Computational acceleration methods based on EEG data features

Abstract

The application provides an electroencephalogram data feature-based operation acceleration method, belongs to the field of electroencephalogram data processing, and is used for solving the problems of poor noise resistance and unreasonable operation resource scheduling of electroencephalogram data processing in the related art. The method reserves multi-dimensional coupling information by constructing a space-time-frequency third-order tensor, performs high-order decomposition on the tensor and maps the tensor to a special orthogonal group manifold, extracts manifold curvature and geodesic distance as geometric features, dynamically schedules heterogeneous hardware operation modes according to the core tensor rank and the geometric features, realizes efficient and robust acceleration of electroencephalogram data processing, and balances operation accuracy and resource efficiency.

Description

Technical Field

[0001] This application relates to the field of EEG data processing, and more particularly to a computation acceleration method based on EEG data features. Background Technology

[0002] Currently, EEG data processing largely relies on linear transformation methods to extract features, failing to fully consider the nonlinear coupling characteristics of EEG signals and the spatial-temporal-frequency correlation information. This results in poor noise resistance and susceptibility to noise interference, leading to feature misjudgment. Furthermore, existing computing resource scheduling often employs fixed patterns, unable to adapt computing precision and resource consumption to the dynamic changes in EEG rhythm modalities. This leads to low computing efficiency, resource waste, or insufficient accuracy, making it difficult to meet the demands of real-time EEG processing. Summary of the Invention

[0003] This application provides a computation acceleration method based on EEG data features, which can improve computational efficiency and noise resistance by capturing the multi-dimensional coupling features and geometric properties of EEG signals.

[0004] Firstly, this application provides a computation acceleration method based on EEG data features. It includes: constructing a spatial-temporal-frequency third-order tensor of the EEG signal to preserve the coupling and correlation information of the EEG signal in the spatial channel, temporal sequence, and frequency component dimensions; performing high-order decomposition on the third-order tensor to obtain a core tensor and spatial, temporal, and frequency factor matrices, and mapping the factor matrices to a special orthogonal group manifold; extracting manifold curvature and geodesic distance as geometric features of the EEG rhythm modality based on Riemannian metrics; and dynamically scheduling computational resources according to the rank parameter of the core tensor and the geometric features to accelerate the computation of EEG data processing.

[0005] By adopting the above technical solution, a third-order tensor is constructed to preserve the multi-dimensional coupling information of EEG signals. Combined with high-order decomposition and manifold geometric feature extraction, accurate characterization of EEG rhythm modes can be achieved. Furthermore, based on the core tensor rank and geometric features, computational resources can be dynamically scheduled, which can improve computational efficiency while ensuring processing accuracy and solve the problems of poor noise resistance and unreasonable resource scheduling in traditional methods.

[0006] Furthermore, the construction of the space-time-frequency third-order tensor involves performing a short-time Fourier transform on the multi-channel EEG signal to generate a tensor structure containing the dimensions of the number of channels, the number of time points, and the number of frequency bins.

[0007] By adopting the above technical solution, multi-channel EEG signals can be converted into third-order tensors using short-time Fourier transform, which can effectively preserve the correlation between spatial channels, time series and frequency components, providing a more comprehensive data foundation for subsequent feature extraction.

[0008] Furthermore, the mapping of the factor matrix to a special orthogonal group manifold involves mapping the spatial, temporal, and frequency factor matrices to special orthogonal group manifolds of corresponding dimensions, thereby characterizing the topological geometric properties of the factor matrix through the manifold.

[0009] By adopting the above technical solution, the factor matrix is ​​mapped to a special orthogonal group manifold. The topological properties of the manifold can be used to characterize the geometric properties of EEG signals, enhance the robustness of features to noise, and provide a more reliable basis for subsequent modality recognition.

[0010] Furthermore, the manifold curvature and geodesic distance are extracted as geometric features. The manifold curvature is obtained by calculating the norm of the second covariant derivative of the core tensor element, and the geodesic distance is obtained by solving the shortest path length connecting different rhythmic modes on the manifold.

