Distributed brain-computer interface system based on photon label and decoding method

Abstract

The invention discloses a distributed brain-computer interface system based on photon labels and a decoding method, and belongs to the technical field of brain-computer interfaces, the distributed brain-computer interface system comprises a distributed brain-computer interface system and a self-adaptive Riemann decoding algorithm, the system is composed of a plurality of discrete photon labels, electroencephalogram signal acquisition equipment and an edge computing terminal, a hidden stroboscopic stimulation normal form of'high-frequency carrier wave + pseudorandom sequence 'is adopted by a photon label to solve visual fatigue, and efficient decoding of electroencephalogram signals is achieved through Riemannian manifold preprocessing, Transform deep decoding and meta-learning zero sample calibration by means of an algorithm. According to the method, the problems of screen dependence, visual fatigue, difficulty in high-frequency signal decoding and tedious calibration of a traditional brain-computer interface are solved, natural interaction of what you see is what you get under the scene of the Internet of Things is realized, the decoding accuracy and usability are improved, and the method is suitable for application of consumer-level brain-computer interfaces.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to a distributed brain-computer interface system and decoding method based on photonic tags. Background Technology

[0002] Brain-computer interface technology based on steady-state visual evoked potentials (SSVEP) is a relatively mature brain-computer interaction solution. Its core is to trigger the brain's visual cortex to generate synchronous EEG signals by having the user look at a visual stimulus source that flashes at a specific frequency, and then to control external devices by decoding the signal.

[0003] Existing technologies primarily rely on computer screens or LCD displays as stimulus presentation carriers, which, while possessing some practicality, have significant drawbacks in consumer applications: (1) Limited interaction methods ("screen dependence"): Existing technologies require users to look at virtual icons on the screen to control external devices, which is "indirect interaction" and severs the connection between users and the physical world. It is impossible to achieve intuitive control of "what you see is what you get" in the Internet of Things scenario (such as directly looking at a desk lamp to turn it on and off). (2) Visual fatigue and safety hazards: In order to ensure the decoding accuracy, the existing system usually uses low-frequency strong flickering stimulation of 6Hz-15Hz. Long-term viewing of such visible flickering can easily cause visual fatigue, dizziness, and even induce photosensitive epilepsy in certain groups, which seriously affects user experience and safety. (3) Difficulty in decoding high-frequency weak signals: If the flicker frequency is increased to above 60Hz, which is invisible to the human eye, to alleviate fatigue, the amplitude of the SSVEP signal will drop sharply and the signal-to-noise ratio (SNR) will be extremely low. The traditional canonical correlation analysis (CCA) and its variant algorithms will have a significantly reduced recognition accuracy in such low SNR environments, and usually require a cumbersome user calibration process, which cannot meet the consumer application requirements of "wear and use". Summary of the Invention

[0004] The purpose of this invention is to propose a distributed brain-computer interface system and decoding method based on photonic tags to solve the problems mentioned in the background art: overcome the defects of existing SSVEP brain-computer interface technology such as "screen dependence", risk of visual fatigue, low decoding accuracy of high-frequency weak signals and the need for user calibration.

[0005] To achieve the above objectives, the present invention provides a distributed brain-computer interface system based on photonic tags, comprising: Several discrete photon tags, as independent visual stimuli, are attached to the surface of the physical entity to be controlled. EEG signal acquisition equipment is used to acquire EEG signals generated when a user gazes at a discrete photon tag of a target; The edge computing terminal is connected to several discrete photonic tags and EEG signal acquisition devices. The edge computing terminal decodes the received EEG signals, identifies the discrete photonic tags that the user is looking at, and generates corresponding control commands.

[0006] Preferably, each discrete photonic tag is an independent, adhesive micro-hardware module, including a microcontroller (MCU), an LED driver circuit, a Micro-LED array, a wireless communication module, and a power supply module. The physical entity includes smart home devices, industrial control terminals, assistive rehabilitation devices, or vehicle-mounted interactive terminals. The EEG signal acquisition device is a portable EEG acquisition device worn on the user's head. The edge computing terminal is a mobile phone or computer.

[0007] Preferably, the discrete photonic tag uses an invisible stroboscopic stimulation paradigm to solve visual fatigue. The specific process is as follows: the microcontroller (MCU) receives control commands through the wireless communication module and drives the Micro-LED light-emitting array to perform amplitude or phase modulation at a carrier frequency higher than the critical fusion frequency of the human eye, combined with a pseudo-random coding sequence, to generate stroboscopic stimulation that is invisible to the human eye.

