A source-space-based real-time feedback modulation method and system for decoding electroencephalogram (EEG) signals

Abstract

This invention discloses a real-time feedback control method and system for EEG signal decoding based on source space. The acquired EEG data is used as input to an EEG signal decoding network, which consists of a trainable source space transformation layer and a task decoding model. The EEG signal decoding network outputs the task prediction result corresponding to the current EEG data and calculates the task loss based on the task prediction result and task label. Simultaneously, the brain source signals generated by the trainable source space transformation layer are acquired in real time, and source space evaluation, visualization, and source space constraint loss calculation are performed based on the brain source signals. Subsequently, the source space evaluation indicators, automatic feedback control suggestions, and the visualization results of the brain source signals are output to an expert review end. Based on the expert-assisted review results, a target feedback control strategy is determined, thereby achieving real-time feedback control of the EEG signal decoding network and improving the accuracy and stability of the EEG signal decoding results.

Description

Technical Field

[0001] This invention relates to the technical field of brain power imaging and brain signal decoding, and in particular to a real-time feedback control method and system for brain signal decoding based on source space. Background Technology

[0002] Electroencephalogram (EEG) signals, characterized by high temporal resolution, convenient acquisition, and relatively low equipment cost, have been widely applied in tasks such as brain-computer interfaces, motor imagery recognition, emotion recognition, and cognitive state assessment. However, scalp EEG signals are affected by factors such as volume conduction effects, noise interference, individual differences, and limited electrode spatial resolution. EEG signals in the sensor space cannot be directly equated to the source of neural activity in the brain. If decoding is performed directly based on EEG signals in the sensor space, the model is prone to relying on statistical differences or local channel responses at the electrode level, making it difficult to accurately reflect the source spatial distribution of brain activity, thus affecting the accuracy and interpretability of the decoding results.

[0003] To improve the spatial interpretability of EEG decoding models, existing methods typically first construct a source imaging model using information such as head model, lead field matrix, simulated source activity, or prior information about brain regions. Then, the trained source imaging model is used to obtain the brain source signal corresponding to the real EEG signal. Finally, the brain source signal is input into the downstream decoding model for task learning.

[0004] Compared to decoding directly based on sensor spatial EEG signals, the above method can incorporate source spatial information to some extent, but it still has the following technical drawbacks in practical applications:

[0005] 1. Existing methods typically employ a two-stage process where the source imaging model and the downstream decoding model are trained separately. The source imaging process primarily focuses on source signal reconstruction, while the downstream decoding process mainly focuses on classification or regression performance. Because the optimization objectives of the two stages are inconsistent, and the task loss is difficult to reverse-engineer the source space mapping process, the resulting brain source signals may not be suitable for the specific decoding task, thus limiting the final decoding performance.

[0006] 2. Existing source space modeling methods mostly rely on simulation data, template head models, or fixed priors. However, real EEG data is affected by factors such as individual differences, acquisition conditions, and task status, which may lead to deviations from simulation data or prior models. This results in a mismatch between the source space estimation results and the spatial distribution characteristics of real EEG data, affecting the reliability of brain source signal expression.

[0007] 3. Existing methods typically use brain source signals as offline intermediate results, lacking process constraints and adjustment criteria for the spatial distribution of brain source signals during training. When brain source signals exhibit conditions such as overactivation of non-task-related source spatial locations, diffusion of source spatial activity, or insufficient representation of task-related source spatial locations, the model struggles to make timely and targeted adjustments, resulting in insufficient coordination between the brain source signal generation process and the task decoding objective. Summary of the Invention

[0008] The primary objective of this invention is to provide a real-time feedback control method for decoding EEG signals based on source space, in order to solve problems such as the separation of the source imaging process from the downstream decoding task in existing EEG signal decoding methods, the difficulty in adaptively adjusting the source space mapping process according to specific decoding tasks, and the lack of a regulation mechanism for the spatial state of brain source signals during training.

[0009] The second objective of this invention is to provide a real-time feedback control system for decoding electroencephalogram (EEG) signals based on source space.

[0010] The first objective of this invention is achieved through the following technical solution: a real-time feedback modulation method for decoding electroencephalogram (EEG) signals based on source space, comprising the following steps:

[0011] S1: Construct an EEG signal decoding network consisting of a trainable source space transformation layer and a task decoding model. The trainable source space transformation layer is used to map EEG data from the sensor space to the brain source space, and the task decoding model is used to receive the brain source signals output by the trainable source space transformation layer and perform feature extraction and task discrimination on the brain source signals.

[0012] S2: Input the EEG data into the EEG signal decoding network to obtain the task prediction result corresponding to the current EEG data, and calculate the task loss based on the task prediction result and task label;

[0013] S3: Real-time acquisition of brain source signals output by the trainable source space transformation layer in the EEG signal decoding network; calculation of source space evaluation index based on brain source signals; generation of visualization results of brain source signals at preset source space locations; construction of source space constraint loss that can participate in training optimization based on source space evaluation index;

[0014] S4: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules, output source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals to the expert review end, and determine the target feedback control strategy based on the expert review results;

[0015] S5: Using task loss and source space constraint loss as the basis for training optimization, and target feedback regulation strategy as the basis for real-time regulation, the target regulation object and corresponding regulation parameter in the EEG signal decoding network are determined, and the EEG signal decoding network is updated according to the target regulation object and corresponding regulation parameter, so that the brain source signal generation process and task decoding process are adjusted according to the target feedback regulation strategy, thereby realizing real-time feedback regulation.

[0016] Furthermore, the specific steps of step S1 are as follows:

[0017] S1.1: Construct a trainable source space transformation layer. The trainable source space transformation layer sets a transformation matrix. The transformation matrix is ​​used to map EEG data from sensor space to brain source space to generate brain source signal. Each signal component in the brain source signal corresponds to a preset source space position. The preset source space position includes one or more of brain regions, cortical grid points and source space nodes.

[0018] S1.2: Initialize the transformation matrix so that it has an initial mapping relationship from sensor space to brain source space at the start of training. The initialization method includes one or more of the following: based on the lead field matrix, traditional source imaging results, template head model, brain region prior, and random initialization. Among them, the initialization based on the lead field matrix, traditional source imaging results, template head model, and brain region prior is used to make the transformation matrix have source space mapping prior, and random initialization is used to set the initial parameters of the transformation matrix when source space prior information is lacking.

