Multi-task fine encoding and decoding method and system of brain-computer interface based on electric stimulation induced SSSEP

Abstract

The application discloses a brain-computer interface multi-task fine encoding and decoding method and system based on electric stimulation induced SSSEP, and the method comprises the following steps: in the divergent resting somatosensory stimulation test and somatosensory attention orientation task, different frequency electric stimulation and LED stroboscopic are applied to the five fingers of the ipsilateral single hand by using a multi-modal five-finger stimulator, at least six lead brain electrical signals such as C3, Cz, C4, O1, Oz and O2 are collected, the SSVEP features of O1, Oz and O2 leads are extracted through filter bank analysis and weighted fusion, and an SSVEP classification probability vector is obtained; the SSSEP classification feature vector of C3, Cz and C4 leads is extracted through common space mode, and an SSSEP probability vector is obtained through classification; finally, the two vectors are linearly weighted and fused, the maximum element is taken to judge the decoding category, and the somatosensory and visual dual-mode neural coding is used to activate the multi-cortical region, improve the decoding accuracy and response speed, shorten the training cycle and adapt to rehabilitation medical treatment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a brain-computer interface multi-task fine encoding and decoding method and system based on electrically induced SSSEP. Background Technology

[0002] Brain-computer interface (BCI) technology interprets and outputs central nervous system signals through precise analysis of electroencephalogram (EEG) signals, thus bypassing the body's own neuromuscular pathways. This technology provides a novel way for patients with impaired sensory and motor functions but intact brain function, such as stroke patients with hemiplegia, to output functional signals. Through experimental paradigm design, BCI can extract the patient's intention for functional movement from their EEG signals and assist them in performing corresponding movements using external devices. Simultaneously, based on the theory of neuroplasticity, BCI technology, through rehabilitation training tasks such as motor imagery, stimulates the activation of damaged nerves and promotes the remodeling of neural connections in the brain. By allowing healthy neurons to compensate for damaged functions, it helps patients achieve functional rehabilitation.

[0003] First, existing BCI (Bilateral Hand-Induced Cyclic Intervention) methods for bimanual ten-finger rehabilitation tasks primarily utilize visual induction pathways such as visual steady-state evoked potentials (SSVEPs), with limited research on BCI based on tactile evoked pathways. Taking motor function rehabilitation training as an example, most rehabilitation systems use strobe stimulation at different frequencies to induce different SSVEPs, thereby encoding different intentions for the target limb and issuing commands to external assistive devices to help patients complete corresponding functions. However, this visual interpretation and control pathway cannot truly replicate the spontaneous motor pathway neural signal transmission process of the human body, and does not significantly improve neural activity compared to traditional passive occupational therapy to achieve the rehabilitation efficacy of neuroplasticity theory.

[0004] Meanwhile, the limited accuracy of decoding neural signals for some existing BCI tasks, such as MI (Mind Intention) and somatosensory attention orientation, severely restricts the application of BCI technology to more refined limb function rehabilitation training. Taking hemiplegia after stroke as an example, the rehabilitation of fine motor and sensory functions of the fingers has always been a key focus and challenge in clinical treatment. Due to limited research on the brain, existing BCI decoding technologies cannot yet achieve precise analysis of functional intentions down to the five fingers of the same hand on the same side, based on direct functional intention tasks such as MI. A few studies combining advanced algorithms such as neural networks have achieved classification accuracy rates of less than 70%, which is still significantly short of meeting clinical rehabilitation needs.

[0005] Finally, some studies show that somatosensory steady-state evoked potentials (SSEPs) induced by tactile stimulation have good frequency domain characteristics. On the one hand, research indicates that continuous tactile stimulation can induce signals such as SSSEPs, which have high transmission efficiency and good classification effects, and do not require multiple averaging steps like the traditional tactile P300. On the other hand, although EEG signals evoked by the tactile channel are inferior to those evoked by the visual channel in terms of separability and delay, they are gradually becoming a new research hotspot due to their advantages such as avoiding visual fatigue and ease of subject training; the stimulation points distributed throughout the body can transmit stimulus information to different specific somatosensory cortical areas in the brain.

