Brain-computer interface signal augmentation and evaluation methods and systems that fuse data and knowledge

Abstract

The application relates to a brain-computer interface signal enhancement and evaluation method, system, device and storage medium fusing data and knowledge. The method comprises the following steps: acquiring general EEG semantic representation, baseline EEG features, differential equation driven features and geometric prior features of original electroencephalogram signals respectively; using a multi-scale brain physiological feature fusion strategy to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation driven features and geometric prior features, to generate electroencephalogram fusion features; taking the electroencephalogram fusion features as brain physiological consistency guiding features, and embedding the iEEG diffusion generation module to guide the diffusion generation of the iEEG signal, to obtain an electroencephalogram enhanced signal. Under the premise of not relying on invasive recording, the application realizes end-to-end generation from non-invasive scalp EEG to high-fidelity iEEG-like signals, and significantly improves the performance of the electroencephalogram enhanced signal in terms of time domain fidelity, spectral consistency and rationality of cortical spatial distribution.

Description

Technical Field

[0001] This application belongs to the fields of brain-computer interface, neural engineering and clinical neuroscience, and specifically relates to a method, system, device and storage medium for enhancing and evaluating brain-computer interface signals that integrates data and knowledge. Background Technology

[0002] Brain-computer interface (BCI) is a technology that establishes a direct communication pathway between the brain and external devices. Its core objective is to control computers, robots, prostheses, or neurofeedback systems by decoding neural activity signals. In recent years, with the rapid development of neuroscience, signal processing, and artificial intelligence, BCI has shown great application potential in fields such as motor rehabilitation, consciousness monitoring, diagnosis and treatment of neurological diseases, and human-machine collaboration. The quality of neural signal acquisition and representation capabilities directly determine the decoding accuracy, response speed, and universality of a BCI system.

[0003] Among existing neural signal acquisition paradigms, electroencephalography (EEG) and intracranial electroencephalography (IEEG) are the two most commonly used types of brain electrophysiological signals. EEG, by non-invasively recording the electrical activity of the cerebral cortex using electrodes placed on the scalp surface, has advantages such as ease of operation, low cost, and high temporal resolution (down to millisecond levels), and is therefore widely used in consumer and clinical BCI systems. However, EEG signals undergo severe attenuation and spatial ambiguity when passing through high-impedance tissues such as the scalp and skull, resulting in low spatial resolution (typical localization error >2 cm), poor signal-to-noise ratio, and difficulty in separating independent activity sources in adjacent brain regions. In contrast, iEEG, by directly implanting electrodes under the dura mater or within the brain parenchyma, can capture high-frequency activity (such as the high-gamma band) of local neuronal groups with sub-centimeter spatial accuracy, demonstrating significant superiority in tasks such as epileptic focus localization, language function area mapping, and high-dimensional intent decoding. However, obtaining iEEG requires craniotomy, which is invasive, high-risk, high-cost, and ethically restricted. It is only applicable to specific clinical patients and cannot be extended to healthy people or everyday applications.

[0004] Given the aforementioned limitations, reconstructing enhanced signals with iEEG-level spatiotemporal fidelity and physiological plausibility from non-invasive scalp EEG without relying on invasive recording has become a critical technological bottleneck that urgently needs to be overcome in the fields of brain-computer interfaces, neural engineering, and clinical neuroscience. Currently, this technology faces at least the following technical challenges: First, EEG signals themselves lack clear spatial source information, requiring the introduction of physical models or neurodynamic equations to constrain the signal reconstruction process; second, the geometric structure of the individual cerebral cortex has a decisive influence on electric field propagation, while existing enhancement methods generally ignore the anatomical priors provided by T1MRI, resulting in a lack of anatomical plausibility in the generated signals; third, relying solely on data-driven models makes it difficult to simultaneously capture the multi-scale characteristics of EEG, and even more difficult to effectively fuse heterogeneous information from EEG and MRI; fourth, existing generation models, without strong guidance, are prone to producing spurious signals with spectral distortion, unreasonable spatial distribution, or temporal discontinuity.

[0005] To address the aforementioned issues, existing technologies include patent CN117731306A, which discloses a real-time EEG signal enhancement method based on an edge-end Kalman filter network. This method first linearizes the original EEG signal to fit the Kalman filter framework, then inputs it into a lightweight edge-end Kalman filter network, and dynamically calculates the Kalman gain through an adaptive gain estimation module to optimize state updates. This method effectively improves the stability of diffusion model training, significantly enhances the diversity and emotional semantic consistency of generated EEG signals, and provides high-quality data augmentation support for small-sample EEG emotion recognition tasks. Another patent, CN120725066A, discloses a soft-hybrid generative adversarial network model training method for small-sample EEG signal enhancement. This method constructs a generative adversarial framework based on Wasserstein distance constraints, generates new EEG samples through a soft-hybrid strategy, and introduces a multiple attention mechanism in the classification network to enhance feature extraction capabilities from the time series, channel dimension, and spatial topology levels, achieving collaborative optimization of the generation and discrimination processes. This method significantly reduces computational overhead while ensuring enhanced performance, achieving low-latency, high-energy-efficiency real-time EEG signal enhancement, and is suitable for wearable or edge brain-computer interface systems.

[0006] However, existing technologies still have the following shortcomings:

[0007] First, existing generative models generally lack the ability to jointly model the multi-scale neurodynamic characteristics of EEG. For example, while some diffusion-based methods introduce emotion labels or penalty terms to enhance diversity, their feature extraction remains limited to the original signal or shallow transformations. They fail to synergistically utilize the high-level semantics captured by pre-trained networks, the brain's physiological frequency band trends preserved by low-pass filtering, and the neural activity evolution features derived from physically constrained differential equations. This singular or fragmented feature representation makes it difficult for the generated signal to approximate the real iEEG in terms of temporal waveform continuity, frequency domain energy distribution, and cortical dynamic consistency.

[0008] Secondly, mainstream enhancement methods neglect the decisive influence of individualized anatomical structures on EEG propagation and distribution. Although T1-weighted MRI (Nuclear Magnetic Resonance Imaging) can provide high-resolution geometric information of the cerebral cortex, current technology almost completely fails to incorporate this structural prior into the signal generation process. The generated results rely solely on the statistical regularities of EEG data, lacking constraints on inherent topological properties such as cortical curvature and sulcus orientation. This leads to the enhanced signal potentially violating neuroanatomical rules in spatial distribution, reducing its reliability and interpretability in clinical tasks such as epilepsy localization and functional area mapping.

[0009] Furthermore, existing fusion and generation mechanisms struggle to effectively integrate heterogeneous multimodal information. Some methods attempt to introduce attention mechanisms or multi-scale modules, but their designs primarily focus on improving classification performance rather than feature alignment and fusion for high-quality signal reconstruction. Especially when the input contains heterogeneous tensors such as semantic features, filtered features, differential equation features, and geometric modalities, simple concatenation or shallow interactions can easily lead to information redundancy, key feature suppression, or cross-modal mismatch, thus limiting the generative model's ability to reconstruct complex neural representations.