[0019] S1.3: Construct a task decoding model and connect it to a trainable source spatial transformation layer. This enables the task decoding model to receive brain source signals output by the trainable source spatial transformation layer and perform feature extraction and task discrimination on the brain source signals. The task decoding model includes a spatial feature extraction unit, a temporal feature extraction unit, and a task discrimination unit. The spatial feature extraction unit is used to extract the spatial distribution features of brain source signals between different preset source spatial locations. The temporal feature extraction unit is used to extract the temporal dynamic features of brain source signals between different time segments. The task discrimination unit is used to obtain the task prediction result corresponding to the current EEG data based on the spatial distribution features and temporal dynamic features. In the task discrimination unit, the spatial distribution features and temporal dynamic features are first fused to obtain a joint feature vector. Then, the joint feature vector is input into at least one fully connected layer and processed by the linear transformation and activation function of the fully connected layer to map it into a discrimination score equal to the number of preset task categories. Finally, the discrimination score is normalized to a probability distribution by the Softmax function, and the task category corresponding to the maximum probability is taken as the task prediction result corresponding to the current EEG data.

[0020] S1.4: The trainable source space transformation layer and the task decoding model are combined into an EEG signal decoding network, so that the brain source signal generation process and the task decoding process are within the same training framework, and the task loss generated by the task decoding model can be applied to the trainable source space transformation layer through backpropagation.

[0021] Furthermore, the specific steps of step S2 are as follows:

[0022] S2.1: Input the EEG data into the EEG signal decoding network, so that the EEG data first passes through the trainable source space transformation layer to generate brain source signals, and then the task decoding model performs feature extraction and task discrimination on the brain source signals to obtain the task prediction result corresponding to the current EEG data;

[0023] S2.2: Calculate the task loss based on the task prediction result and the task label, so that the task loss can characterize the difference between the task prediction result and the task label; wherein, the task label is provided by the training data or the label input.

[0024] Furthermore, the specific steps of step S3 are as follows:

[0025] S3.1: Real-time acquisition of brain source signals output by the trainable source space transformation layer in the EEG signal decoding network, making brain source signals the common basis for source space evaluation, source space visualization and source space constraint loss construction;

[0026] S3.2: Calculate source space evaluation indicators based on real-time acquired brain source signals. Source space evaluation indicators are used to quantify the source space state of the current brain source signals, including one or more of the following: activity intensity of task-related source space locations, activity proportion of non-task-related source space locations, source space diffusion degree, spatial smoothness, temporal continuity, forward reconstruction consistency, and category-related source region differences.

[0027] S3.3: Generate visualization results based on real-time acquired brain source signals, so that the visualization results show the activity distribution of the current EEG data at preset source spatial locations. When generating visualization results, the activity values ​​corresponding to each preset source spatial location in the brain source signal are mapped to the corresponding brain regions, cortical grid points or source spatial nodes to form a source spatial activity distribution that can be displayed by the expert review end.

[0028] S3.4: Construct source space constraint loss that can participate in training optimization based on source space evaluation index, so that source space constraint loss can constrain the source space representation of brain source signals; wherein, source space constraint loss is constructed based on source space evaluation index and its corresponding differentiable form, including one or more of the following: task-related source space location constraint loss, non-task-related source space location inhibition loss, source space sparsity loss, spatial smoothing loss, temporal continuity loss, and forward reconstruction consistency loss.

[0029] Furthermore, the specific steps of step S4 are as follows:

[0030] S4.1: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules. The preset feedback control rules include triggering conditions for the intensity of activity in task-related source space locations, the proportion of activity in non-task-related source space locations, the degree of source space diffusion, temporal continuity, and forward reconstruction consistency. When the source space evaluation indicators reach or exceed the corresponding triggering conditions, automatic feedback control suggestions are generated. Automatic feedback control suggestions include control targets, control directions, and suggested control strengths. Control targets are used to indicate the source space mapping relationship, source space mask, loss weight, or training parameters that need to be adjusted. Control directions are used to indicate the direction of control, such as enhancement, suppression, expansion, contraction, improvement, or reduction. Suggested control strengths are used to indicate the magnitude of the corresponding control action.

[0031] S4.2: Output the source space evaluation indicators, automatic feedback regulation suggestions and visualization results of brain source signals to the expert review end, and receive the expert auxiliary review results, so that the automatic feedback regulation suggestions can be reviewed based on the visualization results and quantitative evaluation results;

[0032] S4.3: Determine the target feedback control strategy based on the expert-assisted review results. When the expert-assisted review results indicate that the automatic feedback control recommendations are applicable, the automatic feedback control recommendations will be determined as the target feedback control strategy. When the expert-assisted review results indicate that the automatic feedback control recommendations need to be adjusted, the control targets, control directions, and control intensity will be modified based on the expert-assisted review results, and the modified results will be determined as the target feedback control strategy.

[0033] Furthermore, the specific steps of step S5 are as follows:

[0034] S5.1: Using task loss and source space constraint loss as the basis for training optimization, and target feedback regulation strategy as the basis for real-time regulation, the target regulation object in the EEG signal decoding network is determined. The target regulation object is an adjustable object that can affect the brain source signal generation process, source space constraint process, and task decoding process, including one or more of the following: transformation matrix, source space mask, loss weight, and task decoding model training parameters. Among them, the transformation matrix is ​​used to adjust the mapping relationship from sensor space to brain source space, the source space mask is used to limit the preset source space location range for participating in source space evaluation, source space constraint, or source space visualization, the loss weight is used to adjust the strength of the role of task loss and source space constraint loss in the training process, and the task decoding model training parameters are used to adjust the parameter update process of the task decoding model.

[0035] S5.2: Determine the corresponding control parameters based on the target control object, so that the corresponding control parameters can limit the adjustment direction and adjustment range of the target control object; wherein, the control parameters corresponding to the transformation matrix include the update direction, update range, and adjustment coefficients of the parameters corresponding to the specified preset source space position; the control parameters corresponding to the source space mask include the mask range, mask weight, and mask update range; the control parameters corresponding to the loss weight include the task loss weight, source space constraint loss weight, and weights of each source space constraint sub-item; and the control parameters corresponding to the task decoding model training parameters include the learning rate, regularization coefficient, and parameter update range.

