A method for EEG intention recognition based on the representation of functional connectivity entropy of unbalanced brain regions

Abstract

This invention relates to the fields of brain-computer interface and neural signal processing technology, and provides a method for EEG intention recognition based on the representation of non-equilibrium brain region functional connectivity entropy. The method includes: acquiring and preprocessing multi-channel EEG signals, aggregating them into brain region-level EEG signals according to a preset mapping; calculating functional connectivity between brain regions within a time window and constructing a connectivity strength matrix; calculating the functional connectivity entropy of each brain region based on the connectivity strength to quantify the complexity of connectivity information; further, extracting the non-equilibrium distribution characteristics of the functional connectivity entropy in the brain region space, including global statistics, local spatial differences, and concentration measures; finally, performing intention recognition based on this low-dimensional feature vector. This invention transforms functional connectivity from a strength representation to a robust entropy representation and innovatively constructs its spatial distribution structure, achieving high accuracy, high robustness, and low computational complexity in EEG intention recognition, suitable for online brain-computer interface systems.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and neural signal processing technology, and particularly relates to a brainwave intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions. Background Technology

[0002] EEG intention recognition is a key component of brain-computer interface technology, aiming to decode the user's intention from EEG signals. Existing methods mainly fall into two categories:

[0003] The first type of method is based on the time-domain, frequency-domain, or time-frequency-domain features of EEG signals for identification. While these methods are simple, they only characterize local EEG activity and are insufficient to represent the multi-brain region coordination mechanisms necessary to achieve complex intentions, resulting in limited feature discrimination ability and insufficient stability.

[0004] The second type of method introduces functional connectivity analysis, constructing functional connectivity networks by calculating indicators such as the correlation between brain region signals. Although this method can reflect brain region coordination, it still has obvious drawbacks: first, the connectivity strength values ​​are easily affected by noise, transient fluctuations, and individual differences, resulting in poor feature stability; second, the high-dimensional connectivity matrix requires complex model processing, which is computationally burdensome and not conducive to online deployment.

[0005] In summary, existing technologies face a dual bottleneck: first, the robustness of existing features is insufficient; second, they fail to effectively utilize the spatial distribution information of functional connections in brain regions. Different intentions often correspond to different spatial patterns of brain region coordination (such as concentration or dispersion), but existing methods lack effective characterization of this. Therefore, there is an urgent need for a new EEG intention recognition method that can balance feature stability, low-dimensional representation ability, and spatial structure characterization ability. Summary of the Invention

[0006] The purpose of this invention is to provide a brainwave intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions, aiming to solve the problems mentioned in the background art.

[0007] The present invention is implemented as follows: a brainwave intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions includes the following steps:

[0008] Step 1: Acquire multi-channel EEG signals and map the channel set to the target brain region according to a preset brain region division rule. The brain regions are distinguished from each other, and the brain region-level EEG signals of each brain region are obtained by aggregation and then standardized.

[0009] Step 2: Within each time window, based on the standardized brain region-level EEG signals, calculate the functional connectivity values ​​between any two brain regions, map them to non-negative connectivity strength, and construct... The brain region-level connectivity strength matrix;

[0010] Step 3: Based on the connectivity strength matrix, construct the connectivity probability distribution between each brain region and other brain regions, calculate the functional connectivity entropy of each brain region, and form an entropy vector. The functional connectivity entropy is used to characterize the information complexity characteristics of the functional connectivity strength distribution between a brain region and other brain regions. Based on the entropy vector, extract feature vectors to characterize the uneven distribution of entropy in the brain region space.

[0011] Step 4: Based on the feature vector, obtain the EEG intention recognition result through a classification and discrimination model.

[0012] A further technical solution, the specific steps of step one are as follows:

[0013] Assuming the acquired multi-channel EEG signals Represented as:

[0014] (1);

[0015] in, Indicates the number of channels. Indicates the first Each channel in time The EEG signal; the continuous signal is segmented into segments of length [missing information]. The time window, the Data segments corresponding to each time window for:

[0016] (2);

[0017] in, Indicates the first The starting sampling time of each time window;

[0018] Combining neurophysiological constraints, the channel set is mapped to a predetermined brain region division rule. These are distinct brain regions, among which Indicates the brain region index number, denoted as the first. The set of channel indices contained in each brain region is In terms of implementation, an average aggregation method is used to obtain brain region-level EEG signals:

[0019] (3);

[0020] And form brain region-level electrical signal vectors :

[0021] (4);

[0022] in, For the first Within the first time window Brain regions in time index Brain region-level EEG signals, For the first Within the first time window The original brainwave channels in the time index EEG signals at the location, For the first Within a time window, all A brain region-level EEG signal vector is composed of brain region-level EEG signals;

[0023] To reduce the impact of amplitude drift within different time windows on connectivity estimation, the brain region-level EEG signals were standardized.