[0010] Finally, existing technologies struggle to balance real-time performance with the physiological plausibility of the brain. For example, while edge-end methods based on Kalman filtering offer low latency, their linear assumptions and fixed-noise models fail to capture the nonlinear neurodynamics of EEG. On the other hand, while deep generative models can model complex distributions, the lack of anatomical guidance and multi-scale feature support often results in spectral distortion or unreasonable spatial distribution, making it difficult to simultaneously meet the dual requirements of high fidelity and deployability. Summary of the Invention

[0011] This application provides a method, system, device, and storage medium for enhancing and evaluating brain-computer interface signals that integrate data and knowledge, aiming to at least partially solve one of the aforementioned technical problems in the prior art.

[0012] To address the above problems, this application provides the following technical solution:

[0013] A method for enhancing and evaluating brain-computer interface signals by integrating data and knowledge, characterized by comprising:

[0014] The raw EEG signals were encoded using a general EEG representation extraction module to obtain general EEG semantic representations;

[0015] The original EEG signal is processed by a low-pass filtering feature extraction module to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise.

[0016] A three-layer conductive head model was constructed using T1 MRI signals, and the mapping relationship between EEG sensor signals and cortical surface current density was solved by combining a regularized minimum norm estimation framework. A differential equation system was constructed that couples the forward electromagnetic field propagation model with the cortical mean field neural dynamics, and the differential equation driving characteristics reflecting the spatiotemporal distribution of neural activity were obtained.

[0017] Laplace-Beltramian operator feature decomposition was performed on the cortical surface reconstructed by individual T1 MRI to obtain K different frequency bands of brain geometric intrinsic modes. All brain geometric intrinsic modes were projected onto the EEG electrode space to generate multi-scale, frequency domain decoupled geometric prior features.

[0018] A multi-scale brain physiological feature fusion strategy is used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features to generate EEG fusion features.

[0019] The EEG fusion features are used as guiding features for brain physiological consistency, and an iEEG diffusion generation module is embedded to guide the diffusion generation of iEEG signals to obtain enhanced EEG signals.

[0020] The technical solution adopted in this application embodiment further includes: the general EEG representation extraction module uses a specially designed and trained large EEG model as the bone intervention training model, and completes pre-training on a large-scale public EEG dataset; the general EEG representation extraction module encodes the original EEG signal to obtain general EEG semantic representation, specifically as follows:

[0021] Given raw EEG signals in C For the number of channels, T Given the number of time sampling points, a high-level general EEG semantic representation is extracted using stacked, depth-separable modules:

[0022]

[0023] in For feature dimensions.

[0024] The technical solution adopted in this application embodiment further includes: performing low-pass filtering processing on the original EEG signal through the low-pass filtering feature extraction module to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise, specifically:

[0025] The cutoff frequency is set to zero phase distortion IIR or FIR filter. Baseline EEG characteristics were obtained that preserved the main neural rhythms and removed non-neurogenic high-frequency noise:

[0026]

[0027] in Indicates the cutoff frequency as Low-pass filtering operation.

[0028] The technical solution adopted in this application embodiment also includes: constructing a three-layer conductive head model using T1 MRI signals, and solving the mapping relationship between EEG sensor signals and cortical surface current density using a regularized minimum norm estimation framework, constructing a differential equation system coupling the forward electromagnetic field propagation model and cortical mean field neural dynamics, and obtaining differential equation driving characteristics reflecting the spatiotemporal distribution of neural activity, specifically:

[0029] Based on individual T1 MRI images, head tissue structures are segmented to construct an individualized head model. The forward operator characterizing the electromagnetic conduction relationship between scalp electrodes and cortical sources is denoted as G∈ ,in N The number of discrete source points in the cortex;

[0030] Introducing the continuous neural field equations describing the evolution of mean field electrical activity in the cortex:

[0031]

[0032] in This represents the potential state vector of the cerebral cortex. τ It is a time constant. W For local connectivity kernels, σ( () is the sigmoid activation function. I(t) For external input;

[0033] The operator G for forward propagation of the EEG signal is coupled with the continuous neural field equation, and the relationship between scalp observation and internal state is defined as follows:

[0034]

[0035] in x(t) For theoretical scalp EEG, (t) For noise;

[0036] Through embedded numerical simulation or inverse consistency optimization, the potential state trajectory or residual correction quantity consistent with the laws of neuroelectromagnetic dynamics is deduced from the actual observed scalp EEG sequence X, and the differential equation driving features are extracted from this.

[0037]

[0038] The technical solution adopted in this application embodiment further includes: performing Laplace-Beltramian operator feature decomposition on the cortical surface reconstructed by individual T1 MRI to obtain K different frequency bands of brain geometric intrinsic modes, and projecting all brain geometric intrinsic modes onto the EEG electrode space to generate multi-scale, frequency-domain decoupled geometric prior features, specifically:

[0039] Reconstructing the axial surface of the cerebral cortex using individual T1 MRI And define LBO on this non-Euclidean brain surface:

[0040]

[0041] in Tensor for measuring the sulci and gyri of the cerebral cortex. For definition in Scalar functions on the LBO; by solving the eigenvalue problem, a set of orthogonal and complete brain geometric eigenmodes is obtained. and their corresponding brain modality features :

[0042] ,

[0043] Select the decomposition results A number of brain geometric intrinsic modalities, covering anatomical scales significantly correlated with iEEG spatial distribution, are grouped according to the feature values ​​of these brain modalities to construct a multi-scale brain geometric prior feature map:

[0044]

[0045] in The number of source points in the cerebral cortex;

[0046] The cortical-electrode mapping module maps all brain geometric intrinsic modes to the scalp EEG electrode space, and performs interpolation alignment synchronized with the EEG signal along the time dimension to obtain geometric prior features aligned with the EEG time sequence.

[0047]

[0048] The technical solution adopted in this application embodiment further includes: dynamically fusing the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features using a multi-scale brain physiological feature fusion strategy to generate EEG fusion features, specifically:

[0049] Joint temporal coding using L-layer stacked convolutions for local fine-grained modeling and global context awareness. The general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features are then concatenated along the EEG channel dimension and aligned with the time dimension to form the initial fused EEG features:

[0050] :

[0051]

[0052] The final output Z(L) is input into the cross-modal neural electrical signal learning module, which performs feature-level alignment based on the EEG semantic context of the current time step to generate structurally consistent and information-complete EEG fusion features. The core computation of the cross-modal neural electrical signal learning module is as follows:

[0053]

[0054]

[0055] The technical solution adopted in this application embodiment further includes: using the EEG fusion feature as a brain physiological consistency guiding feature, and embedding an iEEG diffusion generation module to guide the diffusion generation of iEEG signals, ultimately outputting an enhanced EEG signal, specifically:

[0056] Let the target EEG enhancement signal Y be... The iEEG-like signal to be generated is defined as follows:

[0057]

[0058] in These are the preset noise scheduling parameters;

[0059] The reverse generation process is achieved by a parameterized neural network. Learning, its conditional generation distribution is:

[0060]

[0061] The training objective is to minimize the noise prediction loss.

[0062]

[0063] The spectral consistency loss and spatial distribution consistency loss are jointly optimized on the basis of the base loss. The power spectral density of the EEG enhancement signal generated by the spectral consistency loss is forced to approximate the typical distribution of real iEEG in the δ–γ frequency band. The spatial distribution consistency loss is verified by the forward computation model of the neural electrical signal to see whether the projection of the generated iEEG source on the cortex is consistent with the input scalp EEG at the electromagnetic field level.