[0036] S5.3: The EEG signal decoding network is updated according to the target control object and its corresponding control parameters, so that the target control object adjusts according to the target feedback control strategy, thereby regulating the brain source signal generation process and the task decoding process, and realizing real-time feedback control.

[0037] The second objective of this invention is achieved through the following technical solution: a source space-based EEG signal decoding real-time feedback control system, used to implement the above-mentioned source space-based EEG signal decoding real-time feedback control method, comprising a source space conversion module, a brain source signal output module, a task decoding module, a source space evaluation module, a source space visualization module, an expert-assisted review module, and a feedback control module;

[0038] The source space conversion module is used to receive EEG data and map the EEG data from the sensor space to the brain source space through a trainable source space conversion layer to generate brain source signals. The source space conversion module is equipped with a conversion matrix, which is used to establish a mapping relationship between the EEG acquisition channel and the preset source space position, so that each signal component in the generated brain source signal corresponds to the preset source space position.

[0039] The brain source signal output module is connected to the source space conversion module and is used to acquire the brain source signal output by the source space conversion module in real time, and synchronously output the brain source signal to the task decoding module, the source space evaluation module and the source space visualization module.

[0040] The task decoding module is connected to the brain source signal output module and the feedback control module respectively. It receives the brain source signal output by the brain source signal output module through the task decoding model, performs feature extraction and task discrimination on the brain source signal, obtains the task prediction result corresponding to the current EEG data, calculates the task loss based on the task prediction result and the task label, and outputs the task loss to the feedback control module; wherein, the task label is provided by the training data or the label input terminal.

[0041] The source space evaluation module is connected to the brain source signal output module, the expert-assisted review module, and the feedback control module, respectively. It is used to calculate the source space evaluation index, construct a source space constraint loss that can participate in training optimization based on the source space evaluation index, and generate automatic feedback control suggestions based on the source space evaluation index and preset feedback control rules. Then, the source space evaluation index and the automatic feedback control suggestions are output to the expert-assisted review module, and the source space constraint loss is output to the feedback control module. Among them, the source space evaluation index is used to quantify the source space state of the current brain source signal, the source space constraint loss is used to constrain the source space expression of the brain source signal, and the automatic feedback control suggestions are used to indicate the source space mapping relationship, source space mask, loss weight, or task decoding model training parameters to be adjusted.

[0042] The source space visualization module is connected to the brain source signal output module and the expert-assisted review module respectively. It is used to generate visualization results of the brain source signals output by the brain source signal output module, so that the visualization results show the activity distribution of the current EEG data at the preset source space location, and output the visualization results to the expert-assisted review module.

[0043] The expert-assisted review module is used to receive source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals, output source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals to the expert review terminal, receive the expert-assisted review results returned by the expert review terminal, determine the target feedback control strategy based on the expert-assisted review results, and output the target feedback control strategy to the feedback control module.

[0044] The feedback control module is connected to the task decoding module, the source space evaluation module, and the expert-assisted review module, respectively. It receives the task loss output by the task decoding module, the source space constraint loss output by the source space evaluation module, and the target feedback control strategy output by the expert-assisted review module. Using the task loss and source space constraint loss as the basis for training optimization and the target feedback control strategy as the basis for real-time control, it determines the target control object and its corresponding control parameters in the EEG signal decoding network. Then, it updates the EEG signal decoding network according to the target control object and its corresponding control parameters, so that the brain source signal generation process and the task decoding process are adjusted according to the target feedback control strategy, thereby realizing real-time feedback control.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. This invention enables the brain source signal generation process and the task decoding process to be jointly optimized within the same training framework, avoiding the problem of inconsistent optimization objectives caused by the separate training of the source imaging model and the downstream decoding model in existing methods. This allows the source space mapping relationship to be adaptively adjusted according to the specific decoding task, thereby improving the adaptability between brain source signals and task requirements.

[0047] 2. This invention introduces source space evaluation, source space constraint and feedback regulation during the training process, which can dynamically adjust the spatial representation of brain source signals. This is beneficial to enhance the representation of task-related source regions, suppress non-task-related source space activities, reduce the impact of abnormal activation or excessive diffusion of source space on decoding results, and thus improve the accuracy and stability of EEG signal decoding results.

[0048] 3. This invention can output the source spatial activity distribution of brain source signals during training, so that the network prediction results can correspond to the activity state on brain regions, cortical grid points or source spatial nodes, thereby improving the spatial interpretability of EEG signal decoding networks.

[0049] 4. This invention combines automatic feedback control suggestions with expert-assisted review results, and can adopt different source space initialization methods and task decoding model structures according to actual data conditions, which can improve the reliability of the source space control process and enhance its applicability in different EEG decoding tasks and data scenarios. Attached Figure Description

[0050] Figure 1 is a flowchart of the overall process of the method of the present invention.

[0051] Figure 2 is a schematic diagram of the module structure of the system of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0053] Example 1

[0054] This embodiment discloses a real-time feedback control method for decoding EEG signals based on source space, which is used to perform real-time feedback control on the source space mapping process and task decoding process of multi-channel EEG signals in motor imagery brain-computer interface tasks.

[0055] Participants performed left-hand and right-hand motor imagery tasks based on visual cues. EEG acquisition devices simultaneously collected multi-channel EEG data, associating each data segment with a corresponding task label. The acquired EEG data served as input to an EEG signal decoding network, which output the task prediction result corresponding to the current EEG data and calculated the task loss based on the prediction result and task label. Simultaneously, brain-source signals generated by a trainable source space transformation layer were acquired in real-time. Based on these signals, source space evaluation, visualization, and source space constraint loss calculation were performed. Subsequently, source space evaluation indicators, automatic feedback control suggestions, and the visualized results of the brain-source signals were output to an expert review panel. Based on the expert review results, a target feedback control strategy was determined, thereby achieving real-time feedback control of the EEG signal decoding network. Figure 1 As shown, the specific steps are as follows:

[0056] S1: Construct an EEG signal decoding network consisting of a trainable source space transformation layer and a task decoding model. The trainable source space transformation layer maps EEG data from sensor space to brain source space. The task decoding model is connected to the trainable source space transformation layer and receives the brain source signals output by the trainable source space transformation layer, and performs feature extraction and task discrimination on the brain source signals; the specific details are as follows:

[0057] S1.1: Construct a trainable source space transformation layer. This layer includes a transformation matrix that maps the multi-channel EEG data corresponding to left-hand and right-hand motor imagery from the sensor space to the brain-source space, thereby generating brain-source signals. Let the input EEG data be... , where the symbol The symbol "belongs to" indicates that the variable preceding the symbol belongs to a set following it. Represents the set of real numbers. Indicates the number of EEG acquisition channels. This indicates the number of time sampling points; the brain-source signal is... , This represents the number of preset source space locations, and the transformation matrix is: The generation process of brain-source signals can be represented as follows: Each signal component in the brain source signal corresponds to a preset source spatial location, which includes one or more brain regions, cortical grid points, and source spatial nodes related to the motor imagery task.

