based dynamic parameter hot start: To avoid the problems of slow convergence and large sample requirements caused by training new task models from random initialization, this invention utilizes learned historical task knowledge and large-scale unlabeled data to construct a pre-training and transfer mechanism based on contrastive predictive coding (CPC).

[0075] If the new task differs significantly from all historical tasks, a general self-supervised learning task is constructed using multi-source heterogeneous EEG data accumulated in the database, independent of specific task labels. A shared backbone network is trained using the CPC framework. :

[0076] Specifically, embedding time-series features Input an autoregressive model and generate context vectors. The model predicts feature representations for the next k time steps by maximizing mutual information. This process forces the network to learn the underlying temporal structure and frequency domain patterns in EEG signals that are common to the current task, as initialization parameters for a shared backbone for new tasks.

[0077] If a certain historical task exists Distance to the new mission (Preset threshold) indicates that the two tasks are similar in nature. The system directly reuses them. Use shared backbone parameters to initialize new tasks:

[0078] (3) Construction of task-specific classification boundaries: After completing the construction and initialization of the feature extraction network, this invention presents a new task. Construct a dedicated decision-making layer and define a joint optimization objective to drive parameter updates across the entire network.

[0079] 1. Classifier structure: Task-specific classifier It consists of a multilayer perceptron and a Softmax normalization layer. Its input is a fused feature vector. .

[0080] First, global max pooling is used to aggregate the temporal features into a fixed-length vector:

[0081] Then, it is mapped to the task-defined category space through a fully connected layer:

[0082] in and These are the weights and biases of the classifier.

[0083] 2. Joint loss function optimization: To maintain the decoupling between shared and specific features while ensuring classification accuracy, this invention designs a joint loss function. Perform end-to-end training on the network.

[0084]

[0085] in, For the primary task of classification, the standard cross-entropy loss function is used to minimize the predicted label. Differences from the true label y:

[0086] The HSIC independence constraint loss is used to penalize the extraction of duplicate information by shared backbones and specific backbones. To balance the hyperparameters, used to adjust the strength of the decoupling constraints.

[0087] During training, the backpropagation algorithm is used to simultaneously update the current task module. The parameters of the shared backbone, specific backbone, and classifier are used until the model converges.

[0088] Specifically, step S112, which calculates the individual fitness score of the initial EEG signal based on the prototype feature vector, includes: Based on the individual-specific backbone parameters of historical individuals that have been learned, a dynamically growing pool of specific experts is established. Copy the model parameters that performed best in the previous time step for the current task and build the initial bootstrap model; The initial guidance model is fine-tuned by comparative prediction coding based on the initial EEG signal to obtain the fine-tuned guidance model. The initial EEG signal is predicted based on the fine-tuned guidance model to obtain the predicted probability distribution, and a high-confidence pseudo-label dataset is constructed based on a pre-set confidence threshold. Calculate the mean confidence level of the samples based on the high-confidence pseudo-label dataset; Retrieve the nearest individual from a specific expert pool based on prototype feature vectors; Calculate the cosine similarity between the prototype feature center and the prototype feature vector of the nearest individual to obtain the prototype similarity component; The individual fitness score of the initial EEG signal is obtained by weighted summation of the sample confidence mean and prototype similarity component based on balanced hyperparameters.

[0089] Specifically, when the system is running under a specific task and receives new individual data, in order to address the performance degradation caused by individual differences, this invention proposes a guided model-based approach. The pseudo-label generation mechanism, combined with a dual-threshold strategy, enables on-demand dynamic expansion of the network structure.

[0090] 1. Individual-level modular expansion: Corresponding to the modular expansion for different task domains in step S108, this step involves defining the task modules. Internally, a modular expansion mechanism is built for different subjects.

[0091] First, maintain a dynamically growing pool of specific experts: Each of them These represent the specific backbone parameters of historical individuals that have already been learned. When a new individual is added, the system no longer blindly trains from scratch. Instead, through an expansion mechanism, it dynamically decides whether to reuse or fine-tune parameters from the existing expert pool, or to add new individual-specific experts. This design ensures the decoupling of the model structure in the task dimension and its extension in the individual dimension.

