Discriminating method and system for complex brain state and storage medium

Abstract

The invention relates to the technical field of brain nerve data driving, in particular to a complex brain state discrimination method and system and a storage medium, and the method comprises the steps: carrying out the hierarchical processing of the time sequence dynamic characteristics of an acquired original EEG signal, so as to extract the time sequence characteristics of each time window; the time sequence features are fused to construct a space graph structure, multi-layer dynamic graph convolution operation is carried out on the space graph structure to generate a multi-layer space feature tensor used for representing the space dependency relationship of the original EEG signals, and the multi-layer space feature tensor is stacked into a space feature three-dimensional tensor according to the hierarchy dimension of space decomposition; performing cross-level feature fusion on the spatial feature three-dimensional tensor by using a gating circulation unit to obtain cross-level spatial fusion features, and calculating discrimination values of different complex brain states; and finally, by simulating a brain cognitive decision loop, determining the type of the complex brain state based on the discrimination value. According to the method and the device, the technical problem of how to efficiently decode the EEG signal to comprehensively and accurately represent the complex cognitive state of the brain is solved.

Description

A method, system, and storage medium for identifying complex brain states. Technical Field

[0001] This invention relates to the field of brain neural data-driven technology, and in particular to a method, system, and storage medium for identifying complex brain states. Background Technology

[0002] Neurophysiology shows that the brain's neural mechanisms produce corresponding functional responses when stimulated. However, these mechanisms are not merely single stimulus responses, but rather involve complex feedback loops for cognitive task processing. Therefore, these functional responses manifest as complex cognitive loops in the brain's decision-making process. Electroencephalogram (EEG) signals, as a non-invasive method of recording brain activity with high temporal resolution, have become a key tool for decoding cognitive functions and are widely used in areas such as auditory attention, mental fatigue, and visual attention.

[0003] Existing technologies leverage the advantages of deep neural networks to learn advanced discriminative features from EEG signals. For example, convolutional neural networks are used to extract temporal features from EEG signals, capturing their temporal information; Long Short-Term Memory (LSTM) networks are used to analyze the time-step dependencies of nonlinear temporal information in EEG signals; Transformer models are used to effectively identify coarse-to-fine temporal patterns, achieving effective recognition of spatial topological patterns in EEG signals, thereby facilitating the decoding process; Graph Convolutional Networks (GCNs) are used to model non-Euclidean relationships between channels, capturing and analyzing the complex spatial correlations of EEG signals; and Graph Attention Networks (GATs) are introduced to highlight fine-grained representations between channels, further capturing discriminative cognitive patterns from the complex spatial characteristics of EEG signals. However, due to the non-stationarity of EEG signals, complex nonlinear causal relationships, and the complexity of the topological spatial distribution of spatial electrode channels, and because brain cognitive activity is essentially a complex system of spatiotemporal coupling, with hierarchical organization of brain neuronal circuits and functional regions, and different brain regions closely related to specific functions, they do not work completely independently; brain region activation involves chain reactions and collaborative patterns. Therefore, focusing on only one aspect is insufficient to reflect dynamic changes and cannot fully utilize neurophysiology to simulate the spatiotemporal coupling nature of brain cognitive activity. This makes it difficult to capture the integrity of EEG signals, easily destroys the intrinsic integrity of neural information, and makes it impossible to efficiently decode EEG signals, resulting in an incomplete and inaccurate representation of the brain's complex cognitive state. Summary of the Invention

[0004] This application aims to provide a method, system, and storage medium for identifying complex brain states. By comprehensively capturing the spatiotemporal dynamic characteristics of EEG signals and simulating the brain's cognitive decision-making circuits, it achieves a comprehensive and accurate representation of complex brain states, solving the technical problem of how to efficiently decode EEG signals to comprehensively and accurately represent complex cognitive states of the brain.

[0005] Specifically, this is achieved through the following technical solution: Firstly, this application provides a method for discriminating complex brain states, comprising: acquiring raw EEG signals; performing hierarchical processing on the temporal dynamic characteristics of the raw EEG signals to extract temporal features of each time window of the raw EEG signals, and fusing the extracted temporal features of each time window to obtain a temporal collaborative feature matrix of the raw EEG signals; constructing a spatial graph structure of the raw EEG signals based on the temporal collaborative feature matrix, performing multi-level dynamic graph convolution operations on the spatial graph structure to generate a multi-level spatial feature tensor for characterizing the spatial dependencies of the raw EEG signals; stacking the multi-level spatial feature tensor according to the hierarchical dimensions of spatial decomposition to form a three-dimensional spatial feature tensor; using a gated recurrent unit to perform cross-level feature fusion on the three-dimensional spatial feature tensor to obtain cross-level spatial fusion features; calculating the discriminating value of different complex brain states based on the cross-level spatial fusion features; and finally, determining the type of complex brain state by simulating the brain's cognitive decision-making circuit and based on the discriminating value.

[0006] Secondly, this application provides a system for discriminating complex brain states, comprising: a signal acquisition module for acquiring raw EEG signals; a temporal collaborative feature extraction module for performing hierarchical processing on the temporal dynamic characteristics of the raw EEG signals, extracting temporal features of each time window of the raw EEG signals, and fusing the extracted temporal features of each time window to obtain a temporal collaborative feature matrix of the raw EEG signals; a deep brain space exploration module for constructing a spatial graph structure of the raw EEG signals based on the temporal collaborative feature matrix, performing multi-level dynamic graph convolution operations on the spatial graph structure to generate a multi-level spatial feature tensor for characterizing the spatial dependencies of the raw EEG signals; a high-level cross-level fusion module for stacking the multi-level spatial feature tensor into a three-dimensional spatial feature tensor according to the hierarchical dimensions of spatial decomposition; performing cross-level feature fusion on the three-dimensional spatial feature tensor using a gated loop unit to obtain cross-level spatial fusion features; calculating the discriminating value of different complex brain states based on the cross-level spatial fusion features; and a complex brain state discrimination module for determining the type of complex brain state by simulating the brain's cognitive decision-making circuit and based on the discriminating value.

[0007] Fourthly, embodiments of this application provide a computer-readable storage medium, wherein a computer program present in the computer-readable storage medium, when executed by a computer device, is used to implement the method for determining a complex brain state as described in the first aspect.