[0008] A decoding method for a distributed brain-computer interface system based on photon tags, specifically including the following steps: S1: Signal Acquisition: Acquire multi-channel EEG signals from the user through an EEG signal acquisition device; S2: Riemannian manifold preprocessing: Convert multi-channel EEG signals into covariance matrices and map them to the Riemannian manifold space; calculate the geometric mean on the Riemannian manifold space; project the data on the Riemannian manifold space to the Euclidean tangent space through tangent space projection to obtain the denoised feature vectors. S3: Deep Decoding: Input the denoised feature vector into the trained spatiotemporal attention neural network model, and then output the identification information of the discrete photon tag of the target being gazed at by the user.

[0009] Preferably, the Riemannian manifold preprocessing in step S2 specifically includes the following steps: S21: Constructing an enhanced covariance matrix: Multichannel EEG signals are... ,in This indicates that the brain signals are all real numbers. Number of EEG signal channels To determine the number of EEG signal sampling points, an enhanced feature matrix is ​​constructed containing the original EEG signal and the adaptive time-delay EEG signal. : ; In complex interactive scenarios, a multi-scale splicing mode is adopted, and the formula is rewritten as follows: ; in, This represents the current EEG signal. For time-delayed EEG signals, adaptive time delay parameters It is obtained through the following formula: ; in, For the floor function, Sampling frequency, The arithmetic mean of the carrier frequencies of all photonic tags. , These are the time delay scale parameters corresponding to different carrier frequencies; Calculate the covariance matrix based on the Ledoit-Wolf shrinkage estimation method : ; in, The original empirical covariance matrix, Represents the average eigenvalue. For matrix trace operations, The shrinkage strength coefficient, It is the identity matrix; S22: Riemannian manifold space geometric metric: using the covariance matrix Mapping to a symmetric positive definite matrix manifold space The affine-invariant Riemannian metric is used to calculate the two covariance matrices in the manifold space. , Distance: ; in, For affine invariant Riemannian metric, Represents the matrix logarithm operation. Represents the covariance matrix The inverse square root, These are the eigenvalues ​​of the matrix. It is the Frobenius norm; S23: Calculate the Riemann geometric mean: Use the Fréchet mean algorithm to find the geometric center of all sample points in the Riemannian manifold space. : ; in, The optimal solution search operator, The number of sample points. This represents the baseline state of background EEG noise in the current environment; S24: Tangent Space Projection and Eigenvectorization: Transforming the Covariance Matrix using the Logarithmic Mapper Projected onto Euclidean tangent space at the tangent point To obtain the linear tangent vector : ; in, Represents the matrix logarithm operation. Represents the geometric mean matrix The square root, Represents the geometric mean matrix The inverse square root; right Perform a semi-vectorization operation to extract the upper triangular elements and construct the feature vector. : ; in, Extract operators for upper triangular elements.

[0010] Preferably, in step S3, the spatiotemporal attention neural network model is a neural network based on the Transformer architecture, which extracts spatiotemporal features from the feature vector through a multi-head self-attention mechanism.

[0011] Preferably, the spatiotemporal attention neural network model is trained using a meta-learning strategy to support zero-shot decoding for new users, eliminating the need for model calibration for new users.

[0012] Preferably, the meta-learning strategy includes domain adversarial training, used to align EEG features from different users to a common feature space.