[0011] By adopting the above technical solution, the curvature and geodesic distance obtained by solving the second-order covariant derivative norm of the core tensor element and the shortest path of the manifold can accurately reflect the geometric differences of different EEG rhythm modalities and improve the discriminativeness of modality recognition.

[0012] Furthermore, the method of processing EEG signals based on geometric features utilizes the noise robustness of manifold topology to perform multi-level high-order decomposition on the EEG signals, and then fuses the features at different scales after assigning dynamic weights. The dynamic weights are adjusted according to the noise statistical characteristics of each scale.

[0013] By adopting the above technical solution and combining the noise resistance of manifold topology to perform multi-level decomposition and dynamic weight fusion, the interference of noise on effective features can be suppressed, and the reliability of feature extraction can be improved, which is especially suitable for high-noise scenarios.

[0014] Furthermore, the multi-level high-order decomposition and fusion involves performing multi-level high-order decomposition on the EEG signal, and weighting the features at each scale through an attention mechanism. The weights of the attention mechanism are negatively correlated with the noise variance at each scale.

[0015] By adopting the above technical solution and using the attention mechanism to dynamically adjust the weights of each scale according to the noise variance, the contribution of effective features can be enhanced, the influence of noise scale can be reduced, and the noise resistance effect of feature fusion can be further improved.

[0016] Furthermore, the identification of EEG rhythm modes based on geometric features involves classifying the feature vectors composed of the manifold curvature and geodesic distance using a classifier, and outputting the probability distribution of each rhythm mode.

[0017] By adopting the above technical solution, based on the classification and probability distribution output of geometric feature vectors, accurate identification of EEG rhythm modes can be achieved, providing a reliable basis for subsequent computing resource scheduling.

[0018] Furthermore, the dynamic scheduling of computing resources based on the rank parameter and geometric features of the core tensor is achieved by quantifying the rank and geometric features of the core tensor to obtain a computing load index, and switching the computing mode of heterogeneous hardware based on the threshold of the computing load index. The heterogeneous hardware includes field-programmable gate arrays, central processing units, and graphics processing units.

[0019] By adopting the above technical solution and switching heterogeneous hardware modes based on computational load indicators, hardware with different computational capabilities can be adapted according to the complexity of EEG signals, thereby achieving efficient utilization of computing resources.

[0020] Furthermore, the computational complexity index obtained by quantizing the core tensor rank and geometric features is obtained by fusing the core tensor volume and the maximum curvature of the current mode through a balance coefficient. The core tensor volume is the product of the rank of the spatial, temporal, and frequency dimensions.

[0021] By adopting the above technical solution, the computational complexity index is obtained by integrating the core tensor volume and modal curvature, which can comprehensively reflect the signal complexity and make the scheduling of computing resources more in line with actual processing needs.

[0022] Furthermore, the method of switching the computing mode of heterogeneous hardware based on the computing load threshold is to enable lightweight fixed-point computing when the computing load is lower than the first threshold and high-precision floating-point computing when the computing load is higher than the second threshold, thereby achieving a dynamic balance between computing precision and resource efficiency.

[0023] By adopting the above technical solution, the computational precision mode can be switched according to the computational load threshold, which can reduce the consumption of ineffective resources while ensuring processing accuracy, and achieve a dynamic balance between accuracy and efficiency.

[0024] In summary, this application has at least the following beneficial effects:

[0025] A computation acceleration method based on EEG data features is provided, which improves the efficiency and noise resistance of EEG data processing;

[0026] By preserving information through multi-dimensional coupling and extracting geometric features, the reliability of modality recognition is enhanced.

[0027] By dynamically scheduling resources, a balance between computational accuracy and resource efficiency is achieved.

[0028] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0029] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0030] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0031] Figure 2 A flowchart of a computation acceleration method based on EEG data features is shown in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0034] This application provides a computation acceleration method based on EEG data features, which preserves multi-dimensional coupling information of EEG, improves noise resistance and modality recognition accuracy through geometric features, dynamically schedules computing resources, balances efficiency and accuracy, and optimizes the EEG data processing effect.