[0008] As shown above, this application extracts temporal features of each time window of the original EEG signal by performing hierarchical processing on the temporal dynamic characteristics of the original EEG signal; it then fuses these temporal features to obtain the temporal co-feature matrix of the original EEG signal, thus completing the hierarchical extraction of temporal features; based on the temporal co-feature matrix, it constructs the spatial graph structure of the original EEG signal, performs multi-level dynamic graph convolution operations on the spatial graph structure, and generates a multi-level spatial feature tensor to represent the spatial dependencies of the original EEG signal, thus completing the hierarchical extraction of spatial features. This multi-level spatial feature tensor reflects the spatial dependencies from low to high levels in the EEG signal, providing rich spatial topological features for subsequent cross-level fusion of spatial features, which is beneficial to improving the accuracy of brain state discrimination; finally, the multi-level spatial feature tensor is stacked according to the hierarchical dimensions of spatial decomposition to form a spatial feature three-dimensional model. This paper proposes a method to perform cross-level feature fusion on the three-dimensional tensor of spatial features using gated recurrent units. This results in cross-level spatial fusion features, enabling the comprehensive capture of the spatiotemporal dynamic characteristics of EEG signals. This lays a crucial feature foundation for simulating the brain's cognitive decision-making circuit and achieving a comprehensive and accurate representation of complex brain states. Finally, the discriminative value of different complex brain states is calculated based on the cross-level spatial fusion features. By simulating the brain's cognitive decision-making circuit, the type of complex brain state is determined, achieving a comprehensive and accurate representation of complex brain states. Furthermore, the discriminative value provides a core foundation for simulating the brain's cognitive decision-making circuit, and the accuracy of brain state discrimination can be continuously optimized through reinforcement learning strategies. This application solves the technical problems of incomplete and inaccurate representation of complex cognitive states in existing brain technologies. Attached Figure Description

[0009] Figure 1 is a flowchart illustrating a method for discriminating complex brain states according to an embodiment of this application; Figure 2 is a radar chart comparing individual accuracy on the DSR dataset according to an embodiment of this application; Figure 3 is a radar chart comparing individual accuracy on the FAT dataset according to an embodiment of this application; Figure 4 is a radar chart comparing individual accuracy on the KUL dataset according to an embodiment of this application; Figure 5 is a structural diagram illustrating a system for discriminating complex brain states according to an embodiment of this application; Figure 6 is a structural diagram illustrating an electronic device according to an embodiment of this application. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] Please refer to Figure 1 for Example 1. Figure 1 is a flowchart illustrating a method for identifying complex brain states provided in this application embodiment. It should be noted that the executing entity of this method can be a complex brain state identification system. As shown in Figure 1, the method for identifying complex brain states includes the following: 101. Acquiring raw EEG signals.

[0012] EEG signals consist of channel and time dimensions, typically represented as two-dimensional time series, images, and spatial graph structures. They are a non-invasive method of recording brain activity with high temporal resolution. However, the primary form of EEG signals is time series, exhibiting non-stationary and irregular changes. Because brain states are dynamically influenced by cognitive tasks, EEG signals exhibit non-stationary characteristics in the time domain. EEG signals do not possess single-domain characteristics; therefore, observing only a single pattern easily leads to the loss of complementary information, as this information is distributed across the temporal dynamics of frequency rhythms and spatial topology. Therefore, this application's embodiments require a comprehensive capture of the spatiotemporal dynamic characteristics of EEG signals, laying a crucial feature foundation for subsequent simulation of the brain's cognitive decision-making circuits to achieve a comprehensive and accurate representation of complex brain states.

[0013] This application embodiment illustrates a method for discriminating complex brain states by collecting three EEG signal datasets related to cognitive states. These include the cognitive attention dataset DSR, the auditory attention dataset KUL, and the driving fatigue dataset FAT; that is, the cognitive attention dataset DSR, the auditory attention dataset KUL, and the driving fatigue dataset FAT serve as the original EEG signals in this application embodiment. Specifically, the DSR dataset is a multimodal brain imaging dataset used to measure three cognitive tasks in healthy subjects. A discrimination / selective response (DSR) task is used for cognitive attention classification. Twenty-six subjects were selected, and to avoid the influence of cross-conversation differences, the first of the three sessions was used for each subject. Each phase consisted of several series of attention task periods (40s) and rest periods (20s). At a sampling rate of 1 kHz, 28 EEG channels and 2 Electrooculogram (EOG) channels were acquired and recorded. The EOG channels recorded bioelectrical signals generated by eye movements via electrodes to eliminate interference from eye movement artifacts in the EEG signal analysis. Dataset KUL: Sixteen healthy subjects were selected, all of whom were asked to focus their attention on one of two speakers. The auditory stimulus consisted of four pieces of content spoken by different male speakers, presented at 90° angles to the left and right of the subjects' ears. At a sampling rate of 8192 Hz, 64 EEG channels were acquired and recorded. Each subject underwent 8 trials, with each trial lasting 6 minutes, resulting in a total of 12.8 hours of EEG signal data for the 16 subjects. Dataset FAT: Ten healthy subjects were selected and used a simulated driving test chamber to approximate a real driving environment, including two different driving tasks: low load and high load. In each driving task, each subject was required to perform a 2-hour driving task. In addition to one reference electrode, EEG signal data from 31 electrode channels can be obtained at a sampling frequency of 1000 Hz.

[0014] It should be noted that in some preferred embodiments, the raw EEG signal data acquired above is preprocessed to improve the accuracy of subsequent EEG signal decoding. For example, for the DSR dataset, a bandpass filter of 0.5 to 50 Hz is applied to remove low-frequency and high-frequency noise. The EOG channels can be removed using the automatic independent component analysis removal method in the magnetoencephalography and electroencephalography toolbox, i.e., retaining the last 28 EEG channels. The data is downsampled to 200 Hz. Only the first half of each attention test is used to balance the data of attention and inattention (rest). Each test is further subdivided into 2-second time windows with a 1-second overlap. Finally, a total of 689 shapes are obtained for each subject. Samples (time window) (Number of channels). For the KUL dataset, a finite impulse response (FIR) filter was used to bandpass filter the EEG signals, limiting the signals to a frequency range of 0.1 Hz to 50 Hz to eliminate irrelevant frequency components. The EEG signals were downsampled to 128 Hz to reduce the amount of data while retaining basic frequency information. Furthermore, Z-score normalization was applied to the signal of each EEG channel to ensure that the signal of each channel was adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1. The preprocessed EEG data was used to generate more EEG data segments through a 1-second sliding window. A fixed step-size strategy was adopted, selecting data points that occupied half of the time window at this point as the step size of the sliding window, i.e., an overlap of 50%. Finally, a total of 6264 shapes were obtained for each subject. The sample was obtained. For the FAT dataset, a bandpass filter was used to filter out the 1-60Hz range from the original data. Independent principal component analysis was used to process the data to remove noise and artifacts. This was done using the EEGLAB toolbox. Segmented EEG data were obtained through a 1-second sliding window with no overlap. Finally, a total of 600 shapes were obtained for each subject. The sample.

[0015] 102. Perform hierarchical processing on the temporal dynamic characteristics of the original EEG signal to extract the temporal features of each time window of the original EEG signal.

[0016] In this embodiment, a first EEG signal is obtained by enhancing the temporal and positional information of the original EEG signal with a learnable positional coding matrix. The original EEG signal is represented as... ,in Indicates the length of the time window. Let R represent the number of electrode channels and R be the set of real numbers. A learnable position encoding matrix is ​​used. The time and location information of the original EEG signal is enhanced, and the enhanced EEG signal is: , is represented as: ;in, Position encoding matrix Each line in , .in, Indicates time step The corresponding positional encoding vector. Positional encoding matrix. elements Indicates at time step The absolute position embedding. Each time step has an independent coded component. This time-location information enhancement operation helps distinguish the temporal sequence of signals by assigning unique location characteristics to different points in time.

[0017] The first EEG signal space is projected onto the query space, key space, and value space using a trainable parameter matrix. The enhanced first EEG signal... By projecting the trainable parameter matrix space onto the query space (Q), key space (K), and value space (V), we can focus on the wave value changes, specific content identifiers, and feature information of the signal in the current time window. The projected result is... .but , Represents the projection matrix. This represents the hidden layer dimension. It is represented as follows: ; ; .

