Accurate electroencephalogram decoding method based on brain evoked activities and brain-computer interface system

Abstract

The invention relates to a brain induced activity-based electroencephalogram accurate decoding method and a brain-computer interface system. The brain induced activity-based electroencephalogram accurate decoding method comprises an electroencephalogram acquisition module for acquiring original electroencephalogram signals in a resting state and a motion state; an electroencephalogram preprocessing module; the brain dynamics model takes the pure resting state EEG signals and the stimulation state EEG signals processed by the electroencephalogram preprocessing module as input, the internal state of brain spontaneous activity and the internal state of brain spontaneous activity and induced activity mixture are obtained respectively, and then the internal state of brain induced activity is obtained; the internal state-EEG conversion network takes the pure signal processed by the electroencephalogram preprocessing module as input, and obtains a reconstructed EEG signal by using a coding and decoding network; the training module is used for optimizing model parameters and internal state-EEG conversion network parameters at the same time; and the decoding module is used for decoding the motion behavior according to the internal state of the brain induced activity output by the brain dynamics model. The accuracy of model parameter and internal state estimation is improved, and interference generated by spontaneous activity is eliminated.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, specifically to a precise EEG decoding method and brain-computer interface system based on brain-evoked activity. Background Technology

[0002] Brain-computer interface (BCI) is a communication interface directly established between the brain and external devices, aiming to achieve interconnection between thoughts and external objects. It is expected to play a significant role in medical rehabilitation, education, and smart homes in the future. Motor BCI based on electroencephalography (EEG) is a non-invasive BCI technology that uses movement stimulation to induce specific EEG signals in the brain. It boasts advantages such as high temporal resolution, convenient signal acquisition, high safety, and high acceptability, and is currently experiencing rapid development.

[0003] In such systems, preprocessing of EEG signals acquired during motion, combined with decoding algorithms to extract key features from evoked activity, allows for the analysis of brain consciousness and the precise mapping of "motor behavior - brain activity - motor behavior." However, the quality of the EEG signal directly determines the accuracy of decoding. EEG signals acquired during motion are typically a mixture of spontaneous and evoked brain activity. Evoked signals containing spontaneous activity backgrounds have significantly reduced signal-to-noise ratios, easily masking key evoked features. Therefore, extracting pure evoked activity from complex signals is crucial for accurate decoding of motor behavior.

[0004] Current research on extracting evoked brain activity commonly employs time-domain superposition averaging, template matching, and frequency-domain transformation. Time-domain superposition averaging aligns and averages multiple EEG data points from the same task over time, eliminating random spontaneous activity and preserving evoked activity. Template matching compares the similarity of known evoked activity templates with the actual acquired EEG signals to extract evoked activity. Frequency-domain transformation uses methods such as Fourier transform to extract evoked activity associated with specific frequency bands, suppressing non-target frequency components of spontaneous activity. However, time-domain superposition averaging relies on multiple repetitions of the task, lacks real-time performance, and has poor anti-interference capabilities; template matching relies on constructed stable templates, making it difficult to adapt to individual differences and dynamic changes; and frequency-domain transformation cannot distinguish between spontaneous and evoked activity with overlapping frequency bands. In summary, the shortcomings of these methods limit their application in motor brain-computer interfaces. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a precise EEG decoding method and brain-computer interface system based on evoked brain activity. This method utilizes a deep learning-based internal state-EEG conversion network to convert internal states into EEG signals, thereby updating the model parameters of the neurodynamic model. This improves the accuracy of model parameter and internal state estimation, enabling the neurodynamic model to effectively and in real-time separate spontaneous and evoked activities. This allows for the extraction of pure evoked brain activity, eliminating interference from spontaneous activity and laying a solid foundation for real-time and accurate decoding based on EEG data. The system integrates EEG signal acquisition, evoked brain activity extraction, and motor behavior decoding, enabling real-time EEG data acquisition and online operation, thus improving the system's temporal resolution and response speed.

[0006] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0007] In a first aspect, the present invention provides a motor brain-computer interface system based on precise decoding of brain-evoked activity, the system comprising:

[0008] The EEG acquisition module is used to acquire raw EEG signals in both resting and motor states.