[0035] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0036] Reference Figure 1The operating environment includes a multi-channel EEG signal acquisition device, used to acquire raw EEG signals and synchronous reference signals, and output multi-channel time-series data. This device connects to the subsequent computing platform via a data transmission interface. The operating environment also includes a heterogeneous computing platform, comprising a central processing unit (CPU), a graphics processing unit (GPU), and a field-programmable gate array (FPGA). The CPU is responsible for overall computation scheduling and data coordination, the GPU handles complex computational tasks such as manifold geometric feature extraction, and the FPGA performs real-time high-order tensor decomposition and lightweight computation. These three components interact and collaborate via an internal bus. The operating environment also includes a hardware resource scheduling and control module, connected to the heterogeneous computing platform, used to receive computational load indicators and generate hardware mode switching instructions, dynamically adjusting the computational precision and resource allocation of each computing unit. Finally, the operating environment includes a data storage device, connected to both the EEG signal acquisition device and the heterogeneous computing platform, used to store raw EEG data, intermediate processing results, and feature parameters, providing data support for the entire computational acceleration process.

[0037] Figure 2 A flowchart of a computation acceleration method based on EEG data features is shown in an embodiment of this application.

[0038] Reference Figure 2 The method specifically includes the following steps:

[0039] S1: Construct a spatial-temporal-frequency third-order tensor for EEG signals to preserve the coupling and correlation information of EEG signals in the spatial channel, temporal sequence, and frequency component dimensions.

[0040] In this step, the construction of the space-time-frequency third-order tensor involves performing a short-time Fourier transform on the multi-channel EEG signal to generate a tensor structure containing the dimensions of the number of channels, the number of time points, and the number of frequency bins. Specifically, the multi-channel EEG signal is first represented as a two-dimensional matrix, where the matrix elements... Representing the Each channel is in The sampled value at time 10:00. The value range is 1 to The total number of channels is determined by the number of electrodes in the EEG acquisition device, and is usually 64 or 128. The value range is 1 to ( The total number of sampling points is determined by the sampling frequency. Determined by the collection duration, Typical values ​​are 256Hz or 512Hz.

[0041] To convert this two-dimensional time series into a three-dimensional structure containing frequency information, a short-time Fourier transform needs to be performed independently on the signal of each channel. During the transform process, a short-time Fourier transform of length [missing information] is used. A sliding window is used to extract segments from a time series; the window function... Hanning windows were selected. Window length The selection needs to meet the temporal resolution requirements of the main EEG rhythms (usually 256, corresponding to 1 second @ 256Hz sampling rate), and the step size between windows... Set as (i.e., 64 points) to balance time resolution and computational efficiency.

[0042] For the The first channel Signal within a window Its short-time Fourier transform result is:

[0043]

[0044] in, For window index (corresponding to the number of time points). For frequency points (corresponding to the number of frequency bins). The unit is the imaginary number. Frequency grid number. Determined by the target frequency range, it typically covers 1Hz to 128Hz, with each sub-box corresponding to a frequency interval. (e.g., at a sampling rate of 256Hz) ,therefore The value is

[0045] The final generated space-time-frequency third-order tensor In, tensor elements (Amplitude value), where Total number of time windows , respectively corresponding to the number of channels Time point number and frequency bin number The three dimensions fully preserve the frequency characteristic changes and interrelationships of different channel signals during the time evolution process.

[0046] S2: Perform higher-order decomposition on the third-order tensor to obtain the core tensor and spatial, temporal, and frequency factor matrices, and map the factor matrices to a special orthogonal group manifold. Extract the manifold curvature and geodesic distance as geometric features of the EEG rhythm modality based on Riemannian metric.

[0047] After extracting manifold curvature and geodesic distance as geometric features, EEG signals are processed based on these geometric features.