[0058] S1.2: Initialize the transformation matrix to ensure it has an initial mapping from sensor space to brain source space at the start of training. Initialization methods include one or more of the following: based on the lead field matrix, traditional source imaging results, template head model, brain region priors, and random initialization. Initialization based on the lead field matrix, traditional source imaging results, template head model, and brain region priors is used to ensure the transformation matrix has a source space mapping prior, while random initialization is used to set the initial parameters of the transformation matrix when source space prior information is lacking.

[0059] S1.3: Construct a task decoding model and connect it to a trainable source space transformation layer. This enables the task decoding model to receive brain source signals output from the trainable source space transformation layer and perform feature extraction and task discrimination on the brain source signals. The task decoding model includes a spatial feature extraction unit, a temporal feature extraction unit, and a task discrimination unit. The spatial feature extraction unit extracts the spatial distribution features of brain source signals corresponding to left-hand and right-hand motor imagery between different preset source spatial locations. The temporal feature extraction unit extracts the temporal dynamic features of brain source signals between different time segments. The task discrimination unit obtains the motor imagery category prediction result corresponding to the current EEG data based on the spatial distribution features and temporal dynamic features. In the task discrimination unit, the spatial distribution features and temporal dynamic features are first fused to obtain a joint feature vector. Then, the joint feature vector is sequentially input into two cascaded fully connected layers. Through linear transformation and activation function processing of the fully connected layers, it is mapped to a discrimination score equal to the number of preset task categories. Finally, the discrimination score is normalized to a probability distribution using the Softmax function, and the task category corresponding to the maximum probability is taken as the motor imagery category prediction result corresponding to the current EEG data.

[0060] S1.4: The trainable source space transformation layer and the task decoding model are combined into an EEG signal decoding network, so that the brain source signal generation process and the motor imagery task decoding process are in the same training framework, and the task loss generated by the task decoding model can be applied to the trainable source space transformation layer through backpropagation, so that the transformation matrix can be updated with the motor imagery task decoding target and the source space regulation target.

[0061] S2: Input the EEG data into the EEG signal decoding network to obtain the task prediction result corresponding to the current EEG data, and calculate the task loss based on the task prediction result and task label; the specific details are as follows:

[0062] S2.1: Input the EEG data in the current training batch into the EEG signal decoding network, so that the EEG data first passes through the trainable source space transformation layer to generate brain source signals, and then the task decoding model performs feature extraction and task discrimination on the brain source signals to obtain the task prediction result of the current EEG data corresponding to left-hand motor imagery or right-hand motor imagery.

[0063] S2.2: Calculate the task loss based on the task prediction results and task labels, ensuring that the task loss can characterize the difference between the predicted results of the motion imagery category and the actual task labels; let the task prediction result output by the task decoding model be... The task is tagged as When the EEG decoding task is a classification task, the task loss is calculated using cross-entropy loss, expressed as follows: ,in, Indicates mission loss. Indicates the number of task categories. Indicates category index, Indicates the task label is in the first position. Values ​​on the class The output of the task decoding model represents the first... Class prediction probability; in the binary classification task of left-handed motor imagery and right-handed motor imagery. The value is 2.

[0064] S3: Real-time acquisition of brain source signals output from the trainable source space transformation layer in the EEG signal decoding network; calculation of source space evaluation metrics based on the brain source signals; generation of visualization results of brain source signals at preset source space locations; construction of a source space constraint loss that can participate in training optimization based on the source space evaluation metrics; details are as follows:

[0065] S3.1: Real-time acquisition of brain source signals output by the trainable source space transformation layer in the EEG signal decoding network, making brain source signals the common basis for source space evaluation, source space visualization, and source space constraint loss construction; In each training iteration or preset training interval, the brain source signals corresponding to left-hand motor imagery or right-hand motor imagery are output as intermediate results, so that the brain source signals are not only input into the task decoding model, but also simultaneously enter the source space evaluation, visualization display, and source space constraint loss calculation process;

[0066] S3.2: Calculate source space evaluation indicators based on real-time acquired brain source signals. These indicators quantify the current source space state of the brain source signals. Let the first... The brain signal time series corresponding to each preset source spatial location are: ,in, Indicates the preset source spatial location index. Indicates the first Time series of brain source signals corresponding to preset source spatial locations. Indicates brain-derived signals The Middle The set of all time-point signals corresponding to each preset source spatial location, and the task-related preset source spatial location set is as follows: The set of non-task-related preset source spatial locations is The set of all preset source spatial locations is In the task of imagining movement, Corresponding to the motor cortex and its adjacent regions. For preset source spatial locations that are weakly or unrelated to the current motion visualization task, the activity intensity of the task-related source spatial location is expressed as: ,in, Represents the number of sets, such as This indicates the number of preset source space locations related to the task. For the first Each preset source spatial location corresponds to a brain source signal time sequence. The norm is used to represent the total activity of brain source signals at the preset source spatial location within the entire time window; the proportion of activity at non-task-related source spatial locations is represented as... ,in Represents a positive constant to prevent the denominator from being zero; the degree of diffusion in the source space is represented as... ,in, Indicates the source space activity threshold. express Total activity level exceeds threshold The number of source locations, This represents the total number of preset source spatial locations; spatial smoothness is represented as... ,in, This represents the set of adjacency relationships between preset source spatial locations. This indicates another preset source spatial location index. This represents the spatial locations of two predefined source locations that are adjacent to each other. Indicates the first Time series of brain source signals corresponding to preset source spatial locations. express The square of the norm is used here to calculate the degree of difference in the time series of brain source signals between adjacent preset source spatial locations; the smaller the difference, the better the spatial smoothness. Let the first norm be... The spatial vector of brain source signal corresponding to each time point is: , Indicates brain-derived signals The Middle The column vector corresponding to each time point is then represented as: ,in, Indicates the first The spatial vector of brain-source signals corresponding to each time point The degree of difference in the spatial vectors of brain source signals between adjacent time points is calculated; a smaller difference indicates better temporal continuity. When forward reconstruction consistency is adopted, the forward model or lead field matrix is ​​assumed to be... Then the forward reconstruction consistency index is expressed as ,in, This represents the spatial EEG signal of the sensor obtained by reconstructing brain-source signals. Let Frobenius norm be the square of the norm; let the mean of the brain source signal categories corresponding to the left-hand motor imagery category be . The mean brain signal category corresponding to the right-hand motor imagery category is The differences in source regions related to the category are then expressed as .