[0064] During the inference phase of EEG signal enhancement, from pure Gaussian noise T Starting from N(0,I), the conditional denoising process is iteratively executed to generate a high-fidelity EEG enhancement signal:

[0065]

[0066] Another technical solution adopted in this application embodiment is: a brain-computer interface signal enhancement and evaluation system that integrates data and knowledge, comprising:

[0067] General EEG Representation Extraction Module: Used to encode raw EEG signals and obtain general EEG semantic representations;

[0068] Low-pass filtering feature extraction module: used to perform low-pass filtering on the original EEG signal to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise;

[0069] Head electrical signal calculation module: It is used to construct a three-layer conductive head model using T1 MRI signals, and solve the mapping relationship between EEG sensor signals and cortical surface current density by combining a regularized minimum norm estimation framework. It constructs a differential equation system that couples the forward electromagnetic field propagation model with the cortical mean field neural dynamics, and obtains the differential equation driving characteristics that reflect the spatiotemporal distribution of neural activity.

[0070] Geometric Calculation Module: Used to perform Laplace-Beltramian operator feature decomposition on the cortical surface reconstructed by individual T1 MRI, obtain K different frequency bands of brain geometric eigenmodes, and project all brain geometric eigenmodes onto the EEG electrode space to generate multi-scale, frequency domain decoupled geometric prior features;

[0071] EEG Feature Fusion Module: Used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features using a multi-scale brain physiological feature fusion strategy to generate EEG fusion features;

[0072] iEEG diffusion generation module: used to use the EEG fusion features as brain physiological consistency guidance features, and to embed the iEEG diffusion generation module to guide the diffusion generation of iEEG signals, and finally output the enhanced EEG signal.

[0073] Another technical solution adopted in this application embodiment is: a device, the device including a processor and a memory coupled to the processor, wherein,

[0074] The memory stores program instructions for implementing the brain-computer interface signal enhancement and evaluation method for fusing data and knowledge;

[0075] The processor is used to execute the program instructions stored in the memory to control the brain-computer interface signal enhancement and evaluation method for fusing data and knowledge.

[0076] Another technical solution adopted in this application embodiment is: a storage medium storing processor-executable program instructions, the program instructions being used to execute the brain-computer interface signal enhancement and evaluation method for fusing data and knowledge.

[0077] Compared to existing technologies, the beneficial effects of the embodiments of this application are as follows: The brain-computer interface signal enhancement and evaluation method, system, device, and storage medium of this application, which integrates data and knowledge, adopts a unified EEG representation enhancement framework that deeply integrates data-driven and mathematical prior knowledge. It achieves efficient alignment and fusion of heterogeneous features through four-way parallel feature extraction and employs stacked dilated convolution and multi-head attention mechanisms. Finally, based on the fused representation, it drives a diffusion model to generate high-fidelity EEG enhancement signals. This application achieves end-to-end generation from non-invasive scalp EEG to high-fidelity iEEG signals without relying on invasive recording, significantly improving the performance of EEG enhancement signals in terms of temporal fidelity, spectral consistency, and the rationality of cortical spatial distribution. It effectively bridges the representation gap between non-invasive EEG and invasive iEEG, providing reliable technical support for high-precision non-invasive brain-computer interfaces and neuroscience research. Attached Figure Description

[0078] Figure 1 This is a schematic flowchart of a brain-computer interface signal enhancement and evaluation method that integrates data and knowledge according to an embodiment of this application;

[0079] Figure 2 This is a schematic diagram of the structure of a brain-computer interface signal enhancement and evaluation system that integrates data and knowledge according to an embodiment of this application;

[0080] Figure 3 This is a schematic diagram of the device structure according to an embodiment of this application;

[0081] Figure 4 This is a schematic diagram of the structure of the storage medium according to an embodiment of this application. Detailed Implementation

[0082] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0083] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0084] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0085] Specifically, please refer to Figure 1 This is a flowchart of a brain-computer interface signal enhancement and evaluation method that integrates data and knowledge, according to an embodiment of this application. The brain-computer interface signal enhancement and evaluation method that integrates data and knowledge, according to an embodiment of this application, includes the following steps:

[0086] S100: Acquire raw EEG signals and encode them using a general EEG representation extraction module to obtain general EEG semantic representations with strong generalization ability;

[0087] In this step, addressing the challenge of effectively capturing high-level time-frequency-space patterns in scalp EEG signals, this application employs a general EEG representation extraction module based on a pre-trained large-scale EEG model to encode the raw EEG signals. This module uses a specially designed and trained large-scale EEG model as the bone intervention training model and completes pre-training on a large-scale public EEG dataset. By utilizing the unique deep EEG semantic separable convolutional structure of the pre-trained EEG model, it can obtain a general EEG semantic representation with strong generalization ability while preserving the functional connectivity characteristics between EEG signal channels. Furthermore, it possesses good cross-subject generalization ability, improving the efficiency and discriminativeness of general EEG semantic representation extraction. Simultaneously, using the pre-trained model can significantly reduce overfitting to small sample data and improve the richness of feature semantics.

[0088] Specifically, the encoding process of the general EEG representation extraction module includes: given the raw EEG signal... in C For the number of channels, T Given the number of time sampling points, a high-level general EEG semantic representation is extracted using stacked, depth-separable modules:

[0089] (1)

[0090] in The feature dimension is [not specified]. This general EEG semantic representation preserves the task-independent common structure of EEG, providing a data-adaptive initial representation for subsequent signal enhancement.

[0091] It should be noted that the general EEG representation extraction module based on the pre-trained large EEG model can also be replaced by other deep feature extraction networks designed specifically for neural electrical signals. As long as the topology of functional connections between electrodes can be effectively preserved and discriminative representations related to cortical source activity can be output, it can be used as the semantic input for iEEG generation.

[0092] S110: The low-pass filtering feature extraction module, which is conducive to EEG trend analysis, performs low-pass filtering on the raw EEG signal to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise.

[0093] In this step, to effectively preserve low-frequency neural oscillation components closely related to deep cortical neural synchronous activity and clinicopathological events during the reconstruction of high-fidelity iEEG signals from scalp EEG signals, such as the δ (1–4 Hz), θ (4–8 Hz), α (8–13 Hz), and β (13–30 Hz) bands, while suppressing high-frequency electromyography artifacts (>40 Hz) and 50 / 60 Hz power line interference that easily contaminate iEEG generation results, this application utilizes a low-pass filtering feature extraction module oriented towards EEG trends to perform low-pass filtering preprocessing on the original EEG signal to optimize iEEG physiological fidelity. Specifically, a zero-phase-distortion IIR or FIR filter is used to set the cutoff frequency to [value missing]. This allows for the acquisition of baseline EEG characteristics that preserve major neural rhythms while removing non-neurogenic high-frequency noise:

[0094] (2)

[0095] in Indicates the cutoff frequency as Low-pass filtering operation. Baseline EEG characteristics after filtering. It prominently depicts the steady-state brain rhythm structures that are highly correlated with iEEG generation, avoiding the erroneous amplification of high-frequency artifacts into false iEEG high-frequency discharges during subsequent diffusion generation. This provides the entire EEG enhancement model with physiologically reliable and noise-robust EEG characteristics, which helps to reliably reconstruct key clinical neural activity patterns such as slow-wave sleep stages or epileptiform discharges under non-invasive conditions.