[0013] Therefore, the present invention employs the above-described distributed brain-computer interface system and decoding method based on photonic tags, which has the following advantages: (1) Natural interaction: Abandoning the traditional screen carrier, the physical entity is directly controlled by looking at it through discrete photonic tags, achieving intuitive interaction of "what you see is what you get" in the Internet of Things scenario and improving ease of use; (2) Healthy and comfortable: It adopts high-frequency invisible strobe stimulation of 60Hz-100Hz, which is perceived by the human eye as a constant bright state without flickering, thus completely avoiding visual fatigue and the risk of photosensitive epilepsy; (3) High robustness and accuracy: By integrating Riemannian manifold geometry denoising and Transformer feature extraction technology, the feature extraction problem of high-frequency weak signals is effectively solved, and the decoding accuracy under low signal-to-noise ratio environment is significantly improved; (4) Zero-shot ease of use: Through the meta-learning zero-shot calibration strategy, new users can use it directly without going through a cumbersome calibration process, which lowers the threshold for using consumer-grade brain-computer interfaces.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating an application scenario of a distributed brain-computer interface system based on photonic tags mentioned in an embodiment of the present invention. Figure 2 This is a structural block diagram of a discrete photonic tag in a distributed brain-computer interface system based on photonic tags mentioned in an embodiment of the present invention; Figure 3 This is an invisible stroboscopic timing diagram of a distributed brain-computer interface system based on photonic tags mentioned in an embodiment of the present invention; Figure 4 This is a flowchart of the adaptive Riemann decoding algorithm in a decoding method for a distributed brain-computer interface system based on photon tags mentioned in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0017] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0018] Example like Figure 1-4 As shown, this embodiment provides a distributed brain-computer interface system based on photonic tags, including: several discrete photonic tags, an EEG signal acquisition device, and an edge computing terminal; Several discrete photonic tags, as independent visual stimuli, are attached to the surface of the physical entity to be controlled, including but not limited to smart home devices, industrial control terminals, assistive rehabilitation devices, or vehicle-mounted interactive terminals. The discrete photonic tags are independent, adhesive micro-hardware modules, including microcontrollers (MCUs), LED driver circuits, Micro-LED arrays, wireless communication modules, and power modules.

[0019] The electroencephalogram (EEG) signal acquisition device is a portable EEG acquisition device worn on the user's head, used to acquire the electroencephalogram signals generated when the user gazes at a discrete photon tag of a target. The edge computing terminal communicates with discrete photonic tags and EEG signal acquisition devices to receive and decode EEG signals, identify the discrete photonic tag the user is looking at, and generate corresponding control commands. The smartphone, acting as the edge computing terminal, synchronizes with all tags via Bluetooth and receives signals from the EEG device via Wi-Fi.

[0020] Discrete photonic tags use an invisible stroboscopic stimulation paradigm to address visual fatigue. The specific process is as follows: The microcontroller (MCU) receives control commands through the wireless communication module and simultaneously drives the Micro-LED light-emitting array to perform amplitude or phase modulation at a carrier frequency (60Hz-100Hz) higher than the critical fusion frequency of the human eye, combined with a pseudo-random coding sequence, to generate stroboscopic stimulation that is invisible to the human eye.

[0021] A decoding method for a photon-tagged distributed brain-computer interface system, applied to such a system, employs an adaptive Riemann decoding algorithm and includes the following steps: S1: Signal Acquisition: Acquire multi-channel EEG signals from the user through an EEG signal acquisition device; S2: Riemannian manifold preprocessing: Convert multi-channel EEG signals into covariance matrices and map them to the Riemannian manifold space; calculate the geometric mean on the Riemannian manifold space; project the data on the Riemannian manifold space to the Euclidean tangent space through tangent space projection to obtain the denoised feature vectors. S3: Deep Decoding: Input the denoised feature vector into the trained spatiotemporal attention neural network model, and then output the identification information of the discrete photon tag of the target being gazed at by the user.

[0022] The specific steps for Riemannian manifold preprocessing are as follows: (1) Construction of the enhanced super-covariance matrix: To capture the timing and phase characteristics of high-frequency carriers and pseudo-random sequences at different frequencies and to solve the parameter matching problem caused by the inconsistency of multiple target frequencies in distributed systems, this invention introduces multi-scale adaptive delay embedding technology and shrinkage estimation optimization mechanism.

[0023] Let the preprocessed multi-channel EEG signal be ,in This indicates that the brain signals are all real numbers. For the number of channels, The number of sampling points is used to construct an enhanced feature matrix containing the original signal and its adaptive time-delay signal. : ; Adaptive delay parameters Determination: To address the potential frequency differences between different photon tags in a distributed system (e.g., tag A is 60Hz, tag B is 75Hz), this invention does not set a fixed frequency. . The value is determined using a frequency band center adaptive strategy: set ,in Sampling frequency, This is the arithmetic mean of the carrier frequencies of all currently used photonic tags. Furthermore, in complex interactive scenarios, a multi-scale stitching mode can be further employed, i.e. By introducing multiple time delay scales (such as 1 / 4 cycles corresponding to 60Hz and 80Hz), it is ensured that the enhancement matrix can simultaneously cover the characteristic frequency bands of all tags in the system.