[0067] S3.3: Based on real-time acquired brain source signals, generate visualization results, map the activity values ​​corresponding to each preset source space location in the brain source signal to the corresponding brain region, cortical grid point or source space node, so that the expert review terminal can view the source space activity distribution of the brain source signal corresponding to the current left-hand motor imagery or right-hand motor imagery, the activity concentration of the preset source space location related to movement, and the abnormal activation of the preset source space location not related to the task.

[0068] S3.4: Construct a source space constraint loss that can participate in training optimization based on source space evaluation metrics, so that the source space constraint loss can constrain the source space representation of brain source signals; the source space constraint loss is constructed based on source space evaluation metrics and their corresponding differentiable forms, including one or more of the following: task-related source space location constraint loss, non-task-related source space location suppression loss, source space sparsity loss, spatial smoothing loss, temporal continuity loss, and forward reconstruction consistency loss; the task-related source space location constraint loss is expressed as... ,in, The threshold for activity intensity at spatial locations of task-related sources; the spatial location suppression loss for non-task-related sources. The source space sparsity loss is expressed as Spatial smoothing loss Time continuity loss Forward Reconstruction Consistency Loss The source space constraint loss is expressed as ,in, These represent the weights corresponding to the task-related source spatial location constraint loss, non-task-related source spatial location suppression loss, source spatial sparsity loss, spatial smoothing loss, temporal continuity loss, and forward reconstruction consistency loss, respectively.

[0069] S4: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules. Output the source space evaluation indicators, automatic feedback control suggestions, and visualization results of brain source signals to the expert review end, and determine the target feedback control strategy based on the expert review results; the specific details are as follows:

[0070] S4.1: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules. When the activity intensity of task-related source spatial locations is lower than a preset threshold, it indicates that the brain source signals corresponding to left-hand or right-hand motor imagery are insufficiently represented at the preset source spatial locations related to movement. Automatic feedback regulation suggestions are generated to enhance the mapping weights of the preset source spatial locations related to movement or increase the constraint loss weights of the task-related source spatial locations. When the proportion of activity in non-task-related source spatial locations is higher than a preset threshold, it indicates overactivation in the non-task-related preset source spatial locations. Automatic feedback regulation suggestions are generated to reduce the mapping weights of the non-task-related preset source spatial locations or increase the inhibition loss weights of the non-task-related source spatial locations. When the source spatial diffusion degree is higher than a preset threshold, it indicates that brain source signals are distributed across too many preset source spatial locations. Automatic feedback regulation suggestions are generated to increase the source spatial sparsity loss weights or adjust the source spatial mask range. The automatic feedback regulation suggestions include regulation targets, regulation directions, and suggested regulation strengths. The regulation target indicates the source spatial mapping relationship, source spatial mask, loss weights, or training parameters that need to be adjusted. The regulation direction indicates the direction of regulation, such as enhancement, inhibition, expansion, contraction, increase, or decrease. The suggested regulation strength indicates the amplitude of the corresponding regulation action.

[0071] S4.2: Output source space evaluation indicators, automatic feedback regulation suggestions, and visualization results of brain source signals to the expert review end, and receive expert auxiliary review results, so that the automatic feedback regulation suggestions can be reviewed based on the visualization results and quantitative evaluation results; for example, when the visualization results show that the brain source signal activity in the preset source space location related to movement is insufficient, and the source space evaluation indicators show that the activity intensity of the task-related source space location is lower than the preset threshold, the expert review end simultaneously displays automatic feedback regulation suggestions to enhance the representation of the preset source space location related to movement for confirmation or adjustment.

[0072] S4.3: Determine the target feedback control strategy based on the expert-assisted review results; when the expert-assisted review results indicate that the automatic feedback control recommendations are applicable, the automatic feedback control recommendations are determined as the target feedback control strategy; when the expert-assisted review results indicate that the automatic feedback control recommendations need to be adjusted, the control target, control direction, or control intensity is modified based on the expert-assisted review results, and the modified result is determined as the target feedback control strategy; for example, when the visualization results indicate that only some motion-related preset source spatial positions need to be enhanced, the control target is modified from all motion-related preset source spatial positions to the corresponding partial preset source spatial positions, and a target feedback control strategy is formed.

[0073] S5: Using task loss and source space constraint loss as training optimization criteria, and target feedback regulation strategy as real-time regulation criteria, the target regulation object and corresponding regulation parameters in the EEG signal decoding network are determined. The EEG signal decoding network is then updated based on the target regulation object and corresponding regulation parameters, so that the brain source signal generation process and task decoding process are adjusted according to the target feedback regulation strategy, thereby achieving real-time feedback regulation. The specific details are as follows:

[0074] S5.1: Using task loss and source space constraint loss as the basis for training optimization, and target feedback regulation strategy as the basis for real-time regulation, the target regulation objects in the EEG signal decoding network are determined. For example, when the target feedback regulation strategy is to enhance the representation of motion-related preset source space locations, the parameters, source space masks, loss weights, or task decoding model training parameters corresponding to the motion-related preset source space locations in the transformation matrix are determined as the target regulation objects; when the target feedback regulation strategy is to suppress the activation of non-task-related preset source space locations, the parameters, source space masks, or loss weights corresponding to the non-task-related preset source space locations in the transformation matrix are determined as the target regulation objects; when the target feedback regulation strategy is to suppress source space diffusion, the source space mask and source space sparse loss weights are determined as the target regulation objects.