[0009] The EEG preprocessing module is used to acquire pure resting-state EEG signals. r and stimulated EEG signals m ;

[0010] Brain dynamics model, using pure resting-state EEG signals processed by the EEG preprocessing module. r and stimulated EEG signals m Using the input as input, obtain the internal state x of spontaneous brain activity. s Internal states of the brain that are a mixture of spontaneous and induced activity. m According to x e =x m -x s Obtain the internal state of brain-induced activity x e ;

[0011] The internal state-EEG conversion network takes the clean EEG signal processed by the EEG preprocessing module as input and uses an encoding-decoding network to obtain the reconstructed EEG signal.

[0012] The training module calculates the EEG signal reconstruction loss and simultaneously optimizes the parameters of the neurodynamic model and the internal state-EEG transition network using a backpropagation strategy. The neurodynamic model is trained offline in stages. The first stage trains the model using clean resting-state EEG data of duration T1, and the trained model is saved as M1. The second stage trains the model using clean stimulus-state EEG data of duration T2, with training details consistent with the first stage, and the trained model is denoted as M2. M1 and M2 are placed in parallel, and their outputs do not affect each other, and are used to output x. s With x m ;

[0013] The decoding module is used to decode the internal states x of brain-evoked activity output by the brain dynamics model. e Decoding motion behavior.

[0014] The brain dynamics model is a continuous-time brain dynamics model, represented as follows:

[0015]

[0016] In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ IThese are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; and t is the sampling time.

[0017] Secondly, the present invention provides a precise EEG decoding method based on brain-evoked activity, the decoding method comprising the following steps:

[0018] 1) Collect raw EEG signals from resting and motor states, and obtain pure resting-state EEG signals from all channels of the whole brain after preprocessing. r and stimulated EEG signals m The sampling time interval is t. s ;

[0019] 2) Construct a continuous-time brain dynamics model, defined as follows:

[0020]

[0021] In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; t is time;

[0022] 3) Construct an internal state-EEG conversion network, including an encoder module and a decoder module, with input state x. i After spatiotemporal feature extraction by the encoder module and reconstruction by the decoder module, the reconstructed EEG signal is finally obtained.

[0023] 4) Utilize backpropagation strategy to simultaneously optimize the parameters of the brain dynamics model and the internal state-EEG transition network, with the loss function... Defined as:

[0024]

[0025] in, To reconstruct the EEG signal, s t The signal is the actual measured EEG signal, and η is the scaling factor. The total number of samples is T, and the total duration of EEG signal acquisition is T.

[0026] 5) The brain dynamics model adopts a phased offline training strategy: In the first phase, the model is trained with resting-state EEG data of duration T1, and the training set and test set are divided. When the loss function value is less than the specified threshold l, the training is stopped and the trained model is saved as M1. In the second phase, the model is trained with stimulated-state EEG data of duration T2. ​​The ratio of test set to training set is consistent with the first phase. The trained model is recorded as M2.

[0027] 6) Online application phase of the brain dynamics model: M1 and M2 are placed in parallel, with their outputs independent of each other. Resting-state EEG signals of duration Δt are collected during the time interval 0 to Δt to correct M1 until its output remains stable. Stimulated-state EEG signals are collected during the time interval Δt to nΔt to drive M2 and infer the internal state x of the brain, which is a mixture of spontaneous and induced activity. m Simultaneously, M1 was used to accurately infer the internal state of spontaneous brain activity. s The internal state x that induces activity in the brain e Represented as:

[0028] x e =x m -x s ;

[0029] Where n is an integer greater than 1;

[0030] 7) Internal states of the brain that induce activity x e A recurrent neural network is used to decode motion behavior.

[0031] Furthermore, at the start of training, initial values ​​for the model parameters and internal states are given based on prior constraints and neurophysiological common sense. The fourth-order Runge-Kutta method is used to solve the ordinary differential equations at each sampling interval t. s Set k time steps, and calculate the internal state x sequentially in time steps 1 to k. i (1) ... x i (k) The calculation process is as follows: First, calculate x using the known initial value. i (1) , will x i (1) x is used as the initial value for the next time step. i (2) From x i (1) After k-1 iterations, the highest accuracy x is obtained. i (k) The parameters of the brain dynamics model and the internal state-EEG conversion network are updated once after k time steps. The updated model parameters are combined with the state of the previous time step to correct the state of the next time step until the loss between the EEG reconstructed based on the internal state and the actual measured EEG is less than the specified threshold l.