[0048] Specifically, in this step, mapping the factor matrix to a special orthogonal group manifold involves mapping the spatial, temporal, and frequency factor matrices to special orthogonal group manifolds of their corresponding dimensions, thereby characterizing the topological geometric properties of the factor matrix through the manifold. Extracting manifold curvature and geodesic distance as geometric features, the manifold curvature is obtained by calculating the norm of the second covariant derivative of the core tensor elements, and the geodesic distance is obtained by solving for the shortest path length connecting different rhythmic modes on the manifold. Processing EEG signals based on these geometric features utilizes the noise resistance of the manifold topology. Robustness is achieved by performing multi-level high-order decomposition on the EEG signal, and then fusing features at different scales after assigning dynamic weights. The dynamic weights are adjusted according to the noise statistical characteristics of each scale. The multi-level high-order decomposition and fusion involves performing multi-level high-order decomposition on the EEG signal, and weighting the features at each scale through an attention mechanism. The weights of the attention mechanism are negatively correlated with the noise variance at each scale. The identification of EEG rhythm modes based on geometric features involves classifying the feature vectors composed of manifold curvature and geodesic distance through a classifier, and outputting the probability distribution of each rhythm mode.

[0049] When performing higher-order decomposition on a third-order tensor, the Higher-Order Singular Value Decomposition (HOSVD) algorithm is used. The specific process is as follows: the space-time-frequency third-order tensor is decomposed... Decomposed into core tensor and three factor matrices, including the spatial factor matrix. ( The spatial rank is determined by the correlation of the channel signals, and is usually taken as 8-16; the time factor matrix. The time rank is determined by the dynamic rate of change of the signal, and is usually taken as 16-32; the frequency factor matrix. ( The frequency rank, determined by the rhythm frequency band distribution, is typically taken as 4-8. The decomposition satisfies... ,in Indicates the first Tensor-matrix multiplication of dimensions, core tensor The element values ​​reflect the coupling strength of the three dimensional factors.

[0050] When mapping the factor matrix to a special orthogonal group manifold, the spatial factor matrix... Mapped to ( (Special orthogonal group), satisfying and Time factor matrix Mapped to Frequency factor matrix Mapped to The mapping process is achieved through Gram-Schmidt orthogonalization, ensuring that the mapped matrix retains the topological structure of the causal submatrices and possesses rotation invariance. Its topological properties are independent of rotation operations.

[0051] When extracting manifold curvature, the core tensor Each element Calculate its second covariant derivative. The derivative is obtained through the connection coefficient on the manifold. Solve Then calculate its Frobenius norm. manifold curvature For this norm in modal The average value of the corresponding region. When extracting geodesic distance, for any two rhythmic modes... Find the shortest path connecting the two on the manifold. ,in correspond Feature points, correspond Feature points, geodesic distance ,in Let be the tangent vector of the path. Riemannian metric on a manifold (defined as (The tangent vector).

[0052] When performing multi-level high-order decomposition on EEG signals, the number of decomposition layers is set to 5-8, with each layer corresponding to a different frequency scale (layer 1 corresponds to high frequency, layer 8 corresponds to low frequency), and the features obtained from each layer decomposition are... (No. Layer mode The energy) is weighted and fused through an attention mechanism, with weights satisfy ,in For the first The noise variance of a layer (estimated from the variance of the detail coefficients of that layer) is the fused feature. .

[0053] When classifying geometric feature vectors using a classifier, the feature vectors are: The classifier uses a support vector machine (SVM), and its output is transformed into a probability distribution through a softmax function. ,in For classifiers to mode The original output is used to finally output the probability values ​​of each rhythm mode.

[0054] S3: Based on the rank parameter of the core tensor and the geometric features, dynamically schedule computing resources to accelerate the processing of EEG data.