[0024] (5);

[0025] in, To avoid positive constants with a denominator of zero, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the first Within the first time window The time mean of brain region-level EEG signals in each brain region For the first Within the first time window The time standard deviation of brain region-level EEG signals.

[0026] A further technical solution, the specific steps of step two are as follows:

[0027] Within each time window, based on step one... Calculate the functional connectivity between any two brain regions; define brain regions and Functional connection value for:

[0028] (6);

[0029] in, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the functional connectivity operator, the correlation form is used here:

[0030] (7);

[0031] Map functional connectivity values ​​to nonnegative strengths :

[0032] (8);

[0033] And construct a brain region-level connectivity strength matrix :

[0034] (9);

[0035] in, To map functional connectivity values ​​to a mapping function with non-negative strength; To represent brain regions Its connection strength with itself.

[0036] In a further technical solution, in step three, the connection probability distribution is constructed and the functional connection entropy is calculated as follows:

[0037] Targeting brain regions Extract the set of connectivity strengths between it and other brain regions. :

[0038] (10);

[0039] And construct brain regions Connection probability distribution :

[0040] (11);

[0041] Based on this definition of brain regions Functional connection entropy :

[0042] (12);

[0043] In the numerical implementation, when When the value is close to 0, it is passed through a positive constant. right The term is truncated to its lower bound, and then... ,and It does not participate in the normalization calculation of equation (11);

[0044] To eliminate the number of brain regions The resulting scale effect is normalized:

[0045] (13);

[0046] Forming an entropy vector:

[0047] (14);

[0048] in, For the first Within the first time window Normalized functional connectivity entropy values ​​for individual brain regions; For the first All within each time window The brain region functional connectivity entropy vector is formed by the normalized functional connectivity entropy values ​​corresponding to each brain region.

[0049] A further technical solution is to obtain the first step in step three. Brain region functional connectivity entropy vector within a time window Then, the entropy characteristics are further modeled from the perspective of overall spatial distribution;

[0050] First, calculate the global mean of the entropy vector. This is used to describe the overall connection complexity level:

[0051] (15);

[0052] Secondly, calculate the standard deviation of the entropy distribution. This is used to measure the degree of dispersion in the complexity of functional connectivity between different brain regions.

[0053] (16);

[0054] Based on this, we define the normalized entropy imbalance index. To eliminate the impact of different overall complexity levels:

[0055] (17);

[0056] in, It is used to reflect the overall unevenness of entropy in the brain region space. The larger the value, the more concentrated the functional connectivity complexity is in a few brain regions.

[0057] Introducing a spatial adjacency matrix of brain regions The spatial adjacency matrix of brain regions For a pre-constructed fixed matrix, Indicates brain regions With brain regions Spatial proximity in terms of anatomical or functional aspects, and Normalize them to obtain spatial weights :

[0058] (18);

[0059] Based on the aforementioned spatial weights, a local spatial difference energy of entropy is constructed to quantify the differences in functional connectivity complexity between adjacent brain regions:

[0060] (19);

[0061] in, For the first Entropy local spatial difference energy index within a time window; For the first Within the first time window Normalized functional connectivity entropy values ​​for individual brain regions;

[0062] Further set the entropy threshold :

[0063] (20);

[0064] in, This is a proportional coefficient used to adjust the high-entropy judgment threshold;

[0065] Based on this, the proportion of high-entropy brain regions was calculated. :

[0066] (twenty one);

[0067] in, This is an indicator function.

[0068] A further technical solution involves introducing a sorting-based entropy distribution analysis method in step three; firstly, the entropy vector is sorted in ascending order:

[0069] (twenty two);

[0070] in, Indicates the first The brain region functional connectivity entropy vectors within a time window are sorted in ascending order of their values ​​to obtain the [number]th [value]. One component;

[0071] The cumulative percentage function of entropy is defined as follows:

[0072] (twenty three);

[0073] in, This represents the spatial weight coefficient corresponding to the sorted brain region, which is obtained by normalizing the aforementioned brain region spatial adjacency matrix. For the first Within a time window, after sorting the brain region functional connectivity entropy vectors in ascending order of value, the first... The percentage of cumulative entropy of each component;

[0074] Based on this, a Gini-type entropy imbalance index is constructed, which is a measure of Gini imbalance corresponding to the aforementioned weighted discrete Lorentz curve:

[0075] (twenty four);

[0076] in, For the first The spatial concentration index of brain region functional connectivity entropy within a time window indicates that the larger the value, the more concentrated the functional connectivity complexity is in a few brain regions that have a high weight in the spatial structure.