[0018] The raw similarity score between different time steps within the same time window is calculated using the query space and key space. Specifically, the raw similarity score between different time steps within the same time window is calculated using the query space (Q) and key space (K) corresponding to the same time window. The calculation formula is as follows: ;in, express The dimension; Indicates the target time window number The query space corresponding to each time step. Indicates the target time window number The key space corresponding to each time step. This indicates the length of the target time window. In this embodiment, the dot product of the query space (Q) and key space (K) corresponding to the same time window is used to calculate the length of different time steps. and The original similarity score between them, divided by the result of the dot product. The value is used to normalize the variance of the dot product to 1, avoiding errors in subsequent operations. When calculating attention weights during operation, if the input value is too large, the weight distribution will be biased towards extreme values.

[0019] The original similarity score is converted into temporal feature-aware weights for different time windows of the original EEG signal. The original similarity score obtained above is then used as the basis for further calculation. pass Convert to corresponding time feature-aware weights The time feature perception weight This reflects the correlation strength between different time steps. The transformation formula is as follows: Here, the function exp(x) represents the natural exponential function, that is, the exponential function with the real number e (e≈2.71828) as its base, and its expression is: Based on the above formula, the time feature perception weights corresponding to the original similarity scores between different time steps in all time windows of the original EEG signal can be obtained. The time feature perception weights reflect the optimal allocation of the original EEG signal in terms of time structure.

[0020] The temporal features of each time window of the original EEG signal are determined by multiple sensing heads based on the temporal feature sensing weights. To overcome the limitations of single-feature learning to capture multiple time-series features present in the original EEG signal, this application embodiment introduces a method that includes... A collaborative sensing mechanism for multiple sensing heads. The number of channels C in the original EEG signal determines the spatial structure characteristics of the original EEG signal data, providing a foundation for subsequent analysis. Among them, the... Each sensing head determines the temporal characteristics of each time window of the original EEG signal based on the time feature sensing weights, and uses the time feature sensing weights... The temporal features of the current time window are obtained by performing a dot product with the trainable matrix V, and the calculation formula is as follows: ;in, Indicates the first Each sensing head senses weights based on time features. Determine the timing characteristics of the original EEG signal within the current time window.

[0021] 103. The extracted time-series features of each time window are fused to obtain the time-series collaborative feature matrix of the original EEG signal.

[0022] The time-series features extracted from each time window are fused to obtain the time-series fused features of the original EEG signal. As described above, the features include... The collaborative sensing mechanism of the sensing heads concatenates and fuses the temporal features extracted by all sensing heads, as shown below: ;in, This indicates a matrix concatenation operation, which results in a matrix with dimensions equal to the concatenated matrix. This represents the temporal fusion characteristics after fusion.

[0023] A threshold-based nonlinear transformation is applied to the aforementioned temporal fusion features to obtain temporal transformation features. To further enhance the splicing and fusion effect of the aforementioned temporal fusion features, embodiments of this application further refine the temporal fusion features... A nonlinear transformation with a threshold is performed to obtain the time-series transformation characteristics. The transformation formula is as follows: ; where the formula This indicates that negatively correlated feature information is removed from the fused feature matrix and set to 0; This represents the first learnable weight matrix, which assigns weights to temporal fusion features based on feedback during training. This indicates the bias term.

[0024] The timing transformation feature is obtained by concatenating the timing transformation feature with the original EEG signal using residuals. The timing transformation feature is obtained by concatenating the timing transformation feature using residuals. The timing connection characteristics are obtained by comparing the original EEG signal with the original signal. This preserves the original spatial structure information of the original EEG signal, which helps maintain signal stability and the original spatial structure, ensuring that the information of the original EEG signal is not lost, as shown below: The temporal connectivity features are standardized and calibrated to obtain the temporal co-feature matrix of the original EEG signal. Stability is maintained to adapt to various temporal feature signals through a normalization mechanism, such as BatchNorm. ;in, and Representing temporal connectivity features The mean and standard deviation; and These represent the learnable scaling and offset parameters, respectively. Indicates element-wise multiplication. This represents a very small positive number, used to prevent division by zero errors during calculations and to ensure that features are stable and effective during propagation without being excessively amplified. Therefore, this application's embodiments introduce a normalization mechanism to ensure effective feature propagation while suppressing excessive amplification.

[0025] After the above feature extraction steps, all temporal collaborative features obtained will be... The original EEG signal is represented using a time-coordinated feature matrix, which is also used in this embodiment. To indicate, This approach preserves the original spatial structure information of the original EEG signal. In this embodiment, considering the complex temporal dynamics of the original EEG signal, a global feature representation is constructed by learning the intrinsic correlations between different time points in the original EEG signal, thereby accurately extracting long-term dependencies and subtle pattern changes in the EEG signal. Furthermore, a multi-sensor collaborative sensing mechanism is introduced, fusing multiple temporal features extracted from each time window. Through residual connections and standardized feature calibration, the original spatial structure information of the original EEG signal is preserved, which helps maintain signal stability and the original spatial structure.

[0026] 104. Based on the aforementioned time-series collaborative feature matrix, construct the spatial graph structure of the original EEG signal.

[0027] The temporal collaborative feature matrix is ​​then subjected to scale normalization. For the first... Temporal Coordination Features of Layers Scale normalization is performed to unify the scale and enhance the stability of similarity calculation between channels, expressed as: ;in, Indicates the first Temporal Coordination Features of Layers The time-series collaborative features after scale normalization are scaled down to a unit scale to balance the signal strength in different spatial channels. This represents a very small positive number, used for numerical stability and to prevent division by zero errors during calculations. Indicates the first The number of spatial channels within a time window in a layer; This represents an index variable used to iterate through each channel during summation. From 1 to Loop, calculate the first... The sum of squares of the temporal co-features of each channel in each time window of the layer. (Symbol) Indicates the first The first layer of the temporal collaborative feature matrix List.

[0028] Based on the scale-normalized temporal co-feature matrix, a similarity matrix is ​​calculated between each channel of the original EEG signal; this similarity matrix characterizes the adjacency relationships of the spatial graph structure of the original EEG signal. The similarity matrix between each channel is calculated... To construct the spatial graph structure of the original EEG signal: The matrix multiplication results reflect the correlation strength between different electrode channels. It should be noted that in multilayer designs, features for brain state discrimination can be extracted progressively; specifically, if... Then use The output of the first The computation of the temporal collaborative features of a layer depends on the previous layer, i.e., it uses the previous layer. Features; if Then use The output directly uses the input features, meaning the features remain unchanged: .

[0029] 105. Perform multi-level dynamic graph convolution operation on the spatial graph structure to generate a multi-level spatial feature tensor for characterizing the spatial dependencies of the original EEG signal.

[0030] The degree values ​​of each channel of the original EEG signal are calculated based on the similarity matrix, and a diagonal matrix is ​​generated based on the degree values ​​of each channel. The similarity matrix is ​​then converted into a normalized adjacency matrix suitable for graph convolution using the diagonal matrix. Calculate the degree value for each channel. , ;in, Indicates the first and the The similarity between channels. Based on the degree values ​​of all channels. Generate a diagonal matrix To adapt to graph convolution operations, the similarity matrix in this embodiment is... Convert to normalized adjacency matrix , ;in, Representing the inverse square root of a diagonal matrix by using its diagonal elements This constructs a mechanism to mitigate gradient explosion or feature bias issues in graph convolution. .