[0024] Covariance calculation based on Ledoit-Wolf contraction estimation: based on Calculate the sample covariance matrix. For high-frequency SSVEP signals in environments with extremely low signal-to-noise ratios, to prevent issues arising from traditional fixed regularization terms... To address the issues of excessively large covariance matrices masking signal features or excessively small covariance matrices resulting in non-positive matrices, this invention introduces the Ledoit-Wolf contraction estimation method to dynamically construct the covariance matrix. : ; in, The original empirical covariance matrix, Represents the average eigenvalue (shrinkage target). The shrinkage strength coefficient is calculated automatically. It is an identity matrix.

[0025] Adaptive adjustment mechanism of Ledoit-Wolf shrinkage estimation: shrinkage strength coefficient It is not a fixed value, but rather automatically calculated by an algorithm based on the statistical characteristics of the currently acquired EEG signal data (including signal variance, noise power, inter-channel correlation, etc.), and the value range is strictly limited to... Its adjustment logic is entirely driven by the data itself, requiring no manual intervention, thus achieving real-time perception and dynamic adaptation to the noisy environment.

[0026] Adaptive response in high-noise environments: When users are in environments with large fluctuations in ambient light intensity, severe electromagnetic interference (such as outdoors, in scenarios with multiple electronic devices), or unstable scalp impedance (such as loose electrodes, sweating, etc.) leading to a significant increase in noise, the algorithm will automatically identify the original empirical covariance matrix. The volatility increases and the dispersion of the eigenvalue distribution increases. At this time... It will approach 1, enhancing the contraction regularization effect: on the one hand, by increasing The weights will shift the covariance matrix towards the average eigenvalues. The corresponding stable structure shrinks to avoid noise causing the matrix to be non-positive definite or the features to be distorted; on the other hand, it reduces The weights are adjusted to suppress the interference of noise components in the original signal, ensuring that the covariance matrix can stably represent the core spatial correlation of the EEG signal, and providing reliable basic data for subsequent Riemannian manifold mapping.

[0027] Adaptive optimization in low-noise environments: When the user is in a quiet indoor environment with good electrode contact, the original empirical covariance matrix... It can accurately reflect the true characteristics of weak SSVEP signals. It will approach 0, weakening the contraction regularization strength: by increasing The weights are adjusted to retain the effective feature information in the original signal to the maximum extent, avoiding feature loss due to excessive shrinkage; at the same time, the weights are adjusted to retain the effective feature information in the original signal to the maximum extent, avoiding feature loss due to excessive shrinkage; It provides basic stability assurance, preventing minor disturbances to the covariance matrix structure caused by small amounts of noise, and achieving a balance between "fidelity" and "stability".

[0028] (2) Geometric metric of Riemannian manifold space: The covariance matrix generated above Mapping to a symmetric positive definite matrix manifold space In this space, the two covariance matrices The distance between them is no longer Euclidean distance, but an affine invariant Riemannian metric, whose distance formula is defined as: ; in It is a matrix The real eigenvalues. This metric naturally normalizes the signal power, thereby eliminating the impact of impedance differences caused by varying tightness of the fit among different users.

[0029] (3) Calculation of the Riemann geometric mean (reference benchmark): To perform the projection, the geometric center (i.e., the reference tangent point) of the current dataset must first be calculated. This invention employs the Fréchet mean algorithm, which solves the problem by iteratively minimizing the sum of squared distances from all sample points on the manifold to the mean. ; The geometric mean This represents the "background EEG noise baseline state" in the current environment.

[0030] (4) Tangent Space Mapping and Feature Vectorization: Using the log-map operator, the covariance matrix of each signal containing a weak SSVEP signal is... Projected onto tangent space at the tangent point Above. This step unfolds the nonlinear geometry on the curved manifold into linear tangent vectors in Euclidean space. : ; in This represents the matrix logarithm operation.

[0031] because Since it is still a symmetric matrix, it needs to be semi-vectorized to be used as input to the subsequent Transformer network. The upper triangular elements are then extracted to form the final feature vector. : .

[0032] The spatiotemporal attention neural network model is a Transformer-based neural network that extracts spatiotemporal features from feature vectors through a multi-head self-attention mechanism. The model is trained using a meta-learning strategy to support zero-shot decoding for new users, eliminating the need for model calibration for new users. This meta-learning strategy includes domain adversarial training, used to align EEG features from different users to a common feature space.