[0077] Therefore, these features together constitute the first Feature vectors of uneven spatial distribution of functional connectivity entropy in brain regions within a time window :

[0078] (25).

[0079] A further technical solution, the specific steps of step four are as follows:

[0080] Based on the entropy space distribution imbalance feature constructed in step three, EEG intention discrimination is achieved. Let the intention category set be... ,in, To determine the total number of EEG intention categories to be identified, first... Perform a linear mapping:

[0081] (26);

[0082] in, The weight matrix is ​​the linear mapping. Let be the bias vector of the linear mapping. For the first Feature vectors showing uneven entropy spatial distribution within a time window The intermediate discriminant response vector obtained after linear mapping;

[0083] Based on this, construct a discrimination scoring function for intent categories:

[0084] (27);

[0085] in, Represents the weight matrix The row vectors For the first The time window corresponds to the first The discrimination score for each intent category, In order to be with the first Bias parameters corresponding to each intent category;

[0086] And the intent discrimination result for a single time window is obtained through the maximum response criterion:

[0087] (28);

[0088] in, For the first The EEG intent recognition results corresponding to each time window This operation represents the operation of taking the category index that maximizes the discriminant score function;

[0089] For applications involving continuous time windows, a time fusion mechanism is further introduced to address the issue of continuous time windows. The results of the discrimination within each time window were statistically analyzed:

[0090] (29);

[0091] in, For continuous The final EEG intention recognition result is obtained by time fusion of the discrimination results of each time window.

[0092] The present invention provides a brainwave intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions, the beneficial effects of which are as follows:

[0093] (1) Significantly enhanced feature stability and robustness: This invention transforms the original connection strength values, which are susceptible to noise and transient fluctuations, into a probability distribution-based measure of information complexity by calculating the functional connectivity entropy of brain regions. Entropy is insensitive to absolute changes in numerical values ​​but sensitive to changes in distribution patterns, thus effectively suppressing the effects of signal interference and individual differences.

[0094] (2) Enhanced Intent Discrimination and Interpretability: This invention innovatively models functional connectivity entropy from a spatial distribution perspective, extracting distributional imbalance features including global imbalance index, local spatial difference energy, and Gini concentration. These features can quantitatively characterize the structural differences in the spatial distribution of brain functional coordination patterns under different intent states, whether they are "highly dominated by a few brain regions" or "balancedly involved by multiple brain regions." This not only provides the classifier with more discriminative low-dimensional features but also links the recognition results to the spatial organization principles of neurophysiology, enhancing the interpretability of the model.

[0095] (3) High computational efficiency and conducive to engineering deployment: The features finally extracted by this invention are low-dimensional statistical vectors, which replace the high-dimensional functional connectivity matrix that needs to be processed in traditional methods. This greatly reduces the input complexity and computational burden of subsequent classification models, fully meets the requirements of online brain-computer interface systems for real-time performance and lightweight design, and has high practical value on resource-constrained embedded platforms. Attached Figure Description

[0096] Figure 1 A flowchart illustrating an EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions, provided in an embodiment of the present invention;

[0097] Figure 2 This is an overall framework diagram of an EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions, provided for an embodiment of the present invention. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0099] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0100] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a brainwave intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions, comprising the following steps:

[0101] Step 1: EEG signal acquisition and brain region-level representation construction;

[0102] Assuming the acquired multi-channel EEG signals It can be represented as:

[0103] (1);

[0104] in, Indicates the number of channels. Indicates the first Each channel in time The EEG signal. To adapt to online recognition or robust feature construction, the continuous signal is segmented, dividing it into segments of length [missing information]. The time window, the Data segments corresponding to each time window for:

[0105] (2);

[0106] in, Indicates the first The starting sampling time of each time window.

[0107] Combining neurophysiological constraints, the channel set is mapped to a predetermined brain region division rule. These are distinct brain regions, among which Indicates the brain region index number, denoted as the first. The set of channel indices contained in each brain region is In practice, an averaging aggregation method is used to obtain brain region-level EEG signals:

[0108] (3);

[0109] And form brain region-level electrical signal vectors :

[0110] (4);

[0111] in, For the first Within the first time window Brain regions in time index Brain region-level EEG signals, For the first Within the first time window The original brainwave channels in the time index EEG signals at the location, For the first Within a time window, all A brain region-level EEG signal vector composed of brain region-level EEG signals.

[0112] To reduce the impact of amplitude drift within different time windows on connectivity estimation, the brain region-level EEG signals were standardized.

[0113] (5);

[0114] in, To avoid positive constants with a denominator of zero, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the first Within the first time window The time mean of brain region-level EEG signals in each brain region For the first Within the first time window The time standard deviation of EEG signals at the brain region level. For ease of description, the time index within the time window will be uniformly denoted as . .