[0031] Based on the normalized adjacency matrix and graph convolution operator, each layer of temporal collaborative features in the temporal collaborative feature matrix is ​​iteratively refined to obtain a multi-layer spatial feature tensor; wherein, the multi-layer spatial feature tensor is used to characterize the spatial dependencies of the original EEG signal. The dynamic graph convolution process involves multi-layer convolution; therefore, based on the aforementioned normalized adjacency matrix... Each layer iteratively refines the extracted temporal collaborative features using graph convolution operators: ;in, This represents the second weight matrix that can be learned, which assigns weights to temporal collaborative features based on feedback during training. Indicates the bias term; It is the Sigmoid activation function, used to enhance the nonlinear fitting ability after graph convolution calculation. Its formula is: Finally, based on the set number of dynamic brain region exploration layers... It preserves the spatial structural features at different levels and ultimately outputs a multi-layer spatial feature tensor. , The number of exploration layers .

[0032] This application embodiment generates a high-level spatial representation for brain state discrimination through normalization processing, spatial graph structure construction, and multi-layer dynamic graph convolution. It generates a multi-layer spatial feature tensor to characterize the spatial dependencies of the original EEG signal. This multi-layer spatial feature tensor reflects the spatial dependencies from low to high levels in the EEG signal, providing rich spatial topological features for subsequent cross-level fusion of spatial features, which is beneficial to improving the accuracy of brain state discrimination.

[0033] 106. Stack the multi-layer spatial feature tensors according to the hierarchical dimensions of spatial decomposition to form a three-dimensional spatial feature tensor.

[0034] To integrate the aforementioned multi-level spatial features across different layers, embodiments of this application use the generated multi-level spatial feature tensor... Stacked hierarchically, they form a three-dimensional tensor of spatial features. It is mainly used to adapt the timing input format of the gated recurrent unit (GRU): , This represents the stacking operation of tensors. (Through...) The stacking operation preserves the node structure of spatial features of each layer, while treating the evolution between layers as an ordered process, so that GRU can gradually integrate spatial features in a hierarchical order.

[0035] 107. The three-dimensional tensor of the spatial features is fused across levels using a gated loop unit to obtain cross-level spatial fusion features.

[0036] As described above, in order to generate a unified representation of the aforementioned multi-level spatial features, this embodiment of the application introduces a GRU gating mechanism to enhance the consistency and discriminativeness of the multi-level spatial features, while preserving the discriminativeness of the relative spatial relationships between channels. The specific implementation is as follows: the three-dimensional tensor of the aforementioned stacked spatial features... Input the spatial features of each layer into the GRU in hierarchical order. As a time step input of the GRU, the spatial feature information of each level is dynamically integrated using the GRU's update gate and reset gate; wherein, the reset gate determines whether to ignore the spatial features of the previous level: The update gate determines the ratio of retaining spatial features from the previous layer to retaining spatial features from the current layer. Based on the output of the reset gate, candidate spatial features for the current layer are generated. : The ratio of the candidate spatial features to the spatial features output by the update gate from the previous layer is used to generate cross-layer spatial fusion features. : .in, The Sigmoid activation function has the following formula: tanh is the tangent function, and its formula is: , appearing These represent the weight matrix and bias terms in the network layers. The final output of the GRU shows cross-layer spatial fusion features. It integrates all spatial feature information from shallow to deep layers, captures abstract hierarchical dependencies, enhances the consistency and discriminability of cross-level spatial features, and preserves the discriminability of relative spatial relationships between channels.

[0037] 108. Calculate the discriminant value of different complex brain states based on the cross-level spatial fusion features.

[0038] The extracted cross-level spatial fusion features The input is passed to a fully connected layer for discrimination. This fully connected layer consists of a feature weight matrix. With bias term This fully connected layer is composed of components that calculate the discriminative value for the current brain state type. : ;in, This indicates the discriminant value of the raw EEG data belonging to two types of brain states. Represents the fused feature vector. The Sigmoid activation function has the following formula: , It is a discriminative value for different categories, such as cognitive attention, auditory attention, and driving fatigue. This discriminative value is used for subsequent model training in an environment that simulates the brain's cognitive decision-making circuits, thereby improving the accuracy and generalization ability of brain state discrimination.

[0039] 109. By simulating the brain's cognitive decision-making circuits, and based on the discriminative value, the type of complex brain state is determined.

[0040] Traditional deep learning state decoders simply input the network's output into a fully connected layer to obtain a classification result. The network parameters are then updated through gradient iteration based on the classification result and the true labels. However, relying solely on parameters fed back from the labels limits the network's robustness to multi-state learning. Therefore, this application's embodiments optimize the cognitive decoding strategy for raw EEG signals by simulating the dynamic feedback process of brain activity circuits, revealing more discriminative pattern representations. The specific implementation is as follows: Based on the discriminative value and greedy strategy, the optimal discriminative behavior is determined. The current EEG sample state is defined as... The sample state at the next moment is defined as This reflects the continuous physiological processes of EEG signals, for example, The discriminative value based on the current brain state type The optimal discrimination behavior is determined by using a greedy strategy: ;in, Represents a set of behaviors; Indicates from the set of behaviors Choose a target to discriminate behavior, such that the discriminative value of the current brain state type is [missing information]. The action that maximizes the discrimination value is considered the optimal discrimination action, such as discrimination types like cognitive attention, auditory attention, and driving fatigue. During training, the model interacts with the brain environment through a pre-designed neural feedback mechanism, continuously adjusting its strategy to improve state discrimination accuracy. Then, based on the optimal discrimination action and the discrimination value, the discrimination type for complex brain states is determined.

[0041] The training method for model back-optimization is described in detail below: To optimize model decision-making, this embodiment also incorporates a neural feedback reward mechanism. The matching degree between the discriminant type and the corresponding true label type is calculated; a reward / penalty signal is generated based on the matching degree. The calculation formula is as follows: Where sgn is the sign function, and That is, a positive match is 1, and a negative match is -1; This indicates a deterministic choice of the optimal decision-making action; Indicates the true label; This represents the indicator function, which is 1 for a match and 0 otherwise. This reward mechanism provides feedback on the model's actual cognitive state using EEG signals immediately after the model makes a judgment, gradually shifting it from random exploration to selecting the correct judgment strategy. This frequent feedback helps the model converge to the optimal strategy more quickly. Simultaneously, the environmental termination condition stops when the signal input ends.

[0042] By simulating the dynamic feedback process of brain activity circuits, the reward and punishment signals are used to dynamically optimize the discrimination behavior of complex brain states. During the training process, this embodiment employs... The strategy allows the model to explore and find usable features as freely as possible in the early stages of training, and then switches to a strategy to fit effective features and make state decisions at a certain stage of training. The expected behavior of this strategy is... It is expressed as follows: ; ; .

[0043] in, The value represents the set exploration rate. This represents the decay rate of free exploration; n is the number of training epochs, as the model focuses on policy fitting in the later stages of training. This represents the deterministic choice of the optimal decision-making action.