[0075] S5.2: Determine the corresponding control parameters based on the target control object, so that the corresponding control parameters are used to limit the adjustment direction, adjustment range, and adjustment magnitude of the target control object. Here, the transformation matrix corresponds to the update direction, update magnitude, and adjustment coefficient of the parameter corresponding to the specified preset source space position; the source space mask corresponds to the mask range, mask weight, and mask update magnitude; the loss weight corresponds to the task loss weight, source space constraint loss weight, and weights of each source space constraint sub-item; the task decoding model training parameters correspond to the learning rate, regularization coefficient, and parameter update magnitude. When the target feedback control strategy is directed to enhance the representation of motion-related preset source space positions, the corresponding control parameters are used to increase the update magnitude of the parameters corresponding to the motion-related preset source space positions, or increase the weight of the task-related source space position constraint loss; when the target feedback control strategy is directed to suppress the activation of non-task-related preset source space positions, the corresponding control parameters are used to decrease the update magnitude of the parameters corresponding to the non-task-related preset source space positions, or increase the weight of the non-task-related source space position suppression loss.

[0076] S5.3: The EEG signal decoding network is updated based on the target object and its corresponding control parameters, so that the target object adjusts according to the target feedback control strategy, and this adjustment acts on the brain signal generation process or task decoding process to achieve real-time feedback control; let the parameters to be updated in the EEG signal decoding network be... The corresponding control parameter is The parameter update process can then be represented as: ,in, It is the gradient operator, representing the gradient with respect to the set of parameters. Find the partial derivative. Indicates the overall training objective Relative to the parameter to be updated The gradient; when the target feedback regulation strategy is to enhance the representation of motion-related preset source spatial locations, the contribution of motion-related preset source spatial locations to the brain source signal generation process is increased by updating the transformation matrix, source spatial mask or loss weight; when the target feedback regulation strategy is to suppress the activation of non-task-related preset source spatial locations, the influence of non-task-related preset source spatial locations on the brain source signal generation process and task decoding process is reduced by updating the transformation matrix, source spatial mask or loss weight.

[0077] Example 2

[0078] This embodiment discloses a source-space-based real-time feedback control system for EEG signal decoding, used to implement the source-space-based real-time feedback control method for EEG signal decoding described in Embodiment 1. In left-hand and right-hand motor imagery tasks, it provides real-time feedback control over the source-space mapping process and task decoding process of multi-channel EEG signals. Figure 2 As shown, it includes a source space conversion module, a brain source signal output module, a task decoding module, a source space evaluation module, a source space visualization module, an expert-assisted review module, and a feedback control module;

[0079] The source space conversion module is used to receive EEG data corresponding to left-hand or right-hand motor imagery, and to map the EEG data from sensor space to brain source space through a trainable source space conversion layer to generate brain source signals. The source space conversion module is equipped with a conversion matrix, which is used to establish a mapping relationship between the EEG acquisition channel and the preset source space position, so that each signal component in the generated brain source signal corresponds to the preset source space position.

[0080] The brain source signal output module is connected to the source space conversion module and is used to acquire the brain source signal output by the source space conversion module in real time, and synchronously output the brain source signal to the task decoding module, the source space evaluation module and the source space visualization module.

[0081] The task decoding module is connected to the brain source signal output module and the feedback control module respectively. It receives the brain source signal output by the brain source signal output module through the task decoding model, performs feature extraction and task discrimination on the brain source signal, obtains the task prediction result corresponding to the current EEG data, calculates the task loss based on the task prediction result and task label, and outputs the task loss to the feedback control module so that the task discrimination error can participate in the subsequent network update; wherein, the task label is provided by the training data or the label input terminal;

[0082] The source space evaluation module is connected to the brain source signal output module, the expert-assisted review module, and the feedback control module, respectively. It is used to calculate the source space evaluation index, construct a source space constraint loss that can participate in training optimization based on the source space evaluation index, and generate automatic feedback control suggestions based on the source space evaluation index and preset feedback control rules. Then, the source space evaluation index and the automatic feedback control suggestions are output to the expert-assisted review module, and the source space constraint loss is output to the feedback control module. Among them, the source space evaluation index is used to quantify the source space state of the current brain source signal, the source space constraint loss is used to constrain the source space expression of the brain source signal, and the automatic feedback control suggestions are used to indicate the source space mapping relationship, source space mask, loss weight, or task decoding model training parameters to be adjusted.

[0083] The source space visualization module is connected to the brain source signal output module and the expert-assisted review module respectively. It is used to generate visualization results of the brain source signals output by the brain source signal output module, so that the visualization results show the activity distribution of the current EEG data at the preset source space location, and output the visualization results to the expert-assisted review module.

[0084] The expert-assisted review module is used to receive source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals, output source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals to the expert review terminal, receive the expert-assisted review results returned by the expert review terminal, determine the target feedback control strategy based on the expert-assisted review results, and output the target feedback control strategy to the feedback control module.

[0085] The feedback control module is connected to the task decoding module, the source space evaluation module, and the expert-assisted review module, respectively. It receives the task loss output by the task decoding module, the source space constraint loss output by the source space evaluation module, and the target feedback control strategy output by the expert-assisted review module. Using the task loss and source space constraint loss as the basis for training optimization and the target feedback control strategy as the basis for real-time control, it determines the target control object and its corresponding control parameters in the EEG signal decoding network. Then, it updates the EEG signal decoding network according to the target control object and its corresponding control parameters, so that the brain source signal generation process and the task decoding process are adjusted according to the target feedback control strategy, thereby forming a real-time feedback control closed loop composed of brain source signal output, source space evaluation, expert-assisted review, and feedback control.

[0086] Through the collaboration of the above modules, the brain source signal output module serves as the signal distribution node between the source space conversion module, the task decoding module, the source space evaluation module, and the source space visualization module. The expert-assisted review module serves as the review node between the automatic feedback regulation suggestions and the target feedback regulation strategy. The feedback regulation module serves as the execution node for the target feedback regulation strategy to act on the EEG signal decoding network. This enables the source space evaluation results to act on the brain source signal generation process and the task decoding process through the target regulation object and its corresponding regulation parameters, thereby realizing real-time feedback regulation based on the source space brain source signal.