[0115] Step 2: Constructing functional connectivity relationships at the brain region level;

[0116] Within each time window, based on step one... Calculate the functional connectivity between any two brain regions. Define brain regions. and Functional connection value for:

[0117] (6);

[0118] in, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the functional connectivity operator, the correlation form is used here:

[0119] (7);

[0120] Considering the need to construct a probability distribution later, the functional connectivity values ​​will be mapped to non-negative strengths. :

[0121] (8);

[0122] And construct a brain region-level connectivity strength matrix :

[0123] (9);

[0124] in, To map functional connectivity values ​​to a mapping function with non-negative strength, thus eliminating the influence of connectivity symbols on the subsequent probability distribution construction; To represent brain regions In this embodiment, the connection strength between itself and the brain region is set to 0 to avoid interference from self-connection on the functional coordination modeling of brain regions.

[0125] Step two yields a connectivity matrix describing the collaborative relationships between brain regions, providing foundational data for subsequent modeling of "connectivity information complexity".

[0126] Step 3: Extraction of brain region functional connectivity entropy and its spatial distribution imbalance features;

[0127] Existing technologies typically use connection strength or connection matrix as discriminative features, which struggles to characterize the intrinsic differences in brain region collaboration patterns from the perspectives of information complexity and spatial organization. Therefore, this method systematically models brain region functional connectivity features from two dimensions: "connectivity information complexity modeling" and "spatial distribution structure characterization." Specifically, this step includes three progressive sub-processes: First, for a single brain region, the connection probability distribution between it and other brain regions is constructed, and the corresponding functional connectivity entropy is calculated to quantify the complexity level of brain region functional connectivity information. Second, based on the functional connectivity entropy vector formed by multiple brain regions, the spatial distribution imbalance feature of entropy is extracted from both overall statistical characteristics and local spatial structure perspectives to characterize the concentration or dispersion of brain region collaboration activities under different intentional states. Furthermore, a ranking-based Lorenz-Gini entropy distribution imbalance characterization method is introduced to globally measure the concentration of functional connectivity complexity in brain region space, thereby enhancing the overall representation ability of functional collaboration patterns dominated by a few brain regions.

[0128] (1) The connection probability distribution is constructed and the functional connection entropy is calculated as follows:

[0129] Targeting brain regions Extract the set of connectivity strengths between it and other brain regions. :

[0130] (10);

[0131] And construct brain regions Connection probability distribution :

[0132] (11);

[0133] Based on this definition of brain regions Functional connection entropy :

[0134] (12);

[0135] In the numerical implementation, when When the value is close to 0, it is passed through a positive constant. right The term is truncated to its lower bound, and then... ,and It does not participate in the normalization calculation of equation (11).

[0136] To eliminate the number of brain regions The resulting scale effect is normalized:

[0137] (13);

[0138] Forming an entropy vector:

[0139] (14);

[0140] in, For the first Within the first time window The normalized functional connectivity entropy value of a brain region is used to characterize the relative level of functional connectivity information complexity between that brain region and other brain regions. For the first All within each time window The brain region functional connectivity entropy vector is formed by the normalized functional connectivity entropy values ​​corresponding to each brain region.

[0141] (2) Spatial uneven distribution of entropy: global unevenness + local concentration;

[0142] In obtaining the first Brain region functional connectivity entropy vector within a time window Subsequently, the entropy characteristics were further modeled from the perspective of overall spatial distribution to characterize the degree of concentration or dispersion of brain region functional coordination under different intentional states.

[0143] First, calculate the global mean of the entropy vector. This is used to describe the overall connection complexity level:

[0144] (15);

[0145] Secondly, calculate the standard deviation of the entropy distribution. This is used to measure the degree of dispersion in the complexity of functional connectivity between different brain regions.

[0146] (16);

[0147] Based on this, we define the normalized entropy imbalance index. To eliminate the impact of different overall complexity levels:

[0148] (17);

[0149] in, It is used to reflect the overall unevenness of entropy in the brain region space. The larger the value, the more concentrated the functional connectivity complexity is in a few brain regions.

[0150] To further demonstrate the local concentration of entropy distribution in spatial structure, a spatial adjacency matrix of brain regions is introduced. The spatial adjacency matrix of brain regions For a pre-constructed fixed matrix, Indicates brain regions With brain regions Spatial proximity in terms of anatomical or functional aspects, and The spatial weights are obtained by normalizing them. :

[0151] (18);

[0152] Based on the aforementioned spatial weights, a local spatial difference energy of entropy is constructed to quantify the differences in functional connectivity complexity between adjacent brain regions:

[0153] (19);

[0154] in, For the first The entropy local spatial difference energy index within a time window is used to quantify the overall intensity of the difference in functional connectivity complexity between adjacent brain regions in terms of spatial structure. For the first Within the first time window Normalized functional connectivity entropy values ​​for individual brain regions.