[0044] This application embodiment also introduces two discriminators: a policy discriminator. and target discriminator , and These represent the model utility used for policy and objective directions, respectively. The two network model architectures are identical, and when the state... When inputting and passing through the network, respectively, we get and That is, the discriminative value of the corresponding output. , The values ​​are backpropagable parameters in the model. First, the target discriminator... It is the expectation of decision-making value between timely rewards and future states. This can be represented as: ;in The probability of state discrimination; is the discount factor, representing the degree to which the strategy emphasizes future effects; a larger value indicates a greater focus on long-term judgments. E represents the integral of the uncertainty of the state discrimination probability. In the state... Execute action Then, the next case The discrimination is determined by the model This generates the probability for state discrimination. Expectation. right The average of all possible reward transfer outcomes (weighted by probability) is used to comprehensively evaluate the current action. The long-term value of policy discriminators. Policy discriminators are used to determine decision-making actions and obtain their corresponding discriminative values. parameters By minimizing the following loss function The value is used to update the algorithm, while the target discriminator is used to estimate the discriminative value of the brain state at future time steps. This guides the updating of the policy network. To prevent unexpected outliers from occurring during training. The value exhibits sensitive fluctuations, addressing the issue of... Use the following constraints: When the loss value is less than 1, the loss function... The update rule for parameter w is shown in the following formula. The learning rate set for training.

[0045] ; To prevent the target network The rapid updating of intermediate parameters caused the target value to oscillate. To address this, an update interval C was set, where the policy network was updated every training step (a multiple of C). With the target network Synchronization: It should be noted that, in this application, the embodiments introduce a priority experience replay array based on the memory consolidation interaction between the hippocampus and the cerebral cortex in the brain's physiological mechanisms. Unlike the random sampling in typical reinforcement learning environments, which is similar to the short-term storage of contextual memory in the hippocampus, the model will determine the state and its outcome each time. Stored in a memory buffer. Based on the modulating effect of arousal level on memory encoding strength, i.e., the error between the current state judgment and the actual situation. The formula for this judgment is as follows: ;in, For each empirical segment, a priority is assigned based on the error, calculated as the absolute value. During training, sampling is performed according to this priority probability. This mechanism enables the model to learn EEG feature representations that contain more information more effectively, while preventing over-focusing on outliers.

[0046] In summary, this embodiment of the application extracts temporal features of each time window of the original EEG signal by performing hierarchical processing on the temporal dynamic characteristics of the original EEG signal; it then fuses these temporal features to obtain a temporal collaborative feature matrix of the original EEG signal, thus completing the hierarchical extraction of temporal features; based on the temporal collaborative feature matrix, it constructs a spatial graph structure of the original EEG signal, performs multi-level dynamic graph convolution operations on the spatial graph structure, and generates a multi-level spatial feature tensor to characterize the spatial dependencies of the original EEG signal, thus completing the hierarchical extraction of spatial features. This multi-level spatial feature tensor reflects the spatial dependencies from low to high levels in the EEG signal, providing rich spatial topological features for subsequent cross-level fusion of spatial features, which is beneficial to improving the accuracy of brain state discrimination; finally, the multi-level spatial feature tensor is stacked according to the hierarchical dimensions of spatial decomposition to form a spatial feature tensor. This paper describes a method for representing a three-dimensional tensor of spatial features. A gated recurrent unit is used to perform cross-level feature fusion on the three-dimensional tensor, resulting in cross-level spatial fusion features. This completes the cross-level feature fusion of temporal and spatial features of EEG signals, achieving a comprehensive capture of the spatiotemporal dynamic characteristics of EEG signals. This lays a crucial feature foundation for simulating the brain's cognitive decision-making circuit and realizing a comprehensive and accurate representation of complex brain states. Finally, the discriminative value of different complex brain states is calculated based on the cross-level spatial fusion features. By simulating the brain's cognitive decision-making circuit, the type of complex brain state is determined, achieving a comprehensive and accurate representation of complex brain states. Simultaneously, the discriminative value provides a core foundation for simulating the brain's cognitive decision-making circuit, and the accuracy of brain state discrimination can be continuously optimized through reinforcement learning strategies. This application solves the technical problems of incomplete and inaccurate representation of complex cognitive states in existing brain technologies.

[0047] The following is a detailed explanation of the technical effects achieved by the method for discriminating complex brain states described in this application, using a set of experimental data: Experimental setup: For individual-specific EEG experiments, a 10-fold cross-validation method was used for evaluation. In individual-specific experiments, both training and test data came from the same individual. To avoid potential data leakage problems caused by improper random shuffling in individual-specific experiments, shuffling was performed on different conversational experimental segments rather than directly shuffling all data. For continuous cognitive processes in the brain, such as attention and auditory discrimination, adjacent data segments in a trial are highly correlated. If these segments are randomly shuffled before the training-test split, highly correlated segments will appear in both the training and test data. Therefore, very high classification results are observed. However, in reality, when the model has never seen these highly correlated segments, the accuracy (ACC) will decrease. For the attention and auditory datasets, each trial was split into shorter segments. The shuffling of intra-session stages ensures that highly correlated segments within a single trial do not appear simultaneously in both the training and test data in a cross-validation fold. Nested cross-validation is used to avoid biased evaluation. The average ACC and F1 score (F1-Score) of all subjects were used as the final evaluation criteria. The F1-Score is an evaluation metric used in deep learning for classification models, and it is an important standard for testing the true negative example prediction distribution of the data. Its range is [0,1], with a value closer to 1 indicating greater effectiveness in classifying different categories. All experiments were conducted on an NVIDIA GTX1660 SUPER, and the obtained data were used for 10-fold cross-validation. Mean Squared Error (MSE) was chosen as the objective function to guide the training process. MSE is the most commonly used loss function in supervised learning, and it naturally corresponds to the squared form of the Bellman error in Q-learning. Optimizing MSE is equivalent to minimizing the expected value of the Bellman error. The maximum number of training epochs was 200. The Adam optimizer was used to optimize the training process, with an initial learning rate of 7e-4 and a weight decay of 1e-4. The eps parameter was used and set to 1e-7, a small constant added to the denominator to avoid division-by-zero errors during the update process. For reinforcement learning training, the initial exploration rate epsilon was set to 1 for fully free exploration, the exploration rate decay was set to 0.7, and the lowest exploration rate was 1e-3. During training, the policy was trained from fully free exploration to the discriminative purpose of the adapted policy. The discount factor was set to 0.9 to maximize the calculation of the overall policy. Gradient clipping was applied during training to address the gradient explosion problem. Regarding the module design, the size of the linear hidden layer and the intermediate layer of the multi-layer GCN in the temporal feature extraction module was set to 32. The number of deep brain space exploration layers was set to 3, a result obtained from multiple experiments (Table 1).Additionally, the pooling size in the model was adjusted based on the shape of the input data. Aside from this, the hyperparameter settings were identical for each subject across all datasets.