[0087] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A real-time feedback modulation method for decoding electroencephalogram (EEG) signals based on source space, characterized in that, Includes the following steps: S1: Construct an EEG signal decoding network consisting of a trainable source space transformation layer and a task decoding model. The trainable source space transformation layer is used to map EEG data from the sensor space to the brain source space, and the task decoding model is used to receive the brain source signals output by the trainable source space transformation layer and perform feature extraction and task discrimination on the brain source signals. S2: Input the EEG data into the EEG signal decoding network to obtain the task prediction result corresponding to the current EEG data, and calculate the task loss based on the task prediction result and task label; S3: Real-time acquisition of brain source signals output by the trainable source space transformation layer in the EEG signal decoding network, calculation of source space evaluation index based on brain source signals, and generation of visualization results of brain source signals at preset source space locations; Construct a source space constraint loss that can participate in training optimization based on the source space evaluation index; S4: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules, output source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals to the expert review end, and determine the target feedback control strategy based on the expert review results; S5: Using task loss and source space constraint loss as the basis for training optimization, and target feedback regulation strategy as the basis for real-time regulation, the target regulation object and corresponding regulation parameter in the EEG signal decoding network are determined, and the EEG signal decoding network is updated according to the target regulation object and corresponding regulation parameter, so that the brain source signal generation process and task decoding process are adjusted according to the target feedback regulation strategy, thereby realizing real-time feedback regulation.

2. The method for real-time feedback modulation of EEG signals based on source space according to claim 1, characterized in that, The specific steps for step S1 are as follows: S1.1: Construct a trainable source space transformation layer. The trainable source space transformation layer sets a transformation matrix. The transformation matrix is ​​used to map EEG data from sensor space to brain source space to generate brain source signal. Each signal component in the brain source signal corresponds to a preset source space position. The preset source space position includes one or more of brain regions, cortical grid points and source space nodes. S1.2: Initialize the transformation matrix so that it has an initial mapping relationship from sensor space to brain source space at the start of training. The initialization method includes one or more of the following: based on the lead field matrix, traditional source imaging results, template head model, brain region prior, and random initialization. Among them, the initialization based on the lead field matrix, traditional source imaging results, template head model, and brain region prior is used to make the transformation matrix have source space mapping prior, and random initialization is used to set the initial parameters of the transformation matrix when source space prior information is lacking. S1.3: Construct a task decoding model and connect it to a trainable source spatial transformation layer. This enables the task decoding model to receive brain source signals output by the trainable source spatial transformation layer and perform feature extraction and task discrimination on the brain source signals. The task decoding model includes a spatial feature extraction unit, a temporal feature extraction unit, and a task discrimination unit. The spatial feature extraction unit is used to extract the spatial distribution features of brain source signals between different preset source spatial locations. The temporal feature extraction unit is used to extract the temporal dynamic features of brain source signals between different time segments. The task discrimination unit is used to obtain the task prediction result corresponding to the current EEG data based on the spatial distribution features and temporal dynamic features. In the task discrimination unit, the spatial distribution features and temporal dynamic features are first fused to obtain a joint feature vector. Then, the joint feature vector is input into at least one fully connected layer and processed by the linear transformation and activation function of the fully connected layer to map it into a discrimination score equal to the number of preset task categories. Finally, the discrimination score is normalized to a probability distribution by the Softmax function, and the task category corresponding to the maximum probability is taken as the task prediction result corresponding to the current EEG data. S1.4: The trainable source space transformation layer and the task decoding model are combined into an EEG signal decoding network, so that the brain source signal generation process and the task decoding process are within the same training framework, and the task loss generated by the task decoding model can be applied to the trainable source space transformation layer through backpropagation.

3. The method for real-time feedback modulation of EEG signals based on source space according to claim 2, characterized in that, The specific steps for step S2 are as follows: S2.1: Input the EEG data into the EEG signal decoding network, so that the EEG data first passes through the trainable source space transformation layer to generate brain source signals, and then the task decoding model performs feature extraction and task discrimination on the brain source signals to obtain the task prediction result corresponding to the current EEG data; S2.2: Calculate the task loss based on the task prediction result and the task label, so that the task loss can characterize the difference between the task prediction result and the task label; wherein, the task label is provided by the training data or the label input.

4. The real-time feedback modulation method for decoding EEG signals based on source space according to claim 3, characterized in that, The specific steps for step S3 are as follows: S3.1: Real-time acquisition of brain source signals output by the trainable source space transformation layer in the EEG signal decoding network, making brain source signals the common basis for source space evaluation, source space visualization and source space constraint loss construction; S3.2: Calculate source space evaluation indicators based on real-time acquired brain source signals. Source space evaluation indicators are used to quantify the source space state of the current brain source signals, including one or more of the following: activity intensity of task-related source space locations, activity proportion of non-task-related source space locations, source space diffusion degree, spatial smoothness, temporal continuity, forward reconstruction consistency, and category-related source region differences. S3.3: Generate visualization results based on real-time acquired brain source signals, so that the visualization results show the activity distribution of the current EEG data at preset source spatial locations. When generating visualization results, the activity values ​​corresponding to each preset source spatial location in the brain source signal are mapped to the corresponding brain regions, cortical grid points or source spatial nodes to form a source spatial activity distribution that can be displayed by the expert review end. S3.4: Construct source space constraint loss that can participate in training optimization based on source space evaluation index, so that source space constraint loss can constrain the source space representation of brain source signals; wherein, source space constraint loss is constructed based on source space evaluation index and its corresponding differentiable form, including one or more of the following: task-related source space location constraint loss, non-task-related source space location inhibition loss, source space sparsity loss, spatial smoothing loss, temporal continuity loss, and forward reconstruction consistency loss.