[0155] Furthermore, to describe the dominance of highly complex brain regions in the whole, an entropy threshold is set. :

[0156] (20);

[0157] in, This is a proportional coefficient used to adjust the high entropy threshold, and it is usually taken as a positive real number.

[0158] Based on this, the proportion of high-entropy brain regions was calculated. :

[0159] (twenty one);

[0160] in, This is an indicator function.

[0161] (3) Characterization of the Lorentz-Gini type entropy distribution imbalance based on sorting;

[0162] To further enhance the characterization of "a few brain regions dominating functional synergistic patterns," this paper introduces a sorting-based entropy distribution analysis method based on brain region functional connectivity entropy modeling to globally measure the concentration of functional connectivity complexity in the brain region space. First, the entropy vectors are sorted in ascending order:

[0163] (twenty two);

[0164] in, Indicates the first The brain region functional connectivity entropy vectors within a time window are sorted in ascending order of their values ​​to obtain the [number]th [value]. Each component.

[0165] The cumulative percentage function of entropy (discrete Lorentz curve) is defined as follows:

[0166] (twenty three);

[0167] in, The spatial weight coefficients corresponding to the sorted brain regions are obtained by normalization of the aforementioned brain region spatial adjacency matrix and are used to reflect the relative importance of different brain regions in the spatial organization structure. For the first Within a time window, after sorting the brain region functional connectivity entropy vectors in ascending order of value, the first... The percentage of cumulative entropy of each component;

[0168] Based on this, a Gini-type entropy imbalance index is constructed. This index is a measure of Gini imbalance corresponding to the aforementioned weighted discrete Lorentz curve, used to globally characterize the concentration of functional connectivity complexity in brain regions.

[0169] (twenty four);

[0170] in, For the first The spatial concentration index of brain region functional connectivity entropy within a time window indicates that the larger the value, the more concentrated the functional connectivity complexity is in a few brain regions that have a high weight in the spatial structure.

[0171] Therefore, this method uses these features together to constitute the first... Feature vectors of uneven spatial distribution of functional connectivity entropy in brain regions within a time window :

[0172] (25).

[0173] Step 4: EEG intention recognition based on the characteristics of unbalanced entropy spatial distribution;

[0174] Based on the entropy space distribution imbalance feature constructed in step three, EEG intention discrimination is achieved. Let the intention category set be... ,in, To determine the total number of EEG intention categories to be identified, first... Perform a linear mapping:

[0175] (26);

[0176] in, The weight matrix is ​​the linear mapping. Let be the bias vector of the linear mapping. For the first Feature vectors showing uneven entropy spatial distribution within a time window The intermediate discriminant response vector obtained after linear mapping.

[0177] Based on this, construct a discrimination scoring function for intent categories:

[0178] (27);

[0179] in, Represents the weight matrix The row vectors For the first The time window corresponds to the first The discrimination score for each intent category, In order to be with the first Bias parameters corresponding to each intent category.

[0180] And the intent discrimination result for a single time window is obtained through the maximum response criterion:

[0181] (28);

[0182] in, For the first The EEG intent recognition results corresponding to each time window This operation represents the operation of taking the category index that maximizes the discriminant score function;

[0183] For applications involving continuous time windows, a time fusion mechanism is further introduced to address the issue of continuous time windows. The results of the discrimination within each time window were statistically analyzed:

[0184] (29);

[0185] in, For continuous The final EEG intention recognition result is obtained by time fusion of the discrimination results of each time window.

[0186] Through the above steps, EEG intention recognition based on the representation of functional connectivity entropy of unbalanced brain regions was achieved.