[0032] Furthermore, the encoder module sequentially includes a spatial feature capture module, a temporal feature capture module, and a fully connected layer for fusing information. The spatial feature capture module consists of four cascaded 2D convolutions, used to extract spatial features between different brain region channels.

[0033] The temporal feature capture module contains three cascaded 1D convolutions to extract temporal features at different time steps;

[0034] The input state is adjusted for dimensionality and then added to the output of the spatial feature capture module via a residual connection;

[0035] The output of the temporal feature capture module is fed into a fully connected layer for feature fusion and dimensionality reduction to obtain the output of the encoder module;

[0036] The decoder module consists of a temporal reconstruction module composed of three cascaded 1D convolutions, a spatial reconstruction module composed of four cascaded 2D transposed convolutions, and a fully connected layer for adjusting the dimensions.

[0037] Furthermore, the recurrent neural network uses a two-layer gated recurrent unit (GRU) to capture temporal dependencies, and sets a temporal attention mechanism to focus on key motion temporal features. The output of the second layer GRU is introduced into the temporal attention mechanism, and the output features of the temporal attention mechanism are output through two cascaded fully connected layers to predict probabilities, thereby obtaining the EEG motion decoding results.

[0038] Thirdly, the present invention provides a method for extracting brain-induced activity, the extraction method comprising the following steps:

[0039] 1) Collect raw EEG signals from resting and motor states, and obtain pure resting-state EEG signals from all channels of the whole brain after preprocessing. r and stimulated EEG signals m The sampling time interval is t. s ;

[0040] 2) Construct a continuous-time brain dynamics model, defined as follows:

[0041]

[0042] In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; t is time;

[0043] 3) Construct an internal state-EEG conversion network, including an encoder module and a decoder module, with input state x. i After spatiotemporal feature extraction by the encoder module and reconstruction by the decoder module, the reconstructed EEG signal is finally obtained.

[0044] 4) Utilize backpropagation strategy to simultaneously optimize the parameters of the brain dynamics model and the internal state-EEG transition network, with the loss function... Defined as:

[0045]

[0046] in, To reconstruct the EEG signal, s t The signal is the actual measured EEG signal, and η is the scaling factor. The total number of samples is T, and the total duration of EEG signal acquisition is T.

[0047] 5) The brain dynamics model adopts a phased offline training strategy: In the first phase, the model is trained with resting-state EEG data of duration T1, and the training set and test set are divided. When the loss function value is less than the specified threshold l, the training is stopped and the trained model is saved as M1. In the second phase, the model is trained with stimulated-state EEG data of duration T2. ​​The ratio of test set to training set is consistent with the first phase. The trained model is recorded as M2.

[0048] 6) Online application phase of the brain dynamics model: M1 and M2 are placed in parallel, with their outputs independent of each other. Resting-state EEG signals of duration Δt are collected during the time interval 0 to Δt to correct M1 until its output remains stable. Stimulated-state EEG signals are collected during the time interval Δt to nΔt to drive M2 and infer the internal state x of the brain, which is a mixture of spontaneous and induced activity. m Simultaneously, M1 was used to accurately infer the internal state of spontaneous brain activity. s The internal state x that induces activity in the brain e Represented as:

[0049] x e =x m -x s ;

[0050] Where n is an integer greater than 1.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] (1) The method of the present invention overcomes the limitations of existing induced activity extraction methods such as time domain superposition averaging, template matching and frequency domain transformation, which cannot separate induced components in real time, have poor individual adaptability and incomplete separation of induced components.