[0033] A specific implementation process is as follows: like Figure 1-4 As shown, this embodiment provides a distributed brain-computer interface system and decoding method based on photonic tags. The specific implementation process is as follows: (a) Implementation of discrete photonic tags: The discrete photonic tag adopts a micro-encapsulation design, measuring 2cm × 2cm × 0.5cm, making it easy to attach to various physical surfaces. The microcontroller (MCU) selected is STM32L476, which is responsible for receiving synchronization instructions and controlling the LED driver circuit; The wireless communication module uses Bluetooth 5.0 to achieve low-latency data transmission with the edge computing terminal; The Micro-LED array adopts a 3×3 layout and emits light at a wavelength of 505nm (green light, with optimal SSVEP response sensitivity). The power module is powered by a button battery, with a battery life of ≥30 days, and supports wireless charging.

[0034] (II) Parameter settings for the invisible stroboscopic stimulation paradigm: The high-frequency carrier frequency is set to 80Hz (higher than the critical fusion frequency of the human eye, ensuring no flicker perception). The pseudo-random sequence selected is an m-sequence of length 31 (long period, strong autocorrelation, easy to identify signal). The modulation method employs amplitude fine-tuning, with the adjustment amplitude being ±5% of the carrier amplitude, ensuring that it is imperceptible to the human eye while guaranteeing that the visual cortex can capture signal characteristics.

[0035] (III) Specific implementation of the adaptive Riemann decoding algorithm: Riemannian manifold pretreatment: The acquired 8-channel EEG signal (sampling rate 250Hz) was segmented, with each segment lasting 1 second and without overlap; Calculate the covariance matrix (8×8 dimensions) for each signal segment; The center vector of the manifold space is calculated based on the Riemann geometric mean, and the covariance matrix is ​​transformed into an 8×(16+1)=136-dimensional Euclidean eigenvector by tangent space projection. Transformer Spatiotemporal Attention Network: The network input is a 136-dimensional feature vector sequence (length 5, corresponding to 5 continuous signal segments). The number of heads in the multi-head self-attention mechanism is set to 4, and the hidden layer dimension is set to 128. Layer normalization and dropout (dropout rate 0.2) are used to prevent overfitting; Meta-learning zero-shot calibration: The general model library is pre-trained based on EEG data from 100 users, covering different age and gender groups; Domain adversarial training employs a gradient inversion layer (GRL) to align the feature distributions of new users with those of pre-trained users. Decoding latency ≤200ms, meeting the requirements of real-time interaction.

[0036] (iv) System workflow: The user wears a portable EEG acquisition device, and the edge computing terminal completes the device initialization and wireless synchronization. The user observes a discrete photonic tag on the surface of the target physical entity, and the tag outputs an invisible strobe signal with an 80Hz carrier superimposed with an m-sequence. EEG acquisition equipment collects SSVEP signals from the occipital lobe region and transmits them to the edge computing terminal; The terminal runs an adaptive Riemann decoding algorithm, sequentially completing Riemann manifold preprocessing, Transformer feature extraction, and meta-learning feature alignment, and outputs the label ID recognition result. The terminal sends control commands (such as switches or parameter adjustments) to the corresponding physical entity based on the recognition results to complete the interaction.

[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A distributed brain-computer interface system based on photonic tags, characterized in that, include: Several discrete photon tags, as independent visual stimuli, are attached to the surface of the physical entity to be controlled. EEG signal acquisition equipment is used to acquire EEG signals generated when a user gazes at a discrete photon tag of a target; The edge computing terminal is connected to several discrete photonic tags and EEG signal acquisition devices. The edge computing terminal decodes the received EEG signals, identifies the discrete photonic tags that the user is looking at, and generates corresponding control commands.

2. The distributed brain-computer interface system based on photonic tags according to claim 1, characterized in that: Each discrete photonic tag is an independent, adhesive micro-hardware module containing a microcontroller (MCU), LED driver circuit, Micro-LED array, wireless communication module, and power supply module. Physical entities include smart home devices, industrial control terminals, assistive rehabilitation devices, or in-vehicle interactive terminals. The EEG signal acquisition device is a portable EEG acquisition device worn on the user's head. The edge computing terminal is a mobile phone or computer.