[0096] S120: A three-layer conductive head model was constructed using T1 MRI signals, and the mapping relationship between EEG sensor signals and cortical surface current density was solved by combining a regularized minimum norm estimation framework. A differential equation system was constructed that coupled the forward electromagnetic field propagation model and the cortical mean field neural dynamics to obtain the differential equation driving characteristics that reflect the spatiotemporal distribution of neural activity.

[0097] In this step, addressing the problem that traditional EEG signals are spatially ambiguous and unable to reflect deep cortical dynamics due to the volume conduction effect, this application utilizes a neurodynamically guided head electroencephalogram (EEG) calculation module to construct a three-layer conductive head model using T1 MRI signals. This model accurately characterizes the conduction characteristics of current in head tissues. Furthermore, by incorporating the regularized minimum norm estimation framework relied upon by the EEG signal optimization algorithm and introducing a solution strategy for the physical constraints of the human head, noise interference in EEG signal transmission is effectively suppressed, and spatial consistency is improved. This provides a temporal-spatial joint representation basis with neurodynamic foundations for subsequent EEG signal enhancement calculations. Simultaneously, to introduce biophysical consistency constraints that conform to the laws of electromagnetic propagation in the brain and the mechanisms of cortical population neural activity during EEG signal enhancement, this application constructs a differential equation system coupling a forward electromagnetic field propagation model and cortical mean-field neurodynamics. This system generates guided features that are physically interpretable and consistent with real intracranial activity in terms of spatiotemporal evolution.

[0098] Specifically, the head electrical signal calculation module first divides the head tissue structure based on individual T1 MRI images and constructs an individualized head model, where the forward operator characterizing the electromagnetic conduction relationship between the scalp electrodes and the cortical source is denoted as G∈ , N Let be the number of discrete source points in the cortex. Then, the continuous neural field equation describing the evolution of the mean field electrical activity in the cortex is introduced:

[0099] (3)

[0100] in This represents the potential state vector of the cerebral cortex. τ It is a time constant. W For local connectivity kernels, σ( () is the sigmoid activation function. I(t) As an external input, the operator G for forward propagation of the EEG signal is coupled with the continuous neural field equation, defining the relationship between scalp observation and internal state as follows:

[0101] (4)

[0102] in x(t) For theoretical scalp EEG, (t) This is noise. Based on this joint EEG signal generation model, this application uses embedded numerical simulation or inverse consistency optimization to deduce the potential state trajectory or residual correction quantity consistent with the laws of neuroelectromagnetic dynamics from the actually observed scalp EEG sequence X, and thereby extracts the differential equation driving features:

[0103] (5)

[0104] This differential equation-driven feature explicitly encodes the inherent rationality of EEG signals under the dual physical constraints of cortical neural population dynamics and head volume conduction. It can effectively identify and suppress non-biological artifacts that violate neurophysiological laws, thereby providing high-confidence prior guidance for subsequent iEEG generation that conforms to the real brain working mechanism in both spatiotemporal distribution. This improves the brain physiological interpretability, spatiotemporal consistency, and clinical usability of the final iEEG-like signals generated.

[0105] It should be noted that the head electrical signal calculation module can also use other forms of physical constraints or optimization strategies to extract differential equation-driven features. As long as it can provide differential equation-driven features that reflect the spatiotemporal dynamic evolution of neural electrical signals and effectively suppress spatial ambiguity caused by volume conduction, the physiological rationality of iEEG generation can be maintained.

[0106] S130: The geometric computation module for head model frequency band decomposition is used to perform Laplace–Beltrami Operator (LBO) feature decomposition on the cortical surface reconstructed by individual T1 MRI to obtain K different frequency bands of brain geometric intrinsic modes. All brain geometric intrinsic modes are then projected to the EEG electrode space through the cortical-electrode mapping module to generate multi-scale, frequency domain decoupled geometric prior features.

[0107] In this step, to introduce individual-specific cerebral cortical anatomy priors into the reconstruction of iEEG-like signals from scalp EEG, thereby improving the physiological accuracy of the generated iEEG signals in terms of sulcus functional area localization and spatial propagation characteristics, and ensuring the numerical accuracy of the generated iEEG signals at the physical level, this application first uses a geometric calculation module to perform Laplace-Beltramian operator feature decomposition on the cortical surface reconstructed from individual T1 MRI, obtaining K different frequency bands of brain geometric intrinsic modes. These modes naturally encode the global geometric properties of the cerebral cortex. Subsequently, the cortical-electrode mapping module projects all brain geometric intrinsic mode features onto the EEG electrode space, generating a set of multi-scale, frequency-domain decoupled geometric prior features.

[0108] Specifically, the geometric computation module uses individual T1 MRI to reconstruct the axial surface of the cerebral cortex. And define LBO on this non-Euclidean brain surface:

[0109] (6)

[0110] in Tensor for measuring the sulci and gyri of the cerebral cortex. For definition in Scalar functions on the LBO. By solving the eigenvalue problem of the LBO, a set of orthogonal and complete geometric eigenmodes of the brain are obtained. and their corresponding brain modality features :

[0111] , (7)

[0112] Among them, brain modal feature values It directly reflects the spatial frequency characteristics of the corresponding geometric intrinsic modes of the brain on the surface of the subject's cerebral cortex: The smaller the value, the smoother the change in the geometric intrinsic mode of the brain, corresponding to the main structures of the brain's global sulci (such as the central sulcus and lateral fissure). The larger the size, the more intense the oscillations of the brain's geometric intrinsic modes, depicting local wrinkle details.

[0113] Select the decomposition results A multi-scale brain geometric prior feature map was constructed by grouping brain modal feature values ​​according to anatomical scales that are significantly correlated with iEEG spatial distribution:

[0114] (8)

[0115] in This refers to the number of source points in the cerebral cortex.

[0116] Subsequently, all brain geometric intrinsic modes are mapped to the scalp EEG electrode space through the cortical-electrode mapping module, and interpolation alignment synchronized with the EEG signal is performed along the time dimension to finally obtain geometric prior features aligned with the EEG time sequence:

[0117] (9)

[0118] This geometric prior explicitly encodes the modulation effect of individual brain sulcus geometry on the spatial projection of cortical currents onto the scalp. For example, source points located at the top of the gyri contribute stronger EEG signals due to their proximity to scalp electrodes, while source points within deep sulci experience signal attenuation due to geometric masking. By injecting these spatial gain differences determined by real anatomical structures into the generation process as learnable priors, individualized, multi-scale, and frequency-domain decoupled anatomical constraints are provided for the iEEG enhancement task, significantly improving the numerical accuracy, sulcus functional correspondence, and cross-subject generalization ability of generated iEEGs.

[0119] It should be noted that other surface feature extraction techniques that can characterize the main sulci and folds of the brain and the local fold characteristics can also be used for geometric prior feature extraction. As long as the output can encode the spatial modulation effect of cortical geometry on electric field propagation and can be transformed into static guided EEG features of EEG channel space through the electrode mapping mechanism, the interpretability of frequency band calculation for iEEG generation can be supported.