[0053] (2) The continuous-time brain dynamics model in this invention considers the connectivity between brain regions, uses ordinary differential equations to replace the discrete equations of traditional models, and replaces the fixed parameters in traditional models with excitatory decay rates and inhibitory decay rates that dynamically change with time and internal state variables. This improves the accuracy and temporal resolution of the obtained internal states, enhances the response speed and robustness in motion scenarios, and uses an internal state-EEG conversion network that integrates deep learning to convert internal states into EEG signals. This network, leveraging the powerful feature extraction and reconstruction capabilities of deep learning, can achieve higher accuracy conversion than traditional observation matrices, thereby completing the dual update and optimization of model parameters and internal states. Compared with traditional system identification methods such as Kalman filters, this deep learning-integrated model parameter update method has a faster convergence speed and a lower loss function value, improving the accuracy of model parameter and internal state estimation. Furthermore, compared with the neural population model (NMM) commonly used in brain modeling, the parameters in the continuous-time brain dynamics model of this invention can dynamically change according to the differences in brain activity among different individuals, avoiding the problem of poor individual adaptability of general model parameters.

[0054] (3) Compared with directly decoding motor behavior from EEG signals using various deep learning methods, constructing a brain dynamics model can significantly reduce the number of parameters and computational cost. Furthermore, the parameters in the model have real physiological meaning, and the internal states reflect the relative levels of excitability and inhibition at the mesoscopic level. This makes it more interpretable than the black-box networks of deep learning, as it requires no explicit input and can automatically update and iterate given initial values. The internal states in the brain dynamics model contain richer features than the original EEG signals, and using internal states for motor behavior decoding can improve accuracy.

[0055] (4) The trained continuous-time brain dynamics model in this invention can be further used as a "virtual brain" platform to generate realistic internal brain states under unsupervised conditions and drive peripheral devices to work. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working process of a brain-computer interface system that precisely decodes brain-evoked activities.

[0057] Figure 2 This is a schematic diagram of one embodiment of an internal state-EEG transformation network. Detailed Implementation

[0058] The present invention will be further explained below with reference to the embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0059] Example 1

[0060] This invention is based on a precise EEG decoding method for brain-evoked activity, comprising the following steps:

[0061] 1) Raw EEG signals from both resting and motor states were acquired, and these two types of signals were preprocessed by filtering, baseline correction, downsampling, and artifact removal to obtain a clean 64-channel whole-brain EEG signal. r and stimulated EEG signals m In this embodiment, the sampling interval is t. s ;

[0062] 2) Combining brain connectivity maps, a brain dynamics model based on the Wilson-Cowan equations is constructed. The discretized equations are then transformed into differential forms using continuous-time ordinary differential equations, forming a continuous-time brain dynamics model. The model is defined as follows:

[0063]

[0064] In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) This represents the net synaptic input to the inhibitory neurons in the i-th brain region; and β represents the excitatory connectivity matrix and inhibitory connectivity matrix between different brain regions i and j, respectively, including local connectivity (diagonal terms) and long-range connectivity (off-diagonal terms); i E and c represents the local excitation-inhibition connectivity and the local inhibition-inhibition connectivity of the i-th brain region, respectively; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I ψ represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; E and ψ I For nonlinear activation functions, the tanh function is used; s E and s I These are excitatory gain and inhibitory gain, respectively; v i I For the inhibitory bias of the i-th brain region, v jE D represents the excitability bias of the j-th brain region; i E and D i I These are the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, and are closer to the real state. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficients and inhibitory state dependence coefficients, respectively; σ is the sigmoid function;

[0065] The updating and optimization of the continuous-time brain dynamics model includes two aspects: first, identifying unknown model parameters (c i E and c i I ε i E and ε i I , and s E and s I v E and v I γ E and γ I (etc.), and secondly, to realize the internal state variable x. i =(x i E x i I The optimal estimate of ).

[0066] 3) Construct an internal state-EEG conversion network, which includes an encoder module and a decoder module. The encoder module sequentially includes a spatial feature capture module, a temporal feature capture module, and a fully connected layer for fusing information. The spatial feature capture module consists of four cascaded 2D convolutions, which are used to extract spatial features between different brain region channels.

[0067] The temporal feature capture module contains three cascaded 1D convolutions to extract temporal features at different time steps;

[0068] The input state is adjusted by dimension and then added to the output of the spatial feature capture module through a residual connection. The residual connection can preserve the original features and avoid gradient vanishing.

[0069] The output of the temporal feature capture module is fed into a fully connected layer for feature fusion and dimensionality reduction to obtain the output of the encoder module;

[0070] The decoder module consists of a temporal reconstruction module composed of three cascaded 1D convolutions, a spatial reconstruction module composed of four cascaded 2D transposed convolutions, and a fully connected layer for adjusting the dimensions.