[0048] Table 1 Test experiments on exploration layer 1 in the ABSM module

[0049] The experimental comparison methods include: (1) 3D-CNN: a 3D convolutional neural network model based on cognitive attention decoding technology, which has cascaded and parallel convolution operations; BRHNet: a hybrid deep learning model based on wavelet transform to obtain time-frequency domain information of EEG signals; HEMAsNet: a dual-hemispheric temporal feature extraction network that uses a combination of multi-scale convolutional neural networks and long short-term memory blocks to extract the temporal features of the two hemispheres of the brain to analyze and explore cognitive recognition tasks; SCNN-TransE: a cognitive attention evaluation model based on stacked convolutional neural networks and Transformer encoders. LGNet: A local-global dual-mode fusion network that evaluates the cognitive workload of a simulated flight process through a local and global network. It fuses convolutional layers and TransformerLayer, and designs a unique SCCE loss, combining supervised contrastive loss with traditional cross-entropy loss to extract local and global representations from EEG signals. DMAEEG: Introduces a robust EEG cognitive workload recognition model based on self-supervised learning. It combines training strategies of denoising autoencoders and masking autoencoders, and uses temporal convolutional networks and multi-head self-attention mechanisms as the backbone to effectively capture the spatiotemporal features of EEG.

[0050] (2) Decoding techniques based on driver fatigue: CWSTR-Net: a channel-weighted spatiotemporal residual network for fatigue detection. This network combines convolutional neural networks and recurrent neural networks to extract spatiotemporal features from EEG signals and introduces an unsupervised channel-weighted algorithm based on non-smooth non-negative matrix factorization to identify active brain channels; FD-LiteNet: a lightweight automatic fatigue detection search network that uses neural architecture search methods to automatically design lightweight and high-performance convolutional neural network models for fatigue detection; RMSCM: a multi-branch deep learning network based on brain region segmentation strategy, which effectively extracts highly discriminative long-term and short-term feature information and base... A method for weighting the importance of edges in a spatial graph for brain region segmentation; CSF-GTNet: A novel time-frequency domain-based multidimensional feature fusion network for fatigue detection, whose model structure consists of a Gaussian time-domain network and a pure convolutional spatial-frequency domain network; HATNAT: A transfer learning model based on a hybrid attention mechanism, which uses a calibration-based transfer learning strategy to capture local and global temporal patterns using hybrid attention modules, and establishes spatial representations between EEG signal channels through channel attention modules; E-key: A unified framework for dual-task attention convolution, a unified framework structure that combines convolutional neural networks and attention structures to simultaneously perform personal identification and fatigue detection.

[0051] (3) Decoding techniques based on auditory attention: SSF-CNN: a dual-domain feature fusion convolutional neural network, which extracts and combines spectral and spatial features of the EEG alpha band for auditory spatial attention classification; BIA-Net: a cross-modal neural network model based on fast Fourier transform, multi-band frequency embedding, and EEG signal and speech information, used for auditory attention classification; MBSSFCC: a multi-band entropy spectrum spatiotemporal fusion convolutional network, which analyzes multiple frequency bands of EEG, extracts the differential entropy of each frequency band, and combines CNN and convolutional length Short-Term Memory (ConvLSTM) for auditory attention classification; DGSD: Dynamic Graph Convolutional Spatial Decoding Network, a method that uses Dynamic Graph Convolutional Networks (DGCN) to represent the graph structure of EEG signals and extract important features related to auditory spatial attention; STAnet: An end-to-end spatiotemporal attention network for EEG auditory spatial attention detection; ST-GCN: Spatial Topology Graph Convolutional Network, a method for learning complex spatial patterns in multi-channel EEG signals based on the robustness of graph convolutional networks for auditory attention classification.

[0052] (4) Multi-task decoding techniques DeepCNN: Deep convolutional neural network, based on multiple CNN modules, uses convolution to simulate bandpass filtering and spatial filtering for decoding EEG signals; EEGNet: Interpretable convolutional neural network, a novel convolutional neural network that uses separable convolution to perform spatiotemporal sequence processing on EEG signals for decoding; LGG_Net: Multi-scale convolutional neural network, which consists of a series of temporal convolutions with multi-scale one-dimensional convolutional kernels and kernel-level attention fusion. It captures the temporal dynamics of EEG and then uses them as input to the proposed local and global graph filtering layers for decoding EEG signals; SBLEST: End-to-end EEG decoding algorithm, which uses a low-rank weight matrix to encode spatiotemporal filters and classifiers and optimizes them under the principle sparse Bayesian learning framework; LAG: Proposes a strategy called Local Ascend-Global Learning, which starts from the local connectivity of a single brain functional region and develops a k-level adaptive boosting network to dynamically capture strong connectivity patterns between brain regions in different cognitive tasks.

[0053] (5) ours: A method for discriminating complex brain states as described in this application

[0054] As shown in Tables 2, 3, and 4, the complex brain state discrimination method proposed in this application demonstrates significant performance advantages in three representative EEG decoding tasks, highlighting its high efficiency in neural state decoding. Specifically, Table 2 shows the outstanding performance in cognitive attention decoding on the DSR dataset, with an mACC of 71.23% and an mF1-Score of 70.54%. This result is a 1.17% improvement over the 3D-CNN method and significantly higher than other methods such as BRHNet and SCNN-Trans. This is because the embodiments of this application can explore the brain's spatial topology at multiple layers, enhancing the extraction of neural representations of attention states. Similarly, in decoding driving fatigue on the FAT dataset, the mACC is 99.23% and the mF1-Score is 99.16%, showing stable decoding accuracy and high accuracy across different categories. This performance consistency stems from the neural feedback loop learning of the embodiments of this application, whose adaptive optimization strategy can respond in real time to the EEG feature shifts in fatigue states. Furthermore, the spatiotemporal fusion enhances the robust decoding of fatigue neural mechanisms. Table 4 shows that on the KUL dataset, for auditory attention decoding, the mACC reaches 92.09% and the mF1-Score is 91.25%. This represents a 1.79% improvement compared to other methods such as DGSD, and outperforms other state-of-the-art methods, particularly SSF-CNN and BIA-Net. This is because the embodiments of this application enhance the consistency and discriminativeness of cross-level spatial features by introducing a GRU gating mechanism, while preserving the discriminativeness of the relative spatial relationships between channels. Although the embodiments of this application, like DGSD, explore spatial structure information, the embodiments of this application effectively fuse cross-level multi-layer spatial features into multi-layer spatial features under auditory stimulation, thereby enhancing the discriminative ability of key neural states through the fusion of multi-layer spatial structures.

[0055] Table 2. Performance comparison of ours and other cognitive attention decoding techniques on the DSR dataset.

[0056] Table 3. Performance comparison of ours and other driver fatigue decoding technologies on the FAT dataset.

[0057] Table 4. Performance comparison of ours and other auditory attention decoding techniques on the KUL dataset.

[0058] Table 5 shows a quantitative comparison of the performance differences between the embodiments of this application and other advanced multi-task decoding techniques on the DSR, FAT, and KUL datasets. The results demonstrate the comprehensive advantages of the embodiments of this application in multi-task decoding. Specifically, in the DSR attention task, the mACC is 71.23%, which is 1.8% higher than LAG and 2.74% higher than DeepCNN. This is because the embodiments of this application enhance state representation extraction by deeply analyzing the spatial structure of brain regions in EEG signals, while LAG focuses on local-to-global brain region connectivity. In the FAT fatigue task, the mACC reaches 99.23%, which is 1.42% higher than LGG_Net. Unlike the LGG_Net method, which focuses on spatiotemporal structure decoding, the embodiments of this application simulate the state strategy of brain activity circuits and adjust the decoding strategy in real time to better respond to fatigue features. In the KUL auditory task, the mACC reaches 92.09%, significantly better than others such as LAG and LGG_Net, with an average improvement of 1.82%, highlighting the ability of the embodiments of this application to accurately model the spatiotemporal patterns of auditory attention states. This improvement in consistency differs from the SBLEST data space transformation method. The embodiments of this application optimize the generalization ability of EEG signal multi-state decoding by fusing the temporal and spatial features of EEG signals at multiple levels.