[0120] S140: Utilizes an EEG feature fusion module adapted to multi-scale neural characteristics to dynamically fuse general EEG semantic representations, baseline EEG features, differential equation-driven features, and geometric prior features to generate structurally consistent and information-complete EEG fusion features.

[0121] In this step, to effectively integrate general EEG semantic representations, baseline EEG features, differential equation-driven features, and geometric prior features, this application proposes an EEG feature fusion module specifically for reconstructing iEEG-like signals from scalp EEG. First, it captures long-range dependencies of neural oscillations spanning tens to hundreds of milliseconds by stacking temporal receptive fields through multi-scale convolution. Then, it introduces a cross-modal neural electrical signal learning module to dynamically fuse general EEG semantic representations, baseline EEG features, differential equation-driven features, and geometric prior features, achieving feature alignment and information complementarity guided by the iEEG generation goal. This provides fusion-guided features for EEG signal generation that combine local dynamic details with global physiological constraints of the brain.

[0122] Specifically, the feature fusion process of the EEG feature fusion module includes: First, in order to capture the long-range neural oscillation dependence closely related to iEEG generation while controlling model complexity, L-layer stacked convolutions are used for joint temporal encoding of local fine modeling and global context awareness. The general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features are then concatenated along the EEG channel dimension and aligned with the time dimension to form the initial fused EEG features:

[0123] : (10)

[0124] (11)

[0125] This fusion strategy enables the model to simultaneously perceive millisecond-level transient brain events during a single forward pass, ensuring that the reconstructed iEEG possesses key dynamic patterns.

[0126] Subsequently, the final output Z(L) is input into the cross-modal neural electrical signal learning module. This module adaptively weights more reliable neurodynamic states and more guided brain anatomical modalities based on the EEG semantic context of the current time step, thereby achieving feature-level alignment with iEEG spatiotemporal fidelity as the goal, generating structurally consistent and information-complete EEG fusion features. The core computation of the cross-modal neural electrical signal learning module is as follows:

[0127] (12)

[0128] (13)

[0129] This multi-scale brain physiological feature fusion strategy can balance the ability of convolution to accurately characterize local neural dynamics of the brain with the ability of attention to dynamically model the dependence of global cross-modal EEG signals on the brain. The final output is a fused EEG feature. With a consistent structure and complete information, it not only preserves the local details of high-frequency transient brain events, but also embeds the global brain structure prior constrained by the neural field equation and cortical geometry. This provides high-dimensional guiding EEG features that combine dynamic realism and brain anatomical rationality for the subsequent generation of iEEG signals, fundamentally supporting the high-fidelity reconstruction of iEEG-like signals in the three dimensions of time-domain waveform, spectral structure and brain spatial distribution.

[0130] S150: It uses EEG fusion features as guiding features for brain physiological consistency, and embeds an iEEG diffusion generation module guided by EEG high-dimensional fusion representation to guide the diffusion generation of iEEG signals, and finally outputs high-fidelity EEG enhancement signals.

[0131] In this step, to generate high-fidelity, high-signal-to-noise ratio iEEG signals from EEG fusion features that integrate prior knowledge of multi-source brain structures, this application proposes an iEEG diffusion generation module guided by multi-source EEG fusion representation. This module generates iEEG signals by integrating EEG fusion features... As a feature guiding brain physiological consistency, c is embedded in the iEEG diffusion generation module optimized for iEEG signal reconstruction to guide the diffusion generation of iEEG signals. During the training phase, the loss function not only includes temporal waveform reconstruction error but also introduces spectral consistency loss and spatial distribution consistency loss. During the inference phase, only the subject's EEG and T1 MRI are input to output a near-iEEG signal (i.e., an enhanced EEG signal) that highly approximates the real intracranial recording in terms of temporal dynamics, frequency domain energy, and spatial source distribution.

[0132] Specifically, the diffusion generation process of the iEEG diffusion generation module includes: setting the target EEG enhancement signal Y... The diffusion process of the iEEG-like signal to be generated is defined as follows:

[0133] (14)

[0134] in These are the preset noise scheduling parameters.

[0135] The reverse generation process is handled by a parameterized neural network specifically designed for iEEG reconstruction. Learning, its conditional generation distribution is:

[0136] (15)

[0137] Within this framework, EEG fusion features Instead of general auxiliary information, the training dynamically guides each step of the denoising process towards a direction consistent with the rationality of iEEG neural signals. This guidance is derived from the cross-modal neural electrical signal learning module described above. The training objective is to minimize the noise prediction loss.

[0138] (16)

[0139] To further constrain the clinical usability of the generated EEG signals, this application jointly optimizes two regularization terms for iEEG characteristics on top of the basic loss: spectral consistency loss and spatial distribution consistency loss. The spectral consistency loss forces the power spectral density (PSD) of the generated enhanced EEG signal to approximate the typical distribution of real iEEG in the δ–γ frequency band. The spatial distribution consistency loss verifies whether the projection of the generated iEEG source on the cortex is consistent with the input scalp EEG at the electromagnetic field level through a neural electrical signal forward computation model.

[0140] During the inference phase of EEG signal enhancement, only pure Gaussian noise needs to be considered. T Starting from N(0,I), the conditional denoising process is iteratively executed to generate a high-fidelity EEG enhancement signal:

[0141] (17)

[0142] This generation mechanism not only generates EEG-enhanced signals with iEEG-level spatiotemporal resolution, but also strictly preserves the individualized neural activity patterns contained in the original EEG signals. The final output iEEG-like signals closely approximate real intracranial recordings in terms of temporal dynamic evolution, frequency domain energy structure, and spatial distribution of cortical sources. It effectively overcomes the spatial ambiguity bottleneck caused by volume conduction in scalp EEG, providing a high-value and highly reliable alternative data source for clinical and research scenarios such as non-invasive brain function decoding, preoperative epileptogenic zone localization, and closed-loop neuromodulation.

[0143] It should be noted that other controllable generation mechanisms that can use multi-source brain structure priors as strong conditions and gradually optimize the spatiotemporal structure of EEG signals can also be used to generate enhanced EEG signals. As long as they can jointly constrain the consistency of time-domain waveform and spectral distribution during the training phase, and can output enhanced EEG signals that approximate real intracranial recordings by relying solely on EEG and T1 MRI during the inference phase, the same generation effect can be achieved.

[0144] Furthermore, this application has broad transfer potential and can be extended to other neural signal enhancement and cross-modal reconstruction tasks. It is not only applicable to EEG, but also to spatial resolution enhancement of MEG (magnetoencephalography) signals, deep tissue penetration enhancement of fNIRS (functional near-infrared spectroscopy) signals, or EEG-fMRI multimodal joint reconstruction. This application is not dependent on any specific task objective; the generated high-fidelity neural signals can serve as general representation inputs, supporting various downstream applications such as motion decoding, emotion recognition, cognitive state monitoring, neurofeedback training, and even brain-inspired artificial intelligence.