[0071] Input state x i After spatiotemporal feature extraction by the encoder module and reconstruction by the decoder module, the reconstructed EEG signal is finally obtained.

[0072]

[0073] 4) Define the EEG signal reconstruction loss function, combine MSE and MAE to balance local and global errors, and use the backpropagation strategy to simultaneously optimize the brain dynamics model parameters and the internal state-EEG transformation network parameters. The loss function... The definition is as follows:

[0074]

[0075] in, To reconstruct the EEG signal, s t The processed s signal is the actual measured EEG signal (i.e., step 1). r or s m ), where η is the scaling factor. The total number of samples is T, the total duration of EEG signal acquisition is T, and t=1 indicates the first sample.

[0076] At the start of training, initial values ​​for the model parameters and internal states are given based on prior constraints and neurophysiological common sense. The fourth-order Runge-Kutta method is used to solve the ordinary differential equations at each sampling interval t. s Set k time steps, the specific value of k is related to the sampling interval t s Regarding this, the internal state x is calculated sequentially at time steps 1 to k. i (1) ... x i (k) The specific calculation process is as follows: First, calculate x using the known initial value. i (1) , will x i (1) x is used as the initial value for the next time step. i (2) From x i (1) After k-1 iterations, the highest accuracy x is obtained. i (k)The parameters of the brain dynamics model and the internal state-EEG conversion network are updated once after k time steps. The updated model parameters are combined with the state of the previous time step to correct the state of the next time step until the loss between the EEG reconstructed based on the internal state and the actual measured EEG is less than the specified threshold l.

[0077] 5) The brain dynamics model adopts a phased offline training strategy: In the first phase, the model is trained using resting-state EEG data with a duration of T1. The first 0.8T1 duration is used as the training set, and the last 0.2T1 duration is used as the test set. Training is stopped when the loss function value is less than the specified threshold l, and the trained model is saved as M1. In the second phase, the model is trained using stimulated-state EEG data with a duration of T2. The ratio of test set to training set is consistent with that in the first phase, and the trained model is recorded as M2.

[0078] 6) Online application phase of the brain dynamics model: M1 and M2 are placed in parallel, with their outputs independent of each other. Resting-state EEG signals of duration Δt are collected during the time period from 0 to Δt to correct M1 until its output remains stable. Stimulated-state EEG signals are collected during the time period from Δt to nΔt (n is an integer greater than 1) to drive M2 to infer the internal state of the brain, which is a mixture of spontaneous and induced activity. Simultaneously, M1 is used to accurately infer the internal state of spontaneous brain activity. The internal state of the brain that induces activity It can be represented as follows:

[0079] x e =x m -x s

[0080] 8) Internal states of the brain that induce activity x e The motion behavior is decoded using a recurrent neural network (RNN).

[0081] In this embodiment, the RNN sequentially employs two layers of gated recurrent units (GRUs) to capture temporal dependencies, and a temporal attention mechanism to focus on key motion temporal features. The first GRU layer has 64 hidden units, matching the input state x. e The number of channels is [number missing]; the second GRU layer has 32 hidden units to achieve feature dimensionality reduction and focus on core information. Both GRU layers use the tanh activation function, with dropout set to 0.2, randomly discarding 20% ​​of neurons to prevent overfitting. The output of the second GRU layer [is described]. A temporal attention mechanism is introduced. The output features of the temporal attention mechanism are passed through two cascaded fully connected layers to output the predicted probability, thereby obtaining the EEG motion decoding results.

[0082] In the temporal attention mechanism, according to α m =softmax(Wa ·H2(:,m)+b a ) Calculate each sampling time Attention weight α m H2(:,m) represents the column vector consisting of all elements in the m-th column of matrix H2; W a ∈R 1×32 For gain, b a For bias terms;

[0083] Then according to H attn (:,m)=α m H2(:,m) obtains the temporal attention output that fuses temporal features.

[0084] Among them, H attn (:,m) represents matrix H attn The column vector consisting of all elements in the m-th column.