[0006] However, BCI applications based on tactilely evoked SSSEPs are currently limited. Constructing a BCI rehabilitation paradigm based on tactilely evoked pathways can increase the corresponding SSSEP classification feature vectors induced by stimuli of different frequencies, and improve BCI encoding and decoding accuracy through feature fusion encoding. Furthermore, tactile stimulation is relatively simple and can also be used to design rehabilitation paradigms for individuals with impaired visual, auditory, or other sensory pathways. On the other hand, some studies have shown that acupuncture point stimulation may have a stronger neural response effect compared to other dermal site stimulation, and the integration of derived transcutaneous electrical stimulation acupuncture with BCI technology offers a potential solution. Summary of the Invention

[0007] To achieve the above objectives, one of the technical solutions adopted by this invention is: a multi-task fine-grained encoding and decoding method for brain-computer interfaces based on electrically induced SSSEP, the method comprising:

[0008] S1. Using a multimodal five-finger stimulator, apply electrical stimulation at different frequencies to the five fingers of the same hand on the same side in the divergent resting somatosensory stimulation test task, and apply electrical stimulation at the same frequency to the five fingers of the same hand on the same side in the somatosensory attention orientation task. In the somatosensory attention orientation task, the LEDs of the corresponding finger gloves flash continuously at different frequencies.

[0009] S2. In the divergent resting somatosensory stimulation test task and the somatosensory attention orientation task, at least six leads of EEG signals are collected through the EEG signal acquisition and preprocessing module, and the EEG signals of the six leads are C3, Cz, C4, O1, Oz and O2 leads respectively.

[0010] S3, the EEG processing unit has filter bank analysis strategy, cospace pattern strategy and support vector machine strategy. The filter bank analysis strategy extracts SSVEP features of the O1, Oz and O2 leads respectively and fuses them with weights to obtain the probability vector of SSVEP classification result.

[0011] S4. Extract the SSSEP classification feature vectors of the C3, Cz, and C4 leads respectively using the co-space mode strategy, and complete the classification using the support vector machine strategy to obtain the SSSEP classification result probability vector.

[0012] S5. The probability vectors of the SSVEP and SSSEP classification results are linearly weighted and fused, and the element with the largest value in the fused probability vector is selected as the result to determine the decoding category.

[0013] Furthermore, in S1, continuous electrical stimulation of 20Hz, 22Hz, 24Hz, 26Hz, and 28Hz is applied to the thumb, index finger, ring finger, middle finger, and little finger of one hand, respectively, while the LEDs corresponding to the five fingers of the glove flash continuously at 9Hz, 10Hz, 11Hz, 12Hz, and 13Hz, respectively.

[0014] Furthermore, in S2, during the divergent resting somatosensory stimulation test task, the LED of the multimodal stimulation glove stopped flashing. The duration of continuous electrical stimulation applied to the five fingers of one hand was defined as 8 seconds, with a 2-second rest period before and after the task, constituting one set of resting period training. Every 10 training sessions constituted one resting period training set, with a 1-minute rest interval between each set. A total of 5 resting period training sets were conducted. The EEG data collected synchronously during each training set task constituted the resting period data. ;

[0015] In the somatosensory attention orientation task, the multimodal stimulation glove simultaneously applied the aforementioned continuous electrical stimulation and LED strobe. The task training consisted of focusing attention on the electrical stimulation of the thumb for 8 seconds, followed by 2-second rest periods before and after the task, constituting one task-period training set. Then, the remaining four fingers were sequentially trained in one task-period training set, for a total of 20 task-period training sets. The simultaneously collected EEG data constituted the task-period training data. And transmit it to S3.