[0187] To verify the effectiveness and feasibility of this method (SIFCE), the publicly available brain-computer interface datasets BCI Competition IV-2a and BCI Competition IV-2b were selected for experimental validation. The computing platform used an NVIDIA RTX 3090 (24GB VRAM), an x86 architecture general-purpose processor, and at least 32GB of RAM. The software environment was based on Python, the deep learning framework PyTorch, and commonly used scientific computing libraries. Training was performed using the Adam optimizer, with an initial learning rate set to 1×10⁻⁶. -3 The weight decay is set to 1×10. -4 The batch size is set to 64; the maximum number of training epochs is set to 150, and an early stopping strategy is enabled: training is terminated early and rolled back to the optimal weights when the validation set metric shows no improvement for 20 consecutive epochs; the learning rate strategy uses cosine annealing, and the minimum learning rate is set to 1×10. -5 Data preprocessing includes artifact removal, bandpass filtering, and time window segmentation: First, abnormal amplitude segments are removed from the raw EEG signal to suppress significant artifact interference, with the peak-to-peak amplitude threshold set to 150 μV; then, 4-40 Hz bandpass filtering is performed, and a sliding time window of 2 s with a window shift of 1 s is used for segmentation to extract features. Brain region segmentation adopts a preset channel-brain region mapping rule, aggregating multi-channel signals within the same brain region to obtain brain region-level EEG signals. Spatial adjacency is described using a fixed brain region adjacency matrix: when two brain regions are anatomically adjacent or belong to the same functional proximity set, it is set to 1; otherwise, it is set to 0. The numerical stability constant in the entropy calculation and normalization process is taken as 1×10. -12 The proportion coefficient in the high-entropy threshold is set to 1. An in-subject training strategy is used for each participant, with the training and test sets following the official partitioning of the public dataset. Within the training set, a validation set is partitioned at a fixed ratio for early stopping and learning rate scheduling. Performance is evaluated primarily by classification accuracy (%), with the mean and standard deviation of cross-subject statistics provided.

[0188] On the BCI Competition IV-2a and BCI Competition IV-2b datasets, FBCSP-SVM, EEGNet, MIN2Net, FBCNet, and ConvNet were used as baselines for comparison. To ensure fairness, the baselines listed in Table 1 were trained and evaluated using publicly available implementations, and the comparison was validated under the same data partitioning and preprocessing settings. This method uses EEGNet as the classification backbone and introduces the spatial imbalance feature of brain region functional connectivity entropy constructed by SIFCE into its input side, forming EEGNet-SIFCE. The comparison results are shown in Table 1.

[0189] Table 1. Accuracy comparison of different methods on BCI IV-2a and BCI IV-2b

[0190]

[0191] As shown in Table 1, EEGNet-SIFCE achieved higher average accuracy and lower cross-subject variability on both the BCI IV-2a and BCI IV-2b public datasets. Specifically, on BCI IV-2a, the average accuracy of EEGNet-SIFCE reached 80.03%, an improvement of approximately 6.54 percentage points compared to EEGNet, with the standard deviation decreasing from 14.46 to 9.85. On BCI IV-2b, the average accuracy of EEGNet-SIFCE reached 72.58%, an improvement of approximately 5.77 percentage points compared to EEGNet, with the standard deviation decreasing from 6.62 to 4.91. These results indicate that, without altering the classification backbone structure, introducing the unbalanced spatial distribution of brain region functional connectivity entropy constructed by SIFCE can effectively alleviate the sensitivity of traditional connectivity strength features to noise, transient fluctuations, and individual differences, thereby improving recognition accuracy and enhancing cross-subject stability.

[0192] Table 2 Ablation Experiment

[0193]

[0194] To further verify the contribution of the key components of SIFCE to performance improvement, Table 2 presents the ablation results of feature components under the EEGNet-SIFCE framework. It should be noted that "connection entropy" in Table 2 is the core complexity representation of this method. When removing connection entropy (w / o Entropy), to ensure consistency in comparison, the "entropy vector per brain region" is replaced with the "statistical vector of connection strength per brain region (e.g., averaging the connection strength of this brain region with other brain regions)," and global imbalance features, local adjacency energy, and Gini concentration are further calculated on this strength vector to examine the source of gain of the "complexity representation" relative to the "strength representation." w / o Imbalance indicates that only the connection entropy per brain region is retained as the feature input, and spatial distribution modeling indicators such as global imbalance features, local adjacency energy, and Gini concentration are no longer introduced; w / o Λ indicates the removal of adjacent brain region difference measures in spatial distribution modeling; w / o Gini indicates the removal of ranking-based concentration measures.

[0195] As shown in Table 2, removing connection entropy (w / o Entropy) caused the largest performance drop on both datasets (BCI IV-2a: 80.03%→75.60%; BCI IV-2b: 72.58%→68.60%), indicating that the information complexity characterized by the connection probability distribution is the key foundation for the improved performance of this method. When only the connection entropy of each brain region is retained without spatial imbalance modeling (w / o Imbalance), the accuracy is still significantly lower than that of the complete method (BCI IV-2a: 76.90%; BCI IV-2b: 72.58%→68.60%). IV-2b: 70.00%, indicating that complexity characterization alone is insufficient to fully distinguish spatial organization differences under different intention states. Furthermore, removing local adjacency energy or removing Gini concentration both led to a sustained decline in performance (IV-2a: 77.40% / 77.60%; IV-2b: 70.80% / 71.00%), suggesting that local spatial differences and global concentration measurements have complementary roles in characterizing the structural pattern of "a few brain regions dominating / multiple brain regions cooperating". Overall, the ablation results experimentally validate the core innovative mechanism of this method: transforming functional connectivity from "intensity representation" to "complexity representation", and further constructing discriminative features through spatial imbalance modeling, are important reasons for achieving stable gains.