5. The method for real-time feedback modulation of EEG signals based on source space according to claim 4, characterized in that, The specific steps for step S4 are as follows: S4.1: Generate automatic feedback control suggestions based on source space evaluation indicators and preset feedback control rules. The preset feedback control rules include triggering conditions for the intensity of activity in task-related source space locations, the proportion of activity in non-task-related source space locations, the degree of source space diffusion, temporal continuity, and forward reconstruction consistency. When the source space evaluation indicators reach or exceed the corresponding triggering conditions, automatic feedback control suggestions are generated. Automatic feedback control suggestions include control targets, control directions, and suggested control strengths. Control targets are used to indicate the source space mapping relationship, source space mask, loss weight, or training parameters that need to be adjusted. Control directions are used to indicate the direction of control, such as enhancement, suppression, expansion, contraction, improvement, or reduction. Suggested control strengths are used to indicate the magnitude of the corresponding control action. S4.2: Output the source space evaluation indicators, automatic feedback regulation suggestions and visualization results of brain source signals to the expert review end, and receive the expert auxiliary review results, so that the automatic feedback regulation suggestions can be reviewed based on the visualization results and quantitative evaluation results; S4.3: Determine the target feedback control strategy based on the expert-assisted review results. When the expert-assisted review results indicate that the automatic feedback control recommendations are applicable, the automatic feedback control recommendations shall be determined as the target feedback control strategy. When the expert-assisted review results indicate that the automatic feedback control recommendations need to be adjusted, the control objectives, control direction, and control intensity are revised based on the expert-assisted review results, and the revised results are determined as the target feedback control strategy.

6. The method for real-time feedback modulation of EEG signals based on source space according to claim 5, characterized in that, The specific steps for step S5 are as follows: S5.1: Using task loss and source space constraint loss as the basis for training optimization, and target feedback regulation strategy as the basis for real-time regulation, the target regulation object in the EEG signal decoding network is determined. The target regulation object is an adjustable object that can affect the brain source signal generation process, source space constraint process, and task decoding process, including one or more of the following: transformation matrix, source space mask, loss weight, and task decoding model training parameters. Among them, the transformation matrix is ​​used to adjust the mapping relationship from sensor space to brain source space, the source space mask is used to limit the preset source space location range for participating in source space evaluation, source space constraint, or source space visualization, the loss weight is used to adjust the strength of the role of task loss and source space constraint loss in the training process, and the task decoding model training parameters are used to adjust the parameter update process of the task decoding model. S5.2: Determine the corresponding control parameters based on the target control object, so that the corresponding control parameters can limit the adjustment direction and adjustment range of the target control object; wherein, the control parameters corresponding to the transformation matrix include the update direction, update range, and adjustment coefficients of the parameters corresponding to the specified preset source space position; the control parameters corresponding to the source space mask include the mask range, mask weight, and mask update range; the control parameters corresponding to the loss weight include the task loss weight, source space constraint loss weight, and weights of each source space constraint sub-item; and the control parameters corresponding to the task decoding model training parameters include the learning rate, regularization coefficient, and parameter update range. S5.3: The EEG signal decoding network is updated according to the target control object and its corresponding control parameters, so that the target control object adjusts according to the target feedback control strategy, thereby regulating the brain source signal generation process and the task decoding process, and realizing real-time feedback control.

7. A real-time feedback control system for decoding electroencephalogram (EEG) signals based on source space, characterized in that, The method for real-time feedback control of EEG signal decoding based on source space as described in any one of claims 1 to 6 includes a source space conversion module, a brain source signal output module, a task decoding module, a source space evaluation module, a source space visualization module, an expert-assisted review module, and a feedback control module. The source space conversion module is used to receive EEG data and map the EEG data from the sensor space to the brain source space through a trainable source space conversion layer to generate brain source signals. The source space conversion module is equipped with a conversion matrix, which is used to establish a mapping relationship between the EEG acquisition channel and the preset source space position, so that each signal component in the generated brain source signal corresponds to the preset source space position. The brain source signal output module is connected to the source space conversion module and is used to acquire the brain source signal output by the source space conversion module in real time, and synchronously output the brain source signal to the task decoding module, the source space evaluation module and the source space visualization module. The task decoding module is connected to the brain source signal output module and the feedback control module respectively. It receives the brain source signal output by the brain source signal output module through the task decoding model, performs feature extraction and task discrimination on the brain source signal, obtains the task prediction result corresponding to the current EEG data, calculates the task loss based on the task prediction result and the task label, and outputs the task loss to the feedback control module; wherein, the task label is provided by the training data or the label input terminal. The source space evaluation module is connected to the brain source signal output module, the expert-assisted review module, and the feedback control module, respectively. It is used to calculate the source space evaluation index, construct a source space constraint loss that can participate in training optimization based on the source space evaluation index, and generate automatic feedback control suggestions based on the source space evaluation index and preset feedback control rules. Then, the source space evaluation index and the automatic feedback control suggestions are output to the expert-assisted review module, and the source space constraint loss is output to the feedback control module. Among them, the source space evaluation index is used to quantify the source space state of the current brain source signal, the source space constraint loss is used to constrain the source space expression of the brain source signal, and the automatic feedback control suggestions are used to indicate the source space mapping relationship, source space mask, loss weight, or task decoding model training parameters to be adjusted. The source space visualization module is connected to the brain source signal output module and the expert-assisted review module respectively. It is used to generate visualization results of the brain source signals output by the brain source signal output module, so that the visualization results show the activity distribution of the current EEG data at the preset source space location, and output the visualization results to the expert-assisted review module. The expert-assisted review module is used to receive source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals, output source space evaluation indicators, automatic feedback control suggestions and visualization results of brain source signals to the expert review terminal, receive the expert-assisted review results returned by the expert review terminal, determine the target feedback control strategy based on the expert-assisted review results, and output the target feedback control strategy to the feedback control module. The feedback control module is connected to the task decoding module, the source space evaluation module, and the expert-assisted review module, respectively. It receives the task loss output by the task decoding module, the source space constraint loss output by the source space evaluation module, and the target feedback control strategy output by the expert-assisted review module. Using the task loss and source space constraint loss as the basis for training optimization and the target feedback control strategy as the basis for real-time control, it determines the target control object and its corresponding control parameters in the EEG signal decoding network. Then, it updates the EEG signal decoding network according to the target control object and its corresponding control parameters, so that the brain source signal generation process and the task decoding process are adjusted according to the target feedback control strategy, thereby realizing real-time feedback control.

Images