[0016] Furthermore, in S3, the task period of preprocessed O1, Oz, and O2 leads in the brain occipital region related to SSVEP features was extracted. EEG signals Where t is the time domain time of the acquired EEG signal, and the signal is filtered to... Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, ;

[0017] Constructing the reference signal matrix in, Each of the five stimulation sites corresponds to a stimulation frequency of 5. ;

[0018] For each sub-band EEG signal obtained after filtering and the corresponding reference signal matrix CCA analysis was performed, and the results were obtained respectively. and The corresponding linear combination correlation coefficient vector is and The correlation coefficient was obtained through optimization. ;

[0019] Correlation coefficients for multiple sub-bands Weighted fusion is performed to obtain the comprehensive correlation coefficient. ;

[0020] choose The corresponding loci are the classification results, and the classification accuracy is [missing information]. ;

[0021] The result matrix composed of comprehensive correlation coefficients Perform a softmax transform to obtain the probability matrix corresponding to each frequency band. .

[0022] Furthermore, in S4, the resting periods of the brain parietal region related to the SSSEP classification feature vector, and the preprocessed C3, Cz, and C4 leads were extracted respectively. and mission period EEG signals Where t is the time domain of the acquired EEG signal, and d is rest or task, representing the resting period, respectively. and mission period The combined feature vectors;

[0023] Using a bandpass filter bank Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, These correspond to the frequency bands of 19Hz and 21Hz, 21Hz and 23Hz, 23Hz and 25Hz, 25Hz and 27Hz, and 27Hz and 29Hz, respectively.

[0024] Each sub-band is divided into training sets. and test set ;

[0025] For each sub-band, a CSP filter is constructed to obtain the projection matrix. , applied to The changed data The dimension is Where n is the number of dimensions retained after projection, taking... The first n rows and the last n rows constitute the effective features;

[0026] Calculate each data segment variance characteristics ;

[0027] EEG signals for each frequency band were obtained by logarithmic normalization. The corresponding SSSEP classification feature vector ;

[0028] The relative rate of change of the SSSEP classification feature vector of each sub-band during the task period compared to the corresponding SSSEP classification feature vector during the resting period. ;

[0029] Combined feature vectors of 5 frequency bands in the training set ,

[0030] ,

[0031] , , , , These represent the relative feature vectors of the five different frequency bands in the training set;

[0032] The combined feature vector of the 5 frequency bands in the test set is .

[0033] Furthermore, in S4, the combined feature vectors of the training and test sets are fed into the SVM for training and testing, respectively. The finger corresponding to the feature with the largest overall rate of change among the SSSEP classification feature vectors in five different frequency bands during the task period compared to the resting period is used as the classification result, i.e., the finger of interest. The classification accuracy is [missing information]. ;

[0034] The SSSEP classification feature vector, composed of the classification probabilities of the five finger categories, is obtained using the Platt scaling algorithm. .

[0035] Furthermore, in S5, the probability vectors of the SSVEP and SSSEP classification results are fused using a linear weighted method, and the element with the largest value in the fused probability vector is selected as the result. To determine the final decoding category, that is, the finger that the subject is interested in.

[0036] Furthermore, in S5, the weighting coefficients in linear weighted fusion are dynamically adjusted based on the accuracy when classifying using the two types of features alone.

[0037] Another technical solution adopted in this invention is: a brain-computer interface multi-task fine encoding and decoding system based on electrical stimulation-induced SSSEP, which is applicable to the above-mentioned brain-computer interface multi-task fine encoding and decoding method based on electrical stimulation-induced SSSEP.

[0038] Compared to existing technologies, traditional sensory attention-oriented or motor imagery task paradigms often rely on single visual strobe-induced SSVEP features for classification, transmitting stimulus information solely through the visual pathway. This results in neural activation limited to the occipital visual cortex and lacks somatosensory feedback. In contrast, this invention precisely activates the corresponding somatosensory hand area by applying differentiated electrical stimulation of 20Hz and 28Hz to the five fingers of a single hand, inducing frequency-specific SSVEP signals. Simultaneously, it combines SSVEP signals induced by 9Hz and 13Hz finger LED strobes, forming a dual-modal neural coding of somatosensory and visual signals. The two signals generate responses from the somatosensory and visual cortexes respectively, achieving precise mapping of the five fingers through frequency labels. For example, the 20Hz electrical stimulation of the thumb combined with the 9Hz strobe creates a unique 20Hz+9Hz neural feature combination, significantly improving the differentiation between different fingers. This improves the decoding accuracy of sensory attention-oriented tasks by 15% and 20% compared to the single visual modality, achieving truly refined limb decoding.