[0055] In this method, the dynamic scheduling of computing resources based on the rank parameter and geometric features of the core tensor is achieved by quantifying the core tensor rank and geometric features to obtain a computing load index. The computing load index threshold is then used to switch the computing mode of heterogeneous hardware, including field-programmable gate arrays, central processing units, and graphics processing units. The quantification of the core tensor rank and geometric features to obtain the computing load index is achieved by merging the core tensor volume and the maximum curvature of the current mode using a balance coefficient. The core tensor volume is the product of the rank of the spatial, temporal, and frequency dimensions. Switching the computing mode of heterogeneous hardware based on the computing load index threshold involves enabling lightweight fixed-point arithmetic when the computing load is below a first threshold and enabling high-precision floating-point arithmetic when the computing load is above a second threshold, thus achieving a dynamic balance between computing precision and resource efficiency.

[0056] When quantizing the core tensor rank and geometric features to obtain computational complexity indicators, the core tensor volume... The calculation formula is: ,in For spatial dimension rank, For the time dimension rank, The frequency dimension rank is derived from the core tensor parameters obtained after performing a higher-order decomposition on the third-order tensor. The maximum curvature of the current mode. For the currently identified EEG rhythm modes The maximum value of the corresponding manifold curvature originates from geometric features extracted based on Riemannian metrics. (Balance coefficient) The range of values ​​is This value is typically set based on the emphasis on computational efficiency and accuracy in the actual application scenario, with a typical value of 0.6. At this point, the computational load index... The calculation formula is: This indicator comprehensively reflects the computational complexity required for current EEG data processing.

[0057] When switching the computing mode of heterogeneous hardware based on a computing load threshold, the first threshold... Second threshold Statistical analysis of historical computational data determines that it typically meets the following requirements. ,For example ,in This represents the maximum amount of computation the system can handle (determined by the maximum performance of heterogeneous hardware). When When using a Field Programmable Gate Array (FPGA) to perform lightweight fixed-point arithmetic, the data bit width is typically 8-16 bits; When the central processing unit (CPU) is activated, it performs medium-precision arithmetic operations with a data bit width of 16-32 bits; when At this time, the graphics processing unit (GPU) is enabled to perform high-precision floating-point operations with a data bit width of 32 bits. Through this dynamic switching mechanism, the accuracy of EEG data processing is ensured while minimizing the consumption of computing resources.

[0058] By constructing a space-time-frequency three-order tensor, the coupling and correlation information of EEG signals in multiple dimensions is fully preserved, overcoming the limitation of traditional single-dimensional analysis in losing key correlation features. High-order decomposition of the tensor is performed and mapped to a special orthogonal manifold, transforming signal features into topologically robust geometric properties (curvature, geodesic distance). The natural anti-noise properties of manifold topology are utilized to reduce noise interference in feature extraction. An attention mechanism is used to dynamically weight and fuse multi-level decomposed features, further strengthening effective features, suppressing noise scale, and improving the reliability of modality recognition. Based on the core tensor rank and geometric feature quantification computational complexity indicators, the computational mode of heterogeneous hardware is dynamically scheduled, ensuring precise matching of computational resource allocation with the complexity of EEG rhythms—low-complexity modalities are adapted to lightweight computation to save resources, while high-complexity modalities utilize high-precision computation to ensure processing accuracy.

[0059] The above-mentioned technical means form a closed-loop logic from signal modeling, feature extraction to resource scheduling: the preservation of multi-dimensional coupled information provides a data foundation for accurate feature extraction, the noise resistance of manifold geometric features ensures the stability of modality recognition in high-noise scenarios, and dynamic resource scheduling achieves a balance between computational efficiency and accuracy. Finally, it is deduced that this method can achieve efficient and robust processing of EEG data with better resource consumption in complex noisy environments.

[0060] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0061] In summary, this application has at least the following beneficial effects:

[0062] 1. By constructing a space-time-frequency third-order tensor and combining it with higher-order decomposition, the multi-dimensional coupling and correlation information of EEG signals is fully preserved, breaking through the limitation of traditional single-dimensional analysis in losing key features, and providing a data foundation for accurate identification of EEG rhythm modes;

[0063] 2. The factor matrix is ​​mapped to a special orthogonal group manifold and geometric features are extracted. The noise robustness of the manifold topology is utilized to reduce the interference of noise on feature extraction and improve the stability of EEG rhythm modality recognition in high-noise scenarios.