[0059] Table 5 compares the performance of ours with other multi-task decoding techniques on the datasets DSR, KUL, and FAT.

[0060] Finally, it should be noted that in Tables 2, 3, 4, and 5 above, mACC and mF1-Score represent the average recognition accuracy and average F1 score of all participants in the dataset, respectively. Methods marked with an asterisk (*) indicate that the corresponding data was obtained by reproducing the method using the corresponding parameters; methods without an asterisk indicate that the corresponding data is directly cited from existing literature.

[0061] As shown in Figures 2, 3, and 4, Figure 2 is a radar chart comparing individual accuracy on the DSR dataset provided by the embodiments of this application; Figure 3 is a radar chart comparing individual accuracy on the FAT dataset provided by the embodiments of this application; and Figure 4 is a radar chart comparing individual accuracy on the KUL dataset provided by the embodiments of this application. These figures demonstrate the classification results of the embodiments of this application and various advanced multi-task decoding techniques on the DSR, FAT, and KUL datasets for each subject, showing a comparison of the global trends for each subject. The axis labels represent independent subject numbers, the ray length reflects the classification accuracy, and the lower right corner indicates the average accuracy and standard deviation of the embodiments of this application. Overall, the embodiments of this application outperform other methods in most subjects. Specifically, as shown in Figure 2, on the DSR dataset, the mACC is 71.23% (Std=4.75%), and 14 out of 26 subjects have accuracy significantly better than the benchmark method; furthermore, subjects S11 and S22 show general recognition difficulties in each method, which is consistent with the description in the background art. As shown in Figure 3, the embodiment of this application exhibits the best and most stable performance on FAT (mACC = 99.23% Std = 1.24%), with all subjects achieving an accuracy greater than 96%. As shown in Figure 4, on the KUL dataset, the mACC is 92.09% (Std = 4.50%). The embodiment of this application performed excellently in 12 out of 16 subjects' auditory attention recognition tests (75%), with an average accuracy improvement of approximately 2%. In summary, the embodiment of this application demonstrates good performance on most subjects across different datasets, proving its ability to effectively extract key features to determine cognitive states.

[0062] In summary, this application proposes a method for discriminating complex brain states. By introducing a brain-inspired cognitive learning strategy and integrating spatiotemporally multi-level dynamic fusion of EEG signals, it learns the brain activities in various functional regions during cognitive processes and the complex relationships between them. Cross-validation experiments were conducted on three publicly available benchmark datasets: attention, fatigue, and auditory attention. The results show that the method described in this application has higher accuracy and F1 score than other methods.

[0063] Example 2: The above description, in conjunction with Figures 1 to 4, details a method for identifying complex brain states provided in this application. The following description, in conjunction with Figure 5, details a system for identifying complex brain states used to execute the method provided in this application. Figure 5 is a schematic diagram of the structure of a system for identifying complex brain states provided in this application. Referring to Figure 5, the system includes the following: a signal acquisition module 501 for acquiring raw EEG signals; a temporal co-processing feature extraction module 502 for performing hierarchical processing on the temporal dynamic characteristics of the raw EEG signals, extracting temporal features of the raw EEG signals; fusing multiple temporal features to obtain a temporal co-processing feature matrix of the raw EEG signals; and a deep brain spatial exploration module 503 for constructing a spatial graph structure of the raw EEG signals based on the temporal co-processing feature matrix, and processing the spatial graph structure. The system performs multi-level dynamic graph convolution operations to generate multi-level spatial feature tensors that characterize the spatial dependencies of the original EEG signal. An advanced cross-level fusion module 504 stacks the multi-level spatial feature tensors according to the hierarchical dimensions of spatial decomposition into a three-dimensional spatial feature tensor. A gated recurrent unit is used to perform cross-level feature fusion on the three-dimensional spatial feature tensor to obtain cross-level spatial fusion features. The discriminant value for different complex brain states is calculated based on the cross-level spatial fusion features. A complex brain state discrimination module 505 simulates the brain's cognitive decision-making circuits and determines the type of complex brain state based on the discriminant value.

[0064] In a preferred embodiment, the temporal collaborative feature extraction module 502 is specifically used to: enhance the temporal location information of the original EEG signal with a learnable positional encoding matrix to obtain a first EEG signal; project the first EEG signal spatially to a query space, a key space, and a value space using a trainable parameter matrix; calculate the original similarity score between different time steps within the same time window using the query space and the key space; convert the original similarity score into temporal feature perception weights for each time window of the original EEG signal; and determine the temporal features of each time window of the original EEG signal based on the temporal feature perception weights using multiple sensing heads.

[0065] In a preferred embodiment, the time-series collaborative feature extraction module 502 is specifically used to: fuse multiple time-series features extracted from each time window to obtain the time-series fusion features of the original EEG signal; perform a threshold-based nonlinear transformation on the time-series fusion features to obtain time-series transformation features; connect the time-series transformation features and the original EEG signal using residuals to obtain time-series connection features; and perform standardization calibration on the time-series connection features to obtain the time-series collaborative feature matrix of the original EEG signal.

[0066] In a preferred embodiment, the deep brain spatial exploration module 503 is specifically used for: performing scale normalization on the temporal co-operation feature matrix; calculating the similarity matrix between each channel of the original EEG signal based on the scale-normalized temporal co-operation feature matrix; wherein the similarity matrix is ​​used to characterize the adjacency relationship of the spatial graph structure of the original EEG signal; calculating the degree value of each channel of the original EEG signal according to the similarity matrix, and generating a diagonal matrix based on the degree value of each channel; converting the similarity matrix into a normalized adjacency matrix suitable for graph convolution using the diagonal matrix; iteratively refining each layer of temporal co-operation features of the temporal co-operation feature matrix based on the normalized adjacency matrix and the graph convolution operator to obtain a multi-layer spatial feature tensor; wherein the multi-layer spatial feature tensor is used to characterize the spatial dependency relationship of the original EEG signal.

[0067] In a preferred embodiment, the advanced cross-level fusion module 504 is specifically used to: input the spatial feature three-dimensional tensor into a gated loop unit in hierarchical order, with each layer of spatial features serving as a time step input to the gated loop unit; dynamically integrate spatial feature information using the update gate and reset gate of the gated loop unit; wherein, the reset gate determines whether to ignore the spatial features of the previous layer, and the update gate determines the ratio of retaining the spatial features of the previous layer to retaining the spatial features of the current layer; generate candidate spatial features of the current layer based on the output of the reset gate; and fuse the candidate spatial features with the ratio of the spatial features of the previous layer output by the update gate to generate cross-level spatial fusion features.

[0068] In a preferred embodiment, the complex brain state discrimination module 505 is specifically used for: determining the optimal discrimination behavior based on the discrimination value and the greedy strategy; determining the discrimination type of the complex brain state according to the optimal discrimination behavior and the discrimination value; calculating the matching degree between the discrimination type and the corresponding real label type; generating a reward / punishment signal according to the matching degree; and dynamically optimizing the discrimination behavior of the complex brain state using the reward / punishment signal by simulating the dynamic feedback process of brain activity circuits.