[0085] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A motor brain-computer interface system based on precise decoding of brain-evoked activities, characterized in that, The system includes: The EEG acquisition module is used to acquire raw EEG signals in both resting and motor states. The EEG preprocessing module is used to acquire pure resting-state EEG signals. r and stimulated EEG signals m ; Brain dynamics model, using pure resting-state EEG signals processed by the EEG preprocessing module. r and stimulated EEG signals m Using the input as input, obtain the internal state x of spontaneous brain activity. s Internal states of the brain that are a mixture of spontaneous and induced activity. m According to x e =x m -x s Obtain the internal state of brain-induced activity x e ; The internal state-EEG conversion network takes the clean EEG signal processed by the EEG preprocessing module as input and uses an encoding-decoding network to obtain the reconstructed EEG signal. The training module calculates the EEG signal reconstruction loss and simultaneously optimizes the parameters of the neurodynamic model and the internal state-EEG transition network using a backpropagation strategy. The neurodynamic model is trained offline in stages. The first stage trains the model using clean resting-state EEG data of duration T1, and the trained model is saved as M1. The second stage trains the model using clean stimulus-state EEG data of duration T2, with training details consistent with the first stage, and the trained model is denoted as M2. M1 and M2 are placed in parallel, and their outputs do not affect each other, and are used to output x. s With x m ; The decoding module is used to decode the internal states x of brain-evoked activity output by the brain dynamics model. e Decoding motion behavior.

2. The system according to claim 1, characterized in that, The brain dynamics model is a continuous-time brain dynamics model, represented as follows: In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; and t is the sampling time.

3. A precise EEG decoding method based on brain-evoked activity, characterized in that, The decoding method includes the following steps: 1) Collect raw EEG signals from resting and motor states, and obtain pure resting-state EEG signals from all channels of the whole brain after preprocessing. r and stimulated EEG signals m The sampling time interval is t. s ; 2) Construct a continuous-time brain dynamics model, defined as follows: In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; t is time; 3) Construct an internal state-EEG conversion network, including an encoder module and a decoder module, with input state x. i After spatiotemporal feature extraction by the encoder module and reconstruction by the decoder module, the reconstructed EEG signal is finally obtained. 4) Utilize backpropagation strategy to simultaneously optimize the parameters of the brain dynamics model and the internal state-EEG transition network, with the loss function... Defined as: in, To reconstruct the EEG signal, s t The signal is the actual measured EEG signal, and η is the scaling factor. The total number of samples is T, and the total duration of EEG signal acquisition is T. 5) The brain dynamics model adopts a phased offline training strategy: In the first phase, the model is trained with resting-state EEG data of duration T1, and the training set and test set are divided. When the loss function value is less than the specified threshold l, the training is stopped and the trained model is saved as M1. In the second phase, the model is trained with stimulated-state EEG data of duration T2. ​​The ratio of test set to training set is consistent with the first phase. The trained model is recorded as M2. 6) Online application phase of the brain dynamics model: M1 and M2 are placed in parallel, with their outputs independent of each other. Resting-state EEG signals of duration Δt are collected during the time interval 0 to Δt to correct M1 until its output remains stable. Stimulated-state EEG signals are collected during the time interval Δt to nΔt to drive M2 and infer the internal state x of the brain, which is a mixture of spontaneous and induced activity. m Simultaneously, M1 was used to accurately infer the internal state of spontaneous brain activity. s The internal state x that induces activity in the brain e Represented as: x e =x m -x s ; Where n is an integer greater than 1; 7) Internal states of the brain that induce activity x e A recurrent neural network is used to decode motion behavior.

4. The decoding method according to claim 3, characterized in that, At the start of training, initial values ​​for the model parameters and internal states are given based on prior constraints and neurophysiological common sense. The fourth-order Runge-Kutta method is used to solve the ordinary differential equations at each sampling interval t. s Set k time steps, and calculate the internal state x sequentially in time steps 1 to k. i (1) ... x i (k) The calculation process is as follows: First, calculate x using the known initial value. i (1) , will x i (1) x is used as the initial value for the next time step. i (2) From x i (1) After k-1 iterations, the highest accuracy x is obtained. i (k) The parameters of the brain dynamics model and the internal state-EEG conversion network are updated once after k time steps. The updated model parameters are combined with the state of the previous time step to correct the state of the next time step until the loss between the EEG reconstructed based on the internal state and the actual measured EEG is less than the specified threshold l.