[0064] 3. Based on the core tensor rank and geometric features, a dynamic scheduling heterogeneous hardware computing mode is used to achieve precise matching between computing resources and EEG rhythm complexity, thereby reducing the consumption of ineffective resources while ensuring processing accuracy and balancing computing efficiency and accuracy.

[0065] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for accelerating operation based on electroencephalogram data characteristics, characterized in that, The application relates to a method for processing electroencephalogram (EEG) data, comprising the following steps: a spatial-time-frequency third-order tensor of the EEG signal is constructed to retain the coupling information of the EEG signal in the spatial channel, time sequence and frequency component dimensions; a high-order decomposition is performed on the third-order tensor to obtain a core tensor and spatial, time and frequency factor matrices, the factor matrices are mapped to a special orthogonal group manifold, the manifold curvature and geodesic distance are extracted as geometric features of the EEG rhythm mode based on a Riemann metric, the EEG rhythm mode is identified according to the geometric features, and the rank parameter of the core tensor and the geometric features are used to dynamically schedule operation resources to realize operation acceleration of the EEG data processing; the rank parameter of the core tensor and the geometric features are used to dynamically schedule operation resources by quantifying the rank of the core tensor and the geometric features to obtain an operation amount index, the operation mode of heterogeneous hardware is switched according to a threshold value of the operation amount index, and the heterogeneous hardware comprises a field programmable gate array (FPGA), a central processing unit (CPU) and a graphics processing unit (GPU); the rank of the core tensor and the geometric features are quantified to obtain the operation amount index, the operation amount is obtained by balancing the coefficient to fuse the core tensor volume and the maximum curvature of the current mode, the core tensor volume is the product of the ranks of the spatial, time and frequency dimensions, and the operation mode of the heterogeneous hardware is switched according to the threshold value of the operation amount index, that is, the lightweight fixed-point operation is started when the operation amount is lower than a first threshold value, the high-precision floating-point operation is started when the operation amount is higher than a second threshold value, and the dynamic balance of operation accuracy and resource efficiency is realized. 2.The method for accelerating operation based on electroencephalogram data features according to claim 1, characterized in that, The spatial-time-frequency third-order tensor is constructed by performing a short-time Fourier transform on the multi-channel EEG signal to generate a tensor structure comprising the channel number, time point number and frequency bin number dimensions. 3.The method of claim 1, wherein, The factor matrices are mapped to the special orthogonal group manifold by respectively mapping the spatial, time and frequency factor matrices to the special orthogonal group manifolds of the corresponding dimensions, and the topological geometric properties of the factor matrices are described by the manifolds. 4.The method of claim 1, wherein, In the process of extracting the manifold curvature and the geodesic distance as the geometric features of the EEG rhythm mode, the manifold curvature is obtained by calculating the norm of the second covariant derivative of the core tensor elements, and the geodesic distance is obtained by solving the shortest path length connecting different rhythm modes on the manifold. 5.The method for accelerating operation based on electroencephalogram data features according to claim 1, characterized in that, The multi-level high-order decomposition is performed on the EEG signal, the dynamic weight is assigned to the features in different frequency scales after the features are fused, the dynamic weight is adjusted according to the noise statistical characteristics of each frequency scale, and the decomposition layer number is preset when the multi-level high-order decomposition is performed on the EEG signal, and each layer corresponds to a different frequency scale. 6.The method for accelerating operation based on electroencephalogram data features according to claim 5, characterized in that, The multi-level high-order decomposition fusion is a multi-level high-order decomposition of the EEG signal, the features in different frequency scales are weighted through an attention mechanism, and the weight of the attention mechanism is negatively correlated with the noise variance of each frequency scale. 7.The method of claim 1, wherein, The EEG rhythm mode is identified according to the geometric features by classifying the feature vector composed of the manifold curvature and the geodesic distance through a classifier and outputting the probability distribution of each rhythm mode.

Images