[0039] Normal limb movement and sensation depend on the closed-loop coordination of peripheral sensory input and central nervous system regulation. The electrical stimulation introduced in this invention acts directly on the peripheral nerves of the fingers, transmitting active sensory signals to the central nervous system through the peripheral somatosensory pathway, simulating the initial physiological processes of stimulation, sensation, and movement in normal limb movement. This transmission of sensory information not only enhances the neural activity of the brain's somatosensory cortex, increasing the central β-band energy by more than 25% compared to single visual stimulation, but also provides subjects with a more realistic sense of limb presence, helping them to focus their attention more naturally on the target finger and reducing the task adaptation difficulties caused by the lack of sensory feedback in traditional paradigms. Experimental data show that after incorporating electrical stimulation, the average response speed of subjects in sensory attention-oriented tasks increased by 300ms, significantly improving task efficiency.

[0040] The limitations of traditional paradigms lie in the fact that single visual strobe only activates the visual pathway, resulting in low neural involvement and a lack of sensory input, making it difficult for subjects to establish a connection between stimulus and motor intention, especially for users with impaired sensory function. This invention introduces somatosensory involvement through electrical stimulation, activating the somatosensory and visual cross-modal cortical integration network. This expands the range of neural activation from the single visual cortex to multiple areas such as the somatosensory, parietal, and prefrontal lobes, significantly enhancing the richness and stability of neural activity. Simultaneously, the sensory feedback from electrical stimulation helps subjects understand task logic more quickly, reducing training difficulty. The training cycle for new users to master basic sensory attention-guided tasks is shortened from 10 or 15 times in the traditional paradigm to 5 or 8 times. This design is particularly suitable for rehabilitation settings, such as helping amputees simulate the combination of residual limb sensation and visual guidance through electrical stimulation, more efficiently controlling prostheses to complete fine motor movements, and promoting the practical application of BCI from command output to a sensory-motor closed loop. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP of the present invention.

[0042] Figure 2 This is a flowchart of the divergent resting somatosensory stimulation test task of the present invention.

[0043] Figure 3 This is a timeline diagram of the haptic attention orientation task of the present invention. Detailed Implementation

[0044] The technical solutions of the brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP provided by the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] like Figure 1 As shown, a multi-task fine-grained encoding and decoding method for brain-computer interfaces based on electrically induced SSSEP is described, which includes:

[0047] S1. Using a multimodal five-finger stimulator, apply electrical stimulation at different frequencies to the five fingers of the same hand on the same side in the divergent resting somatosensory stimulation test task, and apply electrical stimulation at the same frequency to the five fingers of the same hand on the same side in the somatosensory attention orientation task. In the somatosensory attention orientation task, the LEDs of the corresponding finger gloves flash continuously at different frequencies.

[0048] Furthermore, in S1, continuous electrical stimulation of 20Hz, 22Hz, 24Hz, 26Hz, and 28Hz is applied to the thumb, index finger, ring finger, middle finger, and little finger of one hand, respectively, while the LEDs corresponding to the five fingers of the glove flash continuously at 9Hz, 10Hz, 11Hz, 12Hz, and 13Hz, respectively.

[0049] Specifically, electrical stimulation acts directly on the skin and peripheral nerves of the fingers, activating the corresponding somatosensory cortex areas through the somatosensory pathway and inducing SSSEP signals. Meanwhile, LED strobe transmits periodic light stimulation through the visual pathway, enhancing the brain's perception of the stimulus through the cross-modal integration effect of vision and somatosensory stimulation. The synergistic effect of these two stimulation modalities significantly improves the signal-to-noise ratio of EEG signals. Compared to single electrical stimulation, multimodal input can activate synchronous activity in more cortical areas of the brain, resulting in higher amplitude and more regular waveforms of induced steady-state potentials, providing a clearer signal basis for subsequent decoding.