3. The distributed brain-computer interface system based on photonic tags according to claim 2, characterized in that: Discrete photonic tags use an invisible stroboscopic stimulation paradigm to address visual fatigue. The specific process is as follows: The microcontroller (MCU) receives control commands through the wireless communication module and simultaneously drives the Micro-LED light-emitting array to perform amplitude or phase modulation at a carrier frequency higher than the critical fusion frequency of the human eye, combined with a pseudo-random coding sequence, to generate stroboscopic stimulation that is invisible to the human eye.

4. A decoding method for a distributed brain-computer interface system based on photonic tags, applied to the distributed brain-computer interface system based on photonic tags as described in any one of claims 1-3, characterized in that: The decoding method employs an adaptive Riemann decoding algorithm, which specifically includes the following steps: S1: Signal Acquisition: Acquire multi-channel EEG signals from the user through an EEG signal acquisition device; S2: Riemannian manifold preprocessing: Convert multi-channel EEG signals into covariance matrices and map them to the Riemannian manifold space; calculate the geometric mean on the Riemannian manifold space; project the data on the Riemannian manifold space to the Euclidean tangent space through tangent space projection to obtain the denoised feature vectors. S3: Deep Decoding: Input the denoised feature vector into the trained spatiotemporal attention neural network model, and then output the identification information of the discrete photon tag of the target being gazed at by the user.

5. The decoding method for a distributed brain-computer interface system based on photonic tags according to claim 4, characterized in that: The Riemannian manifold preprocessing in step S2 specifically includes the following steps: S21: Constructing an enhanced covariance matrix: Multichannel EEG signals are... ,in This indicates that the EEG signal is a real number. Number of EEG signal channels To determine the number of EEG signal sampling points, an enhanced feature matrix is ​​constructed containing the original EEG signal and the adaptive time-delay EEG signal. : ； In complex interactive scenarios, a multi-scale splicing mode is adopted, and the formula is rewritten as follows: ； in, This represents the current EEG signal. For time-delayed EEG signals, adaptive time delay parameters It is obtained through the following formula: ； in, For the floor function, Sampling frequency, The arithmetic mean of the carrier frequencies of all photonic tags. , These are the time delay scale parameters corresponding to different carrier frequencies; Calculate the covariance matrix based on the Ledoit-Wolf shrinkage estimation method : ； in, The original empirical covariance matrix, Represents the average eigenvalue. For matrix trace operations, The shrinkage strength coefficient, It is the identity matrix; S22: Riemannian manifold space geometric metric: using the covariance matrix Mapping to a symmetric positive definite matrix manifold space The affine-invariant Riemannian metric is used to calculate the two covariance matrices in the manifold space. , Distance: ； in, For affine invariant Riemannian metric, Represents the matrix logarithm operation. Represents the covariance matrix The inverse square root, These are the eigenvalues ​​of the matrix. It is the Frobenius norm; S23: Calculate the Riemann geometric mean: Use the Fréchet mean algorithm to find the geometric center of all sample points in the Riemannian manifold space. : ； in, The optimal solution search operator, The number of sample points. This represents the baseline state of background EEG noise in the current environment; S24: Tangent Space Projection and Eigenvectorization: Transforming the Covariance Matrix using the Logarithmic Mapper Projected onto Euclidean tangent space at the tangent point To obtain the linear tangent vector : ； in, Represents the matrix logarithm operation. Represents the geometric mean matrix The square root, Represents the geometric mean matrix The inverse square root; right Perform a semi-vectorization operation to extract the upper triangular elements and construct the feature vector. : ； in, Extract operators for upper triangular elements.

6. The decoding method for a distributed brain-computer interface system based on photonic tags according to claim 4, characterized in that: In step S3, the spatiotemporal attention neural network model is a neural network based on the Transformer architecture, which extracts spatiotemporal features from the feature vector through a multi-head self-attention mechanism.

7. The decoding method for a distributed brain-computer interface system based on photonic tags according to claim 4, characterized in that: The spatiotemporal attention neural network model is trained using a meta-learning strategy to support zero-shot decoding for new users, eliminating the need for model calibration for new users.

8. The decoding method for a distributed brain-computer interface system based on photonic tags according to claim 7, characterized in that: Meta-learning strategies include domain adversarial training, which is used to align EEG signal features from different users to a common feature space.

Images