[0092] 2. Pseudo-label generation based on the guided model: Since the EEG data of newly connected individuals often lacks real labels, in order to assess the fit between new individuals and existing models and to conduct subsequent supervised / semi-supervised training, this invention introduces a guided model mechanism to generate high-quality pseudo-labels.

[0093] (1) Initialization of the pilot model: Using individual fitness scores, copy the model parameters that performed best in the previous time step for the current task to construct the initial pilot model. If the current task is running for the first time or the expert pool is empty, the task sharing backbone will be used directly. Use a general pre-trained classification head as the initial bootstrap model.

[0094] (2) Self-supervised domain adaptation: using unlabeled EEG data from new individuals ,right Perform CPC fine-tuning. Utilize an autoregressive prediction task to maximize the mutual information of features at future time steps, forcing... It can quickly adapt to the marginal distribution of new individuals and extract robust temporal features.

[0095] (3) High-confidence screening: Input new individual samples into the adapted guided model Obtain the predicted probability distribution of the classifier output. Set confidence threshold for filtering. Only the predicted probability is retained. A high-confidence pseudo-label dataset is constructed by analyzing the samples and their predicted categories. :

[0096] This step effectively filters out noisy samples located in the ambiguous region of the decision boundary.

[0097] 3. Individual fitness assessment: To more comprehensively and robustly quantify the current model knowledge base's ability to cover new individuals, this invention proposes a comprehensive fitness evaluation mechanism that integrates implicit model confidence and explicit physical prototype similarity.

[0098] Defined as the mean of the prediction confidence scores for all samples in the pseudo-label dataset, it reflects the model's familiarity with new individual data.

[0099] This reflects the physical similarity of the new individual to the historical population in terms of neurophysiological characteristics. First, the prototype feature vector of the new individual is calculated. (Including frequency domain DE, time domain HFD, and spatial domain RASM features). Secondly, a specific expert pool is retrieved. The prototype feature centers of all historical individuals stored in the database .calculate The cosine similarity with the nearest neighbor in the historical feature database is used as the prototype similarity component:

[0100] The higher the index, the more similar the new individual's electroencephalographic properties are to a certain population group known to the system.

[0101] Introducing equilibrium hyperparameters ( The two components are then weighted and summed to obtain the final fitness score. :

[0102] Specifically, step S112 involves modularly expanding the task processing module based on individual fitness scores, including: If an individual's fitness score is greater than or equal to the reuse threshold, the nearest multiple historical expert models are identified in the preset historical model library, and the parameter-weighted ensemble inference of the nearest multiple historical expert models is reused. If an individual's fitness score is less than the reuse threshold but greater than or equal to the expansion threshold, then the model is fine-tuned with a few samples. If an individual's fitness score is less than the expansion threshold, the nearest historical expert model is retrieved from the preset historical model library as the initial model for full-parameter training, and the trained model is used as the expansion model.

[0103] Specifically, embodiments of the present invention preset two judgment thresholds and expand the threshold. With reuse threshold (satisfy ).according to Based on the numerical relationship between these two thresholds, the system divides the adaptation scenarios of the new individual into three intervals and triggers corresponding processing strategies. The highest-level model is copied for building the initial bootstrap model. .

[0104] (1) High fitness interval (exists) ): Weighted reuse mode.

[0105] The new individuals not only have high model prediction confidence, but their physical prototype characteristics (DE / HFD / RASM) also highly match the historical database, making them typical samples. The system does not perform backpropagation training. It directly relies on... The top-K nearest-neighbor historical experts identified during the calculation process are reused for weighted ensemble inference. This achieves new user response with extremely low computational overhead.

[0106] (2) Medium fitness interval (exists) ): Parameter fine-tuning mode.

[0107] The presence of some differences in physical characteristics or a moderate model confidence level indicates a domain shift, but it is within the model's expressive power. The network topology should be kept unchanged. Utilize... As a training set, the task-level shared backbone of the model is used. Specific backbone and task classifier Perform fine-tuning with a small sample size. This corrects for individual-specific biases while preventing catastrophic forgetting of general EEG knowledge.

[0108] (3) Low fitness range (all) ): Network extension mode.