[0145] To evaluate the generation quality of the EEG enhancement signals in the embodiments of this application, the following embodiments employ evaluation methods such as a qualitative evaluation mechanism based on EEG experts, a multi-dimensional frequency domain quantitative evaluation mechanism for EEG signals, and an indirect effectiveness evaluation mechanism based on the performance of downstream tasks using EEG signals to assess the quality of the output EEG enhancement signals. The specific evaluation algorithms for each mechanism include:

[0146] A) Qualitative evaluation mechanism for EEG signals based on experts in the field of EEG: Experts in the field of EEG score each group of enhanced EEG signals across multiple dimensions based on typical interpretable characteristics such as waveform morphology, rhythmicity, artifact level, and spatial distribution consistency, forming a subjective credibility score. This score is used to assist in judging the rationality and usability of enhanced EEG signals in clinical or research applications. Assuming there are K experts, and each expert scores the j-th sample on the m-th feature dimension, then the subjective credibility score of that sample on that dimension is:

[0147] (18)

[0148] Furthermore, clinical importance weights were assigned to the M assessment dimensions. m yields a comprehensive subjective credibility score:

[0149] (19)

[0150] This comprehensive subjective credibility score provides interpretable expert prior verification for signal enhancement and reconstruction results, helping to determine their credibility in real-world application scenarios.

[0151] B) Multi-dimensional frequency domain quantitative evaluation mechanism for EEG signals: The frequency domain characteristics of EEG signals directly reflect the neural oscillation activity of the brain under different states and are the core basis for evaluating the physiological rationality of the signal. Therefore, this evaluation mechanism constructs three frequency domain quantitative evaluation indicators: power spectral density (PSD) matching degree, frequency phase consistency, and spectral distribution distance, systematically measuring the frequency domain consistency between the generated enhanced EEG signal and the real iEEG from different perspectives. Details are as follows:

[0152] Power spectral density (PSD) matching degree: EEG activity corresponds to different cognitive or pathological states in different frequency bands. Therefore, if the reconstructed EEG signal fails to reproduce the energy distribution of the real iEEG in these typical frequency bands, key neurophysiological information may be lost. To quantify this matching degree, the corresponding power spectral densities Preal(f) and Prec(f) of the real iEEG signal xreal(t) and the reconstructed EEG signal xrec(t) are calculated in typical frequency bands. The normalized power matching error of the EEG signal in frequency band B is defined as follows. This index is based on the Bhattacharyya coefficient, and the smaller the value, the more consistent the energy distribution of the EEG signal:

[0153] (20)

[0154] Frequency-phase consistency: The phase relationship of EEG oscillations is an important carrier of neural information transmission and functional connectivity. This application extracts the instantaneous phase by performing Hilbert transforms on the EEG signals before and after enhancement. and Calculate the cyclic variance of its phase difference, where ∈[0,1], the closer the value is to 1, the better the phase consistency of the EEG signal:

[0155] (twenty one)

[0156] In iEEG reconstruction tasks, high phase consistency means that the model not only restores the frequency domain energy of EEG signals, but also preserves the fine temporal structure of neural oscillations, which is crucial for downstream EEG tasks that rely on phase information.

[0157] Spectral Distribution Distance: Besides the local matching of discrete frequency bands in the EEG signal, the overall spectral shape of the EEG signal also contains important information about neurophysiological states. To evaluate the fidelity of the generated enhanced EEG signal in terms of global spectral morphology, the power of each frequency band of the enhanced EEG signal is normalized to a probability distribution. and The differences were quantified using the KL divergence metric.

[0158] (twenty two)

[0159] KL divergence is sensitive to the distribution tail of EEG signals and can effectively capture spectral distortions in EEG signals. In iEEG reconstruction, low... This indicates that the enhanced EEG signal not only closely approximates the real signal in terms of energy and phase, but also retains its overall spectral dynamic characteristics, thereby enhancing its applicability in long-term monitoring or state tracking tasks of EEG signals.

[0160] C) Indirect Effectiveness Evaluation Mechanism Based on Downstream Task Performance of EEG Signals: The core value of EEG signals lies not only in their waveform or spectral morphology, but also in the neural functional information they carry, such as decodeable content related to cognitive states, pathological features, or neural modulation responses. Therefore, relying solely on signal-level similarity (such as waveform fidelity or spectral matching) is insufficient to prove the actual usability of EEG enhancement signals. To this end, this application employs an indirect effectiveness evaluation mechanism based on downstream task performance of EEG signals. Through the performance of downstream neural decoding tasks, the integrity and effectiveness of the signal at the functional information level are verified in reverse. Specifically, the downstream tasks are assumed to be typical EEG classification problems such as epileptic seizure detection, consciousness state recognition, and motor imagery classification. A unified learnable model is used, and the quality of the model output is quantified using general performance indicators.

[0161] Furthermore, due to limitations such as volume conduction, noise interference, and low spatial resolution, raw EEG signals may be masked, potentially obscuring key neural features. To verify whether the generated enhanced EEG signals effectively preserve or even enhance functional information, under the same model architecture, training strategy, and data partitioning, raw EEG signals and enhanced EEG signals were used as inputs respectively to obtain performance data. and The enhancement gain is defined as:

[0162] (twenty three)

[0163] like This indicates that the enhanced EEG signal did not introduce harmful distortions and may have improved the model's discrimination ability by highlighting task-related neural features. This result directly demonstrates the functional advantages of enhanced EEG signals in practical applications, rather than simply making them "look better" in terms of signal morphology.

[0164] Meanwhile, iEEG, due to its high spatiotemporal resolution, is considered the "gold standard" signal for neural decoding and clinical diagnosis. The core objective of this application's non-invasive iEEG reconstruction method is to generate a substitute signal that is functionally equivalent to the real iEEG. To this end, in the same downstream task, both the real iEEG and the reconstructed iEEG were used as model inputs to obtain performance data. and The functional fidelity error is defined as:

[0165] (twenty four)

[0166] If error If the value is less than the preset threshold τ, the reconstructed iEEG can be considered to be highly consistent with the real iEEG in terms of key information for the task, and has the potential to replace the real iEEG for clinical decision-making or scientific research analysis.

[0167] This application has undergone systematic experiments on publicly available datasets to verify its feasibility, covering two aspects: quality assessment of iEEG-like signal reconstruction and practicality verification of enhanced EEG signals in multiple downstream tasks. The results show that this application significantly outperforms existing technologies in both the fidelity of EEG signal generation and task performance. Firstly, regarding iEEG-like signal reconstruction, this application conducted reconstruction experiments and multi-dimensional consistency assessments on two publicly available synchronous EEG-iEEG paired datasets. The experiments used real, synchronously acquired scalp EEG and intracranial iEEG signals as input and supervision targets, utilizing the enhanced EEG signals generated by the framework of this application, and conducting experiments from quantitative and qualitative perspectives such as temporal similarity, frequency distribution consistency, and phase consistency. Experimental results show that the enhanced EEG signals generated by this application significantly outperform baseline methods in multiple metrics, verifying its effectiveness in approximating iEEG signals and its physiological plausibility. Secondly, regarding downstream task decoding performance, this application uses the enhanced EEG signals as input and applies them to several classic EEG-based decoding tasks, including motor imagery classification, emotion state recognition, and epileptic seizure prediction. The above results consistently demonstrate that the enhanced EEG signal generated in this application is not only closer to iEEG at the signal level, but also effectively improves the performance of downstream decoding tasks, thus possessing clear practical value.