[0196] Table 3 Comparison of Complexity and Real-Time Performance

[0197]

[0198] To evaluate the computational overhead of this method in engineering deployment, Table 3 presents a comparison of the number of parameters and computation time between EEGNet and EEGNet-SIFCE. The computation time statistics are based on a single time window input, and under online inference conditions, statistics are calculated using a single-sample input method, with the average value taken from multiple runs.

[0199] As shown in Table 3, after introducing SIFCE, the number of parameters only increased from 0.34M to 0.35M, while the model storage overhead remained almost unchanged. The inference time increased slightly from 1.20 ms to 1.25 ms, with the increased overhead mainly coming from the feature construction stage (approximately 0.35 ms), causing the total time to increase from 1.20 ms to 1.55 ms. These results demonstrate that the proposed method achieves significant improvements in accuracy and stability while maintaining low computational complexity and good real-time performance. It is suitable for deployment in online brain-computer interfaces or resource-constrained scenarios, and reflects the practical engineering value of this method by "replacing high-dimensional connection matrices with low-dimensional statistical features and reducing dependence on complex models."

[0200] In summary, the comparative experiments, ablation experiments, and complexity assessments based on BCI IV-2a and BCI IV-2b demonstrate that the proposed SIFCE method can stably improve the accuracy of EEG intention recognition and reduce performance fluctuations under different public datasets and cross-subject conditions. At the same time, it achieves performance gains with almost no increase in the number of parameters and inference time, thus verifying the effectiveness, feasibility, and engineering deployment value of the proposed method.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A brainwave intention recognition method based on the representation of functional connectivity entropy in unbalanced brain regions, characterized in that, Includes the following steps: Step 1: Acquire multi-channel EEG signals and map the channel set to the target brain region according to a preset brain region division rule. The brain regions are distinguished from each other, and the brain region-level EEG signals of each brain region are obtained by aggregation and then standardized. Step 2: Within each time window, based on the standardized brain region-level EEG signals, calculate the functional connectivity values ​​between any two brain regions, map them to non-negative connectivity strength, and construct... The brain region-level connectivity strength matrix; Step 3: Based on the connectivity strength matrix, construct the connectivity probability distribution between each brain region and other brain regions, calculate the functional connectivity entropy of each brain region, and form an entropy vector; based on the entropy vector, extract feature vectors to characterize the uneven distribution of entropy in the brain region space. Step 4: Based on the feature vector, obtain the EEG intention recognition result through a classification and discrimination model; In step three, a sorting-based entropy distribution analysis method is further introduced; firstly, the entropy vector is sorted in ascending order: （22）； in, Indicates the first The brain region functional connectivity entropy vectors within a time window are sorted in ascending order of their values ​​to obtain the [number]th [value]. One component; The cumulative percentage function of entropy is defined as follows: （23）； in, This represents the spatial weight coefficient corresponding to the sorted brain region, which is obtained by normalizing the aforementioned brain region spatial adjacency matrix; For the first Within a time window, after sorting the brain region functional connectivity entropy vectors in ascending order of value, the first... The percentage of cumulative entropy of each component; To avoid positive constants with a denominator of zero; Based on this, a Gini-type entropy imbalance index is constructed, which is a measure of Gini imbalance corresponding to the weighted discrete Lorentz curve: （24）； in, For the first The spatial concentration index of brain region functional connectivity entropy within a time window indicates that the larger the value, the more concentrated the functional connectivity complexity is in a few brain regions that have a high weight in the spatial structure. Therefore, these features together constitute the first Feature vectors of uneven spatial distribution of functional connectivity entropy in brain regions within a time window : （25） in, Let be the global mean of the entropy vector. Let the standard deviation of the entropy distribution be denoted as . The normalized entropy imbalance index. For the first Entropy local spatial difference energy index within a time window This is the entropy threshold.

2. The EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions according to claim 1, characterized in that, The specific steps of step one are as follows: Assuming the acquired multi-channel EEG signals Represented as: （1）； in, Indicates the number of channels. Indicates the first Each channel in time The EEG signal; the continuous signal is segmented into segments of length [missing information]. The time window, the Data segments corresponding to each time window for: （2）； in, Indicates the first The starting sampling time of each time window; Combining neurophysiological constraints, the channel set is mapped to a predetermined brain region division rule. These are distinct brain regions, among which Indicates the brain region index number, denoted as the first. The set of channel indices contained in each brain region is In terms of implementation, an average aggregation method is used to obtain brain region-level EEG signals: （3）； And form brain region-level electrical signal vectors : （4）； in, For the first Within the first time window Brain regions in time index Brain region-level EEG signals, For the first Within the first time window The original brainwave channels in the time index EEG signals at the location, For the first Within a time window, all A brain region-level EEG signal vector is composed of brain region-level EEG signals; To reduce the impact of amplitude drift within different time windows on connectivity estimation, the brain region-level EEG signals were standardized: （5）； in, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the first Within the first time window The time mean of brain region-level EEG signals in each brain region For the first Within the first time window The time standard deviation of brain region-level EEG signals.