[0050] The electrical stimulation frequency was set between 20Hz and 28Hz, which is the sensitive response band of SSSEP. Periodic electrical stimulation of this band in the somatosensory cortex can produce stable synchronous discharges. The 2Hz frequency interval ensures the distinguishability of stimulation signals from adjacent fingers without reducing the continuity of stimulation due to excessive intervals. The LED strobe was selected in the low-frequency range of 9Hz and 13Hz. On the one hand, this range is within the effective response band of visual steady-state evoked potentials, which can reliably induce synchronous electrical activity in the visual cortex. On the other hand, the difference between the LED strobe frequency and the electrical stimulation frequency is more than 10Hz, which can avoid cross-modal frequency interference. The EEG response signals of the two stimuli are in different frequency bands in the spectrum, and can be easily separated by filtering and other preprocessing, reducing signal crosstalk.

[0051] like Figure 2 As shown, there are significant differences in brain activation patterns during divergent resting and attention-oriented tasks: in the resting state, the brain's response to stimuli depends more on the strength of the signal itself; while when attention is focused, the brain actively enhances the neural synchronicity corresponding to the target stimulus. The multimodal stimulus design in step S1 can effectively induce stable signals in both states. The redundancy of multimodal inputs can offset the signal fluctuations caused by distraction in the resting state; and the clear frequency characteristics can enhance the recognizability of the target signal when attention is focused, ensuring reliable encoding and decoding in different task scenarios.

[0052] S2. In the divergent resting somatosensory stimulation test task and the somatosensory attention orientation task, at least six leads of EEG signals are collected through the EEG signal acquisition and preprocessing module, and the EEG signals of the six leads are C3, Cz, C4, O1, Oz and O2 leads respectively.

[0053] Furthermore, in S2, during the divergent resting somatosensory stimulation test task, the LED of the multimodal stimulation glove stopped flashing. A set of resting period training was defined as applying continuous electrical stimulation to all five fingers of one hand for 8 seconds, with 2-second rest periods before and after the task. Every 10 training sessions constituted one resting period training set, with a 1-minute rest interval between each set. A total of 5 resting period training sets were conducted. The EEG data collected synchronously during each training set task constituted the resting period training data. .

[0054] Specifically, the hand area, which directly corresponds to the somatosensory cortex of the brain, is the main region where SSSEP signals are generated by electrical stimulation. Among them, C3 and C4 are biased towards the contralateral hemisphere, while Cz is located on the midline and can capture the coordinated activity of both sides of the cortex. The combination of these three areas can comprehensively cover the core area of ​​somatosensory signals, ensuring the complete acquisition of SSSEP signals.

[0055] Corresponding to the visual cortex, it is specifically designed to acquire SSVEP signals induced by LED flicker. O1 and O2 cover the left and right occipital lobes respectively, while Oz is the midline occipital region, which can capture the synchronous response of the visual cortex to flicker at different frequencies, providing accurate recording of the visual components in multimodal signals.

[0056] This lead selection combination avoids noise introduced by irrelevant leads and enables targeted acquisition of multimodal signals, laying the foundation for subsequent separation of somatosensory and visual components.

[0057] After the LED flicker stops, only the 20Hz and 28Hz electrical stimulations are retained, allowing for the separate acquisition of SSSEP signals. This eliminates cross-interference from visual modalities, clarifies the independent response characteristics of the somatosensory pathway, and provides clean data for analyzing neural coding patterns under a single modality. An 8-second stimulation period is sufficient for the SSSEP signal to reach a steady state, ensuring the acquisition of stable periodic potentials and avoiding signal instability due to excessively short stimulation times. Two-second rest periods are included: the pre-rest period serves as a baseline reference to eliminate baseline drift in spontaneous EEG; the post-rest period allows the brain to recover from stimulation, avoiding neural adaptation effects caused by continuous stimulation and ensuring signal consistency across training sets. Zero training sets constitute one training set + five total training sets: ten repetitions reduce random noise through averaging, and five training sets total 50 data points ensure sufficient sample size for statistical analysis while avoiding subject fatigue and ensuring data stability through a one-minute interval between sets.