5. The decoding method according to claim 3, characterized in that, The encoder module sequentially includes a spatial feature capture module, a temporal feature capture module, and a fully connected layer for information fusion. The spatial feature capture module consists of four cascaded 2D convolutions, used to extract spatial features between different brain region channels. The temporal feature capture module contains three cascaded 1D convolutions to extract temporal features at different time steps; The input state is adjusted for dimensionality and then added to the output of the spatial feature capture module via a residual connection; The output of the temporal feature capture module is fed into a fully connected layer for feature fusion and dimensionality reduction to obtain the output of the encoder module; The decoder module consists of a temporal reconstruction module composed of three cascaded 1D convolutions, a spatial reconstruction module composed of four cascaded 2D transposed convolutions, and a fully connected layer for adjusting the dimensions.

6. The decoding method according to claim 3, characterized in that, The recurrent neural network uses a two-layer gated recurrent unit (GRU) to capture temporal dependencies, and sets a temporal attention mechanism to focus on key motion temporal features. The output of the second layer GRU is introduced into the temporal attention mechanism. The output features of the temporal attention mechanism are output through two cascaded fully connected layers to predict probabilities, thereby obtaining the EEG motion decoding results.

7. A method for extracting brain-evoked activity, characterized in that, The extraction method includes the following steps: 1) Acquiring raw EEG signals from resting and motor states, and preprocessing them to obtain pure resting-state EEG signals from all channels of the whole brain. r and stimulated EEG signals m The sampling time interval is t. s ; 2) Construct a continuous-time brain dynamics model, defined as follows: In this context, the subscripts i and j represent different brain regions; x i E and x i I These are the excitatory / inhibitory state variables of the i-th brain region, which change over time; p (i) r is the net synaptic input to the excitatory neuron in the i-th brain region; (i) c is the net synaptic input to the inhibitory neuron in the i-th brain region; i E and c i I ε represents the continuous stimulation of excitatory and inhibitory neuronal populations in the i-th brain region, respectively; i E and ε i I D represents the random noise generated by the activity of excitatory and inhibitory neurons in the i-th brain region, respectively; i E and D i I These represent the excitatory decay rate and inhibitory decay rate in the i-th brain region, which change dynamically with the state variable, respectively. and These are the excitatory basal decay rate and the inhibitory basal decay rate, respectively; α E and α I These are the excitatory regulation strength coefficient and the inhibitory regulation strength coefficient, respectively; γ E and γ I These are the excitatory state dependence coefficient and the inhibitory state dependence coefficient, respectively; σ is the sigmoid function; t is time; 4) Construct an internal state-EEG conversion network, including an encoder module and a decoder module, with input state x. i After spatiotemporal feature extraction by the encoder module and reconstruction by the decoder module, the reconstructed EEG signal is finally obtained. 4) Utilize backpropagation strategy to simultaneously optimize the parameters of the brain dynamics model and the internal state-EEG transition network, with the loss function... Defined as: in, To reconstruct the EEG signal, s t The signal is the actual measured EEG signal, and η is the scaling factor. The total number of samples is T, and the total duration of EEG signal acquisition is T. 5) The brain dynamics model adopts a phased offline training strategy: In the first phase, the model is trained with resting-state EEG data of duration T1, and the training set and test set are divided. When the loss function value is less than the specified threshold l, the training is stopped and the trained model is saved as M1. In the second phase, the model is trained with stimulated-state EEG data of duration T2. ​​The ratio of test set to training set is consistent with the first phase. The trained model is recorded as M2. 6) Online application phase of the brain dynamics model: M1 and M2 are placed in parallel, with their outputs independent of each other. Resting-state EEG signals of duration Δt are collected during the time interval 0 to Δt to correct M1 until its output remains stable. Stimulated-state EEG signals are collected during the time interval Δt to nΔt to drive M2 and infer the internal state x of the brain, which is a mixture of spontaneous and induced activity. m Simultaneously, M1 was used to accurately infer the internal state of spontaneous brain activity. s The internal state x that induces activity in the brain e Represented as: x e =x m -x s ; Where n is an integer greater than 1.

Images