3. The EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions according to claim 2, characterized in that, The specific steps of step two are as follows: Within each time window, based on step one... Calculate the functional connectivity between any two brain regions; define brain regions and Functional connection value for: （6）； in, For the first Within the first time window Brain regions in time index Standardized brain region-level EEG signals, For the functional connectivity operator, the correlation form is used here: （7）； Map functional connectivity values ​​to nonnegative strengths : （8）； And construct a brain region-level connectivity strength matrix : （9）； in, To map functional connectivity values ​​to a mapping function with non-negative strength; To represent brain regions Its connection strength with itself.

4. The EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions according to claim 3, characterized in that, In step three, the connection probability distribution is constructed and the functional connection entropy is calculated as follows: Targeting brain regions Extract the set of connectivity strengths between it and other brain regions. : （10）； And construct brain regions Connection probability distribution : （11）； Based on this definition of brain regions Functional connection entropy : （12）； In the numerical implementation, when When the value is close to 0, it is passed through a positive constant. right The term is truncated to its lower bound, and then... ,and It does not participate in the normalization calculation of equation (11); To eliminate the number of brain regions The resulting scale effect is normalized: （13）； Forming an entropy vector: （14）； in, For the first Within the first time window Normalized functional connectivity entropy values ​​for individual brain regions; For the first All within each time window The brain region functional connectivity entropy vector is formed by the normalized functional connectivity entropy values ​​corresponding to each brain region.

5. The EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions according to claim 4, characterized in that, In step three, the first... Brain region functional connectivity entropy vector within a time window Then, the entropy characteristics are further modeled from the perspective of overall spatial distribution; First, calculate the global mean of the entropy vector. This is used to describe the overall connection complexity level: （15）； Secondly, calculate the standard deviation of the entropy distribution. This is used to measure the degree of dispersion in the complexity of functional connectivity between different brain regions. （16）； Based on this, we define the normalized entropy imbalance index. To eliminate the impact of different overall complexity levels: （17）； in, It is used to reflect the overall unevenness of entropy in the brain region space. The larger the value, the more concentrated the functional connectivity complexity is in a few brain regions. Introducing a spatial adjacency matrix of brain regions The spatial adjacency matrix of brain regions For a pre-constructed fixed matrix, Indicates brain regions With brain regions Spatial proximity in terms of anatomical or functional aspects, and Normalize them to obtain spatial weights : （18）； Based on the aforementioned spatial weights, a local spatial difference energy of entropy is constructed to quantify the differences in functional connectivity complexity between adjacent brain regions: （19）； in, For the first Within the first time window Normalized functional connectivity entropy values ​​for individual brain regions; Further set the entropy threshold : （20）； in, This is a proportional coefficient used to adjust the high-entropy judgment threshold; Based on this, the proportion of high-entropy brain regions was calculated. : （21）； in, This is an indicator function.

6. The EEG intention recognition method based on the representation of functional connectivity entropy of unbalanced brain regions according to claim 5, characterized in that, The specific steps of step four are as follows: Based on the entropy space distribution imbalance feature constructed in step three, EEG intention discrimination is achieved, assuming the intention category set is... ,in, To determine the total number of EEG intention categories to be identified, first... Perform a linear mapping: （26）； in, The weight matrix is ​​the linear mapping. Let be the bias vector of the linear mapping. For the first Feature vectors showing uneven entropy spatial distribution within a time window The intermediate discriminant response vector obtained after linear mapping; Based on this, construct a discrimination scoring function for intent categories: （27）； in, Represents the weight matrix The row vectors For the first The time window corresponds to the first The discrimination score for each intent category, In order to be with the first Bias parameters corresponding to each intent category; And the intent discrimination result for a single time window is obtained through the maximum response criterion: （28）； in, For the first The EEG intent recognition results corresponding to each time window This operation represents the operation of taking the category index that maximizes the discriminant score function; For applications involving continuous time windows, a time fusion mechanism is further introduced to address the issue of continuous time windows. The results of the discrimination within each time window were statistically analyzed: （29）； in, For continuous The final EEG intention recognition result is obtained by time fusion of the discrimination results of each time window.

Images