[0109] New individuals are outliers, for example, possessing unique pathological characteristics that prevent the model from making confident predictions, triggering a network structure expansion mechanism. In the current task module... In addition to the existing pool of specific experts, instantiate a new individual-specific backbone network. Its network architecture is consistent with that of other experts in the pool. Fitness scores are used. Search the existing expert pool for historical individual models that most closely resemble the physical characteristics of the new individual. .Will Copy the complete parameters to This serves as the starting point for training, rather than using random initialization. It utilizes a pseudo-labeled dataset of new individuals. right Shared backbone with tasks Perform full-parameter training. After training converges, formally encapsulate the new backbone network and incorporate it into the specific expert pool. This increases the expert pool size from K to K+1. This ensures the system maintains high performance even when facing extreme individual differences, and the new structure does not interfere with the stability of the existing knowledge base.

[0110] If a new individual completes the adaptation process in step S112, the system will bind the individual's ID to a specific backbone. The next time the user visits, the bound parameters will be retrieved directly, without having to rerun the adaptation process each time, unless a long-term update provided in this embodiment of the invention is triggered.

[0111] Specifically, in response to the non-stationary drift of the EEG signal of the same subject during long-term use due to changes in physiological state (such as fatigue, emotional fluctuations) or environmental changes, this invention further provides long-term adaptive capability on the basis of the above-mentioned initial adaptation.

[0112] In practice, when the system detects that an individual is under prolonged continuous monitoring, it triggers an incremental adaptation mechanism, forcibly freezing the specific backbone network corresponding to that individual. All existing encoder layer parameters are preserved to maintain the learned steady-state feature structure of the individual. A lightweight time-varying adaptation layer (e.g., a single-layer Transformer encoder or Adapter module) is dynamically added at the end of the backbone network. The gradient of this new layer is updated instantly using only the newly received online data stream. This mechanism can capture the slow time-varying drift of the signal at extremely low computational cost, enabling the model to continuously track and dynamically calibrate individual state changes without destroying the original feature space.

[0113] As described above, the embodiments of the present invention provide a multi-task EEG analysis method based on dual-stream decoupling and prototype guidance, which has the following technical advantages compared with the prior art: (1) By constructing a dual-stream backbone architecture with task sharing and individual specificity, and introducing HSIC independence constraints, this invention achieves orthogonal decoupling of general features and specific features in the feature space. When adapting to new tasks or new individuals, only specific branches are updated and the shared backbone is frozen, thereby preserving historical knowledge, effectively suppressing catastrophic forgetting, and achieving stable compatibility of multi-task knowledge.

[0114] (2) In view of the non-stationarity and individual differences of EEG signals, this invention proposes a dynamic structural expansion mechanism based on fitness assessment, which integrates model confidence and physical prototype similarity to construct a comprehensive fitness index, automatically selects parameter reuse, fine-tuning or new branches, breaks through the fixed model capacity limit, realizes the model capacity to grow as needed with data complexity, and avoids redundancy.

[0115] (3) This invention explicitly introduces physical prototype features such as frequency domain DE, time domain HFD, and spatial domain RASM, and uses them as priors for the design of hot start and adaptation strategies for new individuals. Under conditions of minimal or even no annotation, rapid evaluation and parameter guidance are achieved through prototype similarity, which significantly reduces the dependence of cold start on labeled data and improves cross-individual adaptation efficiency and system usability.

[0116] (4) By explicitly using physical indicators such as frequency domain energy entropy, fractal dimension and brain functional connectivity in the decision-making process, this invention transforms the model adaptation strategy from black box experience to judgment based on interpretable physiological attributes, improves the interpretability and robustness of the algorithm, provides traceable decision-making basis for clinical and long-term monitoring, and enhances the credibility of the system.

[0117] Figure 4 This is a schematic diagram of an EEG multitasking analysis system based on dual-stream decoupling and prototype guidance, according to an embodiment of the present invention. Figure 4 As shown, the system includes: a preprocessing unit 10, a feature embedding unit 20, a first construction unit 30, a second construction unit 40, a training and analysis unit 50, and a scoring and expansion unit 60.