[0069] The specific implementation method and technical effects of this complex brain state discrimination system can be referred to the aforementioned complex brain state discrimination method, and will not be repeated here.

[0070] Embodiment 3 This embodiment of the invention also provides an electronic device. Figure 6 is a structural diagram of an electronic device according to an embodiment of the invention. As shown in Figure 6, the electronic device includes: a central processing unit (CPU) 100, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 200 or a program loaded from a storage section 800 into a random access memory (RAM) 300. The RAM 300 also stores various programs and data required for system operation. The CPU 100, ROM 200, and RAM 300 are interconnected via a bus 400. An input / output (I / O) interface 500 is also connected to the bus 400. The following components are connected to the I / O interface 500: an input section 600 including a keyboard, mouse, etc.; an output section 700 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 800 including a hard disk, etc.; and a communication section 900 including a network interface card such as a LAN card, modem, etc. The communication section 900 performs communication processing via a network such as the Internet. Drive 1000 is also connected to I / O interface 500 as needed. Removable media 1100, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on drive 1000 as needed so that computer programs read from them can be installed into storage section 800 as needed.

[0071] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0072] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for identifying complex brain states, characterized in that, include: Acquire raw EEG signals; perform hierarchical processing on the temporal dynamic characteristics of the raw EEG signals, and extract the temporal features of each time window of the raw EEG signals; The extracted time-series features of each time window are fused to obtain the time-series co-feature matrix of the original EEG signal; Based on the temporal collaborative feature matrix, a spatial graph structure of the original EEG signal is constructed. Multi-level dynamic graph convolution operation is performed on the spatial graph structure to generate a multi-level spatial feature tensor that represents the spatial dependency of the original EEG signal. The multi-level spatial feature tensor is stacked according to the hierarchical dimensions of spatial decomposition to form a three-dimensional spatial feature tensor. A gated loop unit is used to perform cross-level feature fusion on the three-dimensional tensor of the spatial features to obtain cross-level spatial fusion features; the discriminative value of different complex brain states is calculated based on the cross-level spatial fusion features; the type of complex brain state is determined by simulating the brain's cognitive decision-making circuit and based on the discriminative value.

2. The method for discriminating complex brain states as described in claim 1, characterized in that, The temporal dynamic characteristics of the original EEG signal are processed in layers to extract the temporal features of each time window of the original EEG signal. This includes: enhancing the temporal position information of the original EEG signal with a learnable positional encoding matrix to obtain a first EEG signal; projecting the first EEG signal spatially to a query space, a key space, and a value space using a trainable parameter matrix; calculating the original similarity score between different time steps within the same time window using the query space and the key space; converting the original similarity score into temporal feature perception weights for each time window of the original EEG signal; and determining the temporal features of each time window of the original EEG signal based on the temporal feature perception weights using multiple sensing heads.

3. The method for discriminating complex brain states as described in claim 2, characterized in that, The extraction of time-series features from each time window is fused to obtain the time-series co-feature matrix of the original EEG signal. This includes: fusing multiple time-series features from each time window to obtain the time-series fusion features of the original EEG signal; performing a threshold-based nonlinear transformation on the time-series fusion features to obtain time-series transformation features; using residual connections to connect the time-series transformation features with the original EEG signal to obtain time-series connection features; and performing normalization calibration on the time-series connection features to obtain the time-series co-feature matrix of the original EEG signal.

4. The method for discriminating complex brain states as described in claim 3, characterized in that, Based on the temporal co-operational feature matrix, a spatial graph structure of the original EEG signal is constructed. Multi-level dynamic graph convolution operations are performed on the spatial graph structure to generate a multi-layer spatial feature tensor representing the spatial dependencies of the EEG signal. This includes: scaling the temporal co-operational feature matrix; calculating the similarity matrix between channels of the original EEG signal based on the scale-normalized temporal co-operational feature matrix; wherein the similarity matrix represents the adjacency relationship of the spatial graph structure of the original EEG signal; calculating the degree value of each channel of the original EEG signal according to the similarity matrix, and generating a diagonal matrix based on the degree values ​​of each channel; converting the similarity matrix into a normalized adjacency matrix suitable for graph convolution using the diagonal matrix; and iteratively refining each layer of temporal co-operational features of the temporal co-operational feature matrix based on the normalized adjacency matrix and the graph convolution operator to obtain a multi-layer spatial feature tensor; wherein the multi-layer spatial feature tensor represents the spatial dependencies of the original EEG signal.

5. The method for discriminating complex brain states as described in claim 4, characterized in that, A gated loop unit is used to perform cross-level feature fusion on the three-dimensional tensor of the spatial features to obtain cross-level spatial fusion features. This includes: inputting the three-dimensional tensor of the spatial features into the gated loop unit in hierarchical order, with each layer of spatial features serving as a time step input to the gated loop unit; dynamically integrating spatial feature information using the update gate and reset gate of the gated loop unit; wherein the reset gate determines whether to ignore the spatial features of the previous layer, and the update gate determines the ratio of retaining the spatial features of the previous layer to retaining the spatial features of the current layer; generating candidate spatial features for the current layer based on the output of the reset gate; and fusing the candidate spatial features with the ratio of the spatial features of the previous layer output by the update gate to generate cross-level spatial fusion features.

6. A method for discriminating complex brain states as described in any one of claims 1 to 5, characterized in that, By simulating the brain's cognitive decision-making circuits and determining the type of complex brain states based on the discriminative value, the method includes: determining the optimal discriminative behavior based on the discriminative value and a greedy strategy; determining the discriminative type of the complex brain state based on the optimal discriminative behavior and the discriminative value; calculating the matching degree between the discriminative type and the corresponding real label type; generating a reward / punishment signal based on the matching degree; and dynamically optimizing the discriminative behavior of the complex brain state using the reward / punishment signal by simulating the dynamic feedback process of the brain activity circuits.

7. A system for discriminating complex brain states, characterized in that, include: The signal acquisition module is used to acquire raw EEG signals; The temporal collaborative feature extraction module is used to perform hierarchical processing on the temporal dynamic characteristics of the original EEG signal, extract the temporal features of the original EEG signal, and fuse multiple temporal features to obtain the temporal collaborative feature matrix of the original EEG signal. The deep brain spatial exploration module constructs a spatial graph structure of the original EEG signal based on the temporal collaborative feature matrix, performs multi-level dynamic graph convolution operation on the spatial graph structure, and generates a multi-level spatial feature tensor to characterize the spatial dependencies of the original EEG signal; the advanced cross-level fusion module is used to stack the multi-level spatial feature tensor into a three-dimensional spatial feature tensor according to the hierarchical dimensions of spatial decomposition. A gated loop unit is used to perform cross-level feature fusion on the three-dimensional tensor of the spatial features to obtain cross-level spatial fusion features; the discriminative value of different complex brain states is calculated based on the cross-level spatial fusion features. A complex brain state discrimination module is used to determine the type of complex brain state by simulating the brain's cognitive decision-making circuit and based on the discrimination value.

8. A computer-readable storage medium, characterized in that, The computer program present in the computer-readable storage medium, when executed by a computer device, is used to implement the discrimination method as described in any one of claims 1 to 6.

Images