[0168] Please see Figure 2 This is a schematic diagram of the structure of a brain-computer interface signal enhancement and evaluation system for fusing data and knowledge, according to an embodiment of this application. The brain-computer interface signal enhancement and evaluation system 40 for fusing data and knowledge, according to an embodiment of this application, includes:

[0169] General EEG Representation Extraction Module 41: Used to encode raw EEG signals and obtain general EEG semantic representations;

[0170] Low-pass filtering feature extraction module 42: used to perform low-pass filtering on the original EEG signal to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise;

[0171] Head electrical signal calculation module 43: It is used to construct a three-layer conductive head model using T1 MRI signals, and solve the mapping relationship between EEG sensor signals and cortical surface current density by combining the regularized minimum norm estimation framework. It constructs a differential equation system that couples the forward electromagnetic field propagation model with the cortical mean field neural dynamics, and obtains the differential equation driving characteristics that reflect the spatiotemporal distribution of neural activity.

[0172] Geometric Calculation Module 44: Used to perform Laplace-Beltramian operator feature decomposition on the cortical surface reconstructed by individual T1 MRI, obtain K different frequency bands of brain geometric eigenmodes, and project all brain geometric eigenmodes onto the EEG electrode space to generate multi-scale, frequency domain decoupled geometric prior features.

[0173] EEG feature fusion module 45: used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features and geometric prior features using a multi-scale brain physiological feature fusion strategy to generate EEG fusion features;

[0174] iEEG diffusion generation module 46: used to use the EEG fusion features as brain physiological consistency guidance features, and to embed the iEEG diffusion generation module to guide the diffusion generation of iEEG signals, and finally output the EEG enhancement signal.

[0175] It should be noted that since the information interaction and execution process between the system embodiments of this application and the above-mentioned methods / devices / modules / units are based on the same concept, their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here.

[0176] Based on the above, the brain-computer interface signal enhancement and evaluation method and system of this application, which integrates data and knowledge, adopts a unified EEG representation enhancement framework that deeply integrates data-driven and mathematical prior knowledge. It achieves efficient alignment and fusion of heterogeneous features through four-way parallel feature extraction and employs stacked dilated convolution and multi-head attention mechanisms. Finally, based on the fused representation, it drives a diffusion model to generate high-fidelity EEG enhancement signals. This application achieves end-to-end generation from non-invasive scalp EEG to high-fidelity iEEG signals without relying on invasive recording, significantly improving the performance of EEG enhancement signals in terms of temporal fidelity, spectral consistency, and the rationality of cortical spatial distribution. It effectively bridges the representation gap between non-invasive EEG and invasive iEEG, providing reliable technical support for high-precision non-invasive brain-computer interfaces and neuroscience research.

[0177] Please see Figure 3 This is a schematic diagram of the device structure according to an embodiment of this application. The device 50 includes:

[0178] Memory 51 storing executable program instructions;

[0179] Processor 52 connected to memory 51;

[0180] Processor 52 is used to call executable program instructions stored in memory 51 and perform the following steps: Encoding the raw EEG signal using a general EEG representation extraction module to obtain a general EEG semantic representation; performing low-pass filtering on the raw EEG signal using a low-pass filtering feature extraction module to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise; constructing a three-layer conductive head model using T1 MRI signals, and solving the mapping relationship between EEG sensor signals and cortical surface current density using a regularized minimum norm estimation framework, constructing a differential equation system coupling the forward electromagnetic field propagation model and cortical mean field neurodynamics, and obtaining differential equation driving features reflecting the spatiotemporal distribution of neural activity; and performing individual T1 MRI on the head. The cortical surface reconstructed by MRI is subjected to Laplace-Beltramian eigenvalue decomposition to obtain K different frequency bands of brain geometric intrinsic modes. All brain geometric intrinsic modes are projected onto the EEG electrode space to generate multi-scale, frequency-domain decoupled geometric prior features. The general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features are dynamically fused using a multi-scale brain physiological feature fusion strategy to generate EEG fusion features. The EEG fusion features are used as brain physiological consistency guiding features and embedded in the iEEG diffusion generation module to guide the diffusion generation of iEEG signals to obtain enhanced EEG signals.

[0181] The processor 52 can also be referred to as a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0182] Please see Figure 4This is a schematic diagram of the structure of the storage medium in an embodiment of this application. The storage medium in this embodiment stores program instructions 61 capable of implementing the following steps: encoding the raw EEG signal using a general EEG characterization extraction module to obtain a general EEG semantic characterization; performing low-pass filtering on the raw EEG signal using a low-pass filtering feature extraction module to obtain baseline EEG features that retain the main neural rhythms and remove non-neurogenic high-frequency noise; constructing a three-layer conductive head model using T1 MRI signals, and solving the mapping relationship between EEG sensor signals and cortical surface current density using a regularized minimum norm estimation framework, constructing a differential equation system coupling the forward electromagnetic field propagation model and cortical mean field neurodynamics, and obtaining differential equation driving features reflecting the spatiotemporal distribution of neural activity; and performing individual T1 MRI on the raw EEG signal. The cortical surface reconstructed by MRI is subjected to Laplace-Beltrami operator eigenvalue decomposition to obtain K different frequency bands of brain geometric intrinsic modes. All brain geometric intrinsic modes are projected onto the EEG electrode space to generate multi-scale, frequency-domain decoupled geometric prior features. A multi-scale brain physiological feature fusion strategy is used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features to generate EEG fusion features. These EEG fusion features are used as brain physiological consistency guiding features and embedded in an iEEG diffusion generation module to guide the diffusion generation of iEEG signals, resulting in enhanced EEG signals. The program instructions 61 can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage media include: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program instructions, or terminal devices such as computers, servers, mobile phones, and tablets. Servers can be independent servers or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.

[0184] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for enhancing and evaluating brain-computer interface signals by integrating data and knowledge, characterized in that, include: The raw EEG signals were encoded using a general EEG representation extraction module to obtain general EEG semantic representations; The original EEG signal is processed by a low-pass filtering feature extraction module to obtain baseline EEG features that preserve neural rhythms and remove non-neurogenic high-frequency noise. A three-layer conductive head model was constructed using T1 MRI signals, and the mapping relationship between EEG sensor signals and cortical surface current density was solved by combining a regularized minimum norm estimation framework. A differential equation system was constructed that couples the forward electromagnetic field propagation model with the cortical mean field neural dynamics, and the differential equation driving characteristics reflecting the spatiotemporal distribution of neural activity were obtained. Laplace-Beltramian operator feature decomposition was performed on the cortical surface reconstructed by individual T1 MRI to obtain K different frequency bands of brain geometric intrinsic modes. All brain geometric intrinsic modes were projected onto the EEG electrode space to generate multi-scale, frequency domain decoupled geometric prior features. A multi-scale brain physiological feature fusion strategy is used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features to generate EEG fusion features. The EEG fusion features are used as guiding features for brain physiological consistency, and an iEEG diffusion generation module is embedded to guide the diffusion generation of iEEG signals to obtain enhanced EEG signals.

2. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to claim 1, characterized in that, The general EEG representation extraction module uses a specially designed and trained large-scale EEG model as the bone intervention training model, and completes pre-training on a large-scale public EEG dataset. The general EEG representation extraction module encodes the raw EEG signals to obtain general EEG semantic representations, specifically as follows: Given raw EEG signals in C For the number of channels, T Given the number of time sampling points, a high-level general EEG semantic representation is extracted using stacked, depth-separable modules: in For feature dimensions.

3. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to claim 2, characterized in that, The step of performing low-pass filtering on the original EEG signal through the low-pass filtering feature extraction module to obtain baseline EEG features that preserve neural rhythms and remove non-neurogenic high-frequency noise specifically involves: The cutoff frequency is set to zero phase distortion IIR or FIR filter. Baseline EEG characteristics were obtained while preserving neural rhythms and removing non-neurogenic high-frequency noise: in Indicates the cutoff frequency as Low-pass filtering operation.

4. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to claim 3, characterized in that, The method utilizes T1 MRI signals to construct a three-layer conductive head model, and combines a regularized minimum norm estimation framework to solve the mapping relationship between EEG sensor signals and cortical surface current density. This constructs a differential equation system coupling the forward electromagnetic field propagation model and cortical mean-field neural dynamics, obtaining the differential equation driving characteristics reflecting the spatiotemporal distribution of neural activity. Specifically: Based on individual T1 MRI images, head tissue structures are segmented to construct an individualized head model. The forward operator characterizing the electromagnetic conduction relationship between scalp electrodes and cortical sources is denoted as G∈ ,in N The number of discrete source points in the cortex; Introducing the continuous neural field equations describing the evolution of mean field electrical activity in the cortex: in This represents the potential state vector of the cerebral cortex. τ It is a time constant. W For local connectivity kernels, σ( () is the sigmoid activation function. I(t) For external input; The operator G for forward propagation of the EEG signal is coupled with the continuous neural field equation, and the relationship between scalp observation and internal state is defined as follows: in x(t) For theoretical scalp EEG, (t) For noise; Through embedded numerical simulation or inverse consistency optimization, the potential state trajectory or residual correction quantity consistent with the laws of neuroelectromagnetic dynamics is deduced from the actual observed scalp EEG sequence X, and the differential equation driving features are extracted from this. 。 5. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to claim 4, characterized in that, The Laplace-Beltramian operator feature decomposition is performed on the cortical surface reconstructed from individual T1 MRI to obtain K different frequency bands of brain geometric eigenmodes. All brain geometric eigenmodes are then projected onto the EEG electrode space to generate multi-scale, frequency-domain decoupled geometric prior features. Specifically: Reconstructing the axial surface of the cerebral cortex using individual T1 MRI And define LBO on a non-Euclidean brain surface: in Tensor for measuring the sulci and gyri of the cerebral cortex. For definition in Scalar functions on the LBO; by solving the eigenvalue problem, a set of orthogonal and complete brain geometric eigenmodes is obtained. and their corresponding brain modality features : ， Select the decomposition results A number of brain geometric intrinsic modalities, covering anatomical scales significantly correlated with iEEG spatial distribution, are grouped according to the feature values ​​of these brain modalities to construct a multi-scale brain geometric prior feature map: in The number of source points in the cerebral cortex; The cortical-electrode mapping module maps all brain geometric intrinsic modes to the scalp EEG electrode space, and performs interpolation alignment synchronized with the EEG signal along the time dimension to obtain geometric prior features aligned with the EEG time sequence. 。 6. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to any one of claims 1 to 5, characterized in that, The method utilizes a multi-scale brain physiological feature fusion strategy to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features to generate EEG fusion features. Specifically: Joint temporal coding using L-layer stacked convolutions for local fine-grained modeling and global context awareness. The general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features are then concatenated along the EEG channel dimension and aligned with the time dimension to form the initial fused EEG features: ： Final output The input is a cross-modal neural electrical signal learning module, which performs feature-level alignment based on the EEG semantic context of the current time step to generate structurally consistent and information-complete EEG fusion features; wherein, the core calculation of the cross-modal neural electrical signal learning module is: 。 7. The brain-computer interface signal enhancement and evaluation method for fusing data and knowledge according to claim 6, characterized in that, The process involves using the EEG fusion features as guiding features for brain physiological consistency, embedding an iEEG diffusion generation module to guide the diffusion generation of iEEG signals, and ultimately outputting an enhanced EEG signal. Specifically: Let the target EEG enhancement signal Y be... The iEEG-like signal to be generated is defined as follows: in These are the preset noise scheduling parameters; The reverse generation process is achieved by a parameterized neural network. Learning, its conditional generation distribution is: The training objective is to minimize the noise prediction loss. The spectral consistency loss and spatial distribution consistency loss are jointly optimized on the basis of the base loss. The power spectral density of the EEG enhancement signal generated by the spectral consistency loss is forced to approximate the typical distribution of real iEEG in the δ–γ frequency band. The spatial distribution consistency loss is verified by the forward computation model of the neural electrical signal to see whether the projection of the generated iEEG source on the cortex is consistent with the input scalp EEG at the electromagnetic field level. During the inference phase of EEG signal enhancement, from pure Gaussian noise T Starting from N(0,I), the conditional denoising process is iteratively executed to generate a high-fidelity EEG enhancement signal: 。 8. A brain-computer interface signal enhancement and evaluation system that integrates data and knowledge, characterized in that, include: General EEG Representation Extraction Module: Used to encode raw EEG signals and obtain general EEG semantic representations; Low-pass filtering feature extraction module: used to perform low-pass filtering on the original EEG signal to obtain baseline EEG features that preserve neural rhythms and remove non-neurogenic high-frequency noise; Head electrical signal calculation module: It is used to construct a three-layer conductive head model using T1 MRI signals, and solve the mapping relationship between EEG sensor signals and cortical surface current density by combining a regularized minimum norm estimation framework. It constructs a differential equation system that couples the forward electromagnetic field propagation model with the cortical mean field neural dynamics, and obtains the differential equation driving characteristics that reflect the spatiotemporal distribution of neural activity. Geometric Calculation Module: Used to perform Laplace-Beltramian operator feature decomposition on the cortical surface reconstructed by individual T1 MRI, obtain K different frequency bands of brain geometric eigenmodes, and project all brain geometric eigenmodes onto the EEG electrode space to generate multi-scale, frequency domain decoupled geometric prior features; EEG Feature Fusion Module: Used to dynamically fuse the general EEG semantic representation, baseline EEG features, differential equation-driven features, and geometric prior features using a multi-scale brain physiological feature fusion strategy to generate EEG fusion features; iEEG diffusion generation module: used to use the EEG fusion features as brain physiological consistency guidance features, and to embed the iEEG diffusion generation module to guide the diffusion generation of iEEG signals, and finally output the enhanced EEG signal.

9. A device, characterized in that, The device includes a processor and a memory coupled to the processor, wherein, The memory stores program instructions; when the program instructions are executed by the processor, the processor performs the brain-computer interface signal enhancement and evaluation method for fusing data and knowledge as described in claim 1.

10. A storage medium, characterized in that, The system stores processor-executable program instructions for performing the brain-computer interface signal enhancement and evaluation method for fusing data and knowledge as described in claim 1.

Images