[0118] Specifically, the preprocessing unit 10 is used to standardize the initial EEG signal data to obtain a standard input tensor and extract the prototype feature vector of the standard input tensor. The feature embedding unit 20 is used to transform the standard input tensor into a high-dimensional semantic vector by a deep convolutional neural network based on shared weights, thereby obtaining an embedded sequence tensor. The first building unit 30 is used to build a dual-stream collaborative backbone architecture based on task-sharing backbone, individual-specific backbone and feature decoupling mechanism; The second building unit 40 is used to build an independent task processing module based on task-level modular extension, a multilayer perceptron-based task classifier, and a similarity-based dynamic parameter hot start for the dual-stream collaborative backbone architecture. The training and analysis unit 50 is used to train and optimize the task processing module based on the joint loss function to obtain the trained task processing module, and to analyze the embedded sequence tensor based on the trained task processing module. The scoring and extension unit 60 is used to calculate the individual fitness score of the initial EEG signal based on the prototype feature vector, and to modularly extend the task processing module based on the individual fitness score.

[0119] The present invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0120] The present invention also provides a computer-readable storage medium storing program code, which can be called by a processor to execute the method provided in the embodiments of the present invention.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0122] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0123] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0126] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-task EEG analysis method based on dual-stream decoupling and prototype guidance, characterized in that, The method includes: The initial EEG signal data is standardized to obtain a standard input tensor, and the prototype feature vector of the standard input tensor is extracted. A deep convolutional neural network based on shared weights transforms the standard input tensor into a high-dimensional semantic vector, resulting in an embedded sequence tensor. Based on task-sharing backbone, individual-specific backbone, and feature decoupling mechanism, a dual-stream collaborative backbone architecture is constructed. An independent task processing module is constructed based on task-level modular extension, a multilayer perceptron-based task classifier, and a similarity-based dynamic parameter hot start for the dual-stream collaborative backbone architecture. The task processing module is trained and optimized based on the joint loss function to obtain the trained task processing module, and the embedded sequence tensor is analyzed based on the trained task processing module. The individual fitness score of the initial EEG signal is calculated based on the prototype feature vector, and the task processing module is modularly extended based on the individual fitness score.

2. The method according to claim 1, characterized in that, Extracting the prototype feature vector of the standard input tensor includes: Perform a Fourier transform on the standard input tensor to obtain multiple frequency band signals; The differential entropy of each frequency band signal is calculated to obtain the frequency domain characteristics of the standard input tensor; Calculate the fractal dimension of the standard input tensor to obtain the temporal features of the standard input tensor; Calculate the rational asymmetry index of the symmetrical electrode pairs of the left and right hemispheres in the standard input tensor to obtain the spatial characteristics of the standard input tensor. The frequency domain features, time domain features, and spatial domain features are concatenated and normalized to obtain the prototype feature vector of the standard input tensor.

3. The method according to claim 1, characterized in that, A deep convolutional neural network based on shared weights transforms the standard input tensor into a high-dimensional semantic vector, resulting in an embedded sequence tensor, including: The standard input tensor is segmented based on local windows divided along the time axis to generate multiple sub-segments; A deep one-dimensional convolutional neural network based on shared weights encodes features for each sub-segment to obtain a set of feature maps; The feature map group is subjected to global average pooling along the time step dimension and rearranged to obtain the embedded sequence tensor.

4. The method according to claim 1, characterized in that, Based on task-sharing backbone, individual-specific backbone, and feature decoupling mechanism, a two-stream collaborative backbone architecture is constructed, including: Two deep neural network modules with identical structures but independent parameter spaces are initialized, serving as a task-shared backbone and an individual-specific backbone, respectively. The task-shared backbone is used to learn the general neurophysiological patterns across subjects in the current task and construct a task-shared subspace. The individual-specific backbone is used to capture the inherent EEG feature shifts of the target subject and construct an individual-specific subspace. A feature independence measurement module is constructed at the output ends of the task-sharing backbone and the individual-specific backbone, and a Hilbert-Schmidt independence criterion constraint index is introduced into the feature space of the feature independence measurement module to construct a decoupling loss term; wherein, the decoupling loss term includes: ； In the formula, For the decoupling loss term, These are the output feature matrices of the task-shared backbone and the individual-specific backbone, respectively, with HSIC representing the Hilbert-Schmidt independence criterion constraint index. The output splicing scheme of the dual-stream collaborative backbone architecture includes: ； In the formula, The joint features output by the dual-stream collaborative backbone architecture are defined as follows: D is the embedding dimension, and N is the number of feature vectors in the embedding sequence tensor.

5. The method according to claim 4, characterized in that, Based on task-level modular extension, a multilayer perceptron-based task classifier, and similarity-based dynamic parameter hot-start of the dual-stream collaborative backbone architecture, an independent task processing module is constructed, including: The task routing module determines the task domain to which the current input signal belongs, and an independent task processing module is constructed based on the task domain; Determine whether the distance between the new task and all historical tasks exceeds a preset threshold; If so, then the task-sharing backbone of the dual-stream collaborative backbone architecture is trained based on the contrastive predictive coding framework; If not, the new task is initialized based on the task-shared backbone parameters of the nearest historical task; A task classifier is constructed based on a multilayer perceptron and a Softmax normalization layer, and a joint loss function is built; wherein, the joint loss function includes: ； In the formula, Let the joint loss function be... Let λ be the cross-entropy loss function, and λ be the balancing hyperparameter.

6. The method according to claim 1, characterized in that, The individual fitness score of the initial EEG signal is calculated based on the prototype feature vector, including: Based on the individual-specific backbone parameters of historical individuals that have been learned, a dynamically growing pool of specific experts is established. Copy the model parameters that performed best in the previous time step for the current task and build the initial bootstrap model; Based on the initial EEG signal, the initial guidance model is fine-tuned by comparative prediction coding to obtain the fine-tuned guidance model; Based on the fine-tuned guidance model, the initial EEG signal is predicted to obtain the predicted probability distribution, and a high-confidence pseudo-label dataset is constructed based on a preset confidence threshold. Based on the high-confidence pseudo-label dataset, calculate the mean confidence level of the samples; Based on the prototype feature vector, the nearest individual is retrieved from the specific expert pool; Calculate the cosine similarity between the prototype feature center of the nearest individual and the prototype feature vector to obtain the prototype similarity component; The individual fitness score of the initial EEG signal is obtained by weighted summation of the mean confidence score of the sample and the prototype similarity component based on the balanced hyperparameters.

7. The method according to claim 1, characterized in that, The task processing module is modularly extended based on the individual fitness score, including: If the individual fitness score is greater than or equal to the reuse threshold, the nearest multiple historical expert models are identified in the preset historical model library, and the parameter weighted ensemble reasoning of the nearest multiple historical expert models is reused. If the individual fitness score is less than the reuse threshold and greater than or equal to the expansion threshold, then the model is fine-tuned with a few samples. If the individual fitness score is less than the expansion threshold, the nearest historical expert model is retrieved from the preset historical model library as the initial model for full parameter training, and the trained model is used as the expansion model.

8. A multi-task EEG analysis system based on dual-stream decoupling and prototype guidance, characterized in that, The system is used to implement the EEG multitasking analysis method based on dual-stream decoupling and prototype guidance as described in any one of claims 1-7; the system includes: a preprocessing unit, a feature embedding unit, a first construction unit, a second construction unit, a training and analysis unit, and a scoring and expansion unit; The preprocessing unit is used to standardize the initial EEG signal data to obtain a standard input tensor and extract the prototype feature vector of the standard input tensor. The feature embedding unit is used to transform the standard input tensor into a high-dimensional semantic vector based on a deep convolutional neural network with shared weights, thereby obtaining an embedded sequence tensor. The first building unit is used to construct a dual-stream collaborative backbone architecture based on task-sharing backbone, individual-specific backbone, and feature decoupling mechanism; The second building unit is used to build an independent task processing module based on task-level modular expansion, a multilayer perceptron-based task classifier, and a similarity-based dynamic parameter hot start for the dual-stream collaborative backbone architecture. The training and analysis unit is used to train and optimize the task processing module based on the joint loss function to obtain the trained task processing module, and to analyze the embedded sequence tensor based on the trained task processing module. The scoring and expansion unit is used to calculate the individual fitness score of the initial EEG signal based on the prototype feature vector, and to modularly expand the task processing module based on the individual fitness score.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images