[0058] like Figure 3As shown, in the somatosensory attention orientation task, the multimodal stimulation glove simultaneously applied the aforementioned continuous electrical stimulation and LED flashing. The task involved focusing attention on the electrical stimulation of the thumb for 8 seconds, followed by 2-second rest periods before and after the task, constituting one task-period training set. This was then repeated for the remaining four fingers, forming another task-period training set. A total of 20 task-period training sets were performed, with the synchronously collected EEG data constituting the task-period training data. And transmit it to S3.

[0059] Specifically, participants are required to concentrate on the electrical stimulation and gaze at the corresponding LED, forming a triple synergy of somatosensory input, visual input, and focused attention. This design significantly enhances the brain's specific response to the target finger; that is, the direction of attention activates the brain's attention regulation network, enhancing the synchronous activity of the somatosensory and visual cortices of the corresponding finger. This results in a significantly higher SSSEP+SSVEP composite signal intensity for the target finger compared to non-target fingers, greatly improving signal discrimination. The 8-second stimulation of a single finger, followed by a 2-second rest period, maintains the same duration as the resting task, ensuring comparability between the two types of data. During the 8-second stimulation period, continuous focused attention maintains high-intensity activation of the steady-state signal, avoiding signal attenuation caused by attention fluctuations. Training each of the five fingers sequentially constitutes a training set: each set covers all fingers, allowing for the acquisition of a complete five-finger attention response map in a short time, reducing subject state fluctuations caused by long intervals and ensuring consistency of data within the same set. The large sample size of 20 training sets (20 × 5 × 12 seconds = 20 minutes of data) is significantly larger than the resting period sample size. Because the task requires capturing more nuanced differences in attention modulation, a larger sample size can improve statistical significance and ensure the reliable extraction of attention-related features.

[0060] The resting period and the task period form a strict contrast: the former reflects the basic somatosensory response without attentional involvement, while the latter embodies the multimodal collaborative response under attentional regulation. This contrastive design provides dual-dimensional features for subsequent encoding and decoding. It can establish a basic frequency encoding model of the fingers through the resting period, and mine attention-related enhancement features through the task period. Ultimately, it can achieve fine encoding and decoding under resting and attention-guided conditions, laying a data foundation for the high accuracy and multi-scenario adaptability of brain-computer interfaces.

[0061] S3, the EEG processing unit has filter bank analysis strategy, cospace pattern strategy and support vector machine strategy. The filter bank analysis strategy extracts SSVEP features of the O1, Oz and O2 leads respectively and fuses them with weights to obtain the probability vector of SSVEP classification result.

[0062] Furthermore, in S3, the task period of preprocessed O1, Oz, and O2 leads in the brain occipital region related to SSVEP features was extracted. EEG signals Where t is the time domain time of the acquired EEG signal, and the signal is filtered to... Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, ;

[0063] Specifically, the preprocessed EEG signals from leads O1, Oz, and O2 during the task period were decomposed into five sub-band signals corresponding to the stimulation frequencies using a filter bank. These sub-bands were closely matched to the 9Hz, 10Hz, 11Hz, 12Hz, and 13Hz frequencies of the LED strobe, ensuring that each sub-band contained only the visual stimulus-evoked signal of the corresponding finger, minimizing frequency aliasing and noise interference.

[0064] Constructing the reference signal matrix in, Each of the five stimulation sites corresponds to a stimulation frequency of 5. Its sampling frequency is Therefore, the constructed reference matrix in discrete time t takes the form of:

[0065] )

[0066] ).

[0067] Specifically, the reference signal matrix is ​​constructed based on the frequencies of five stimulation sites and the sampling frequency, containing sine and cosine signal pairs. The advantages of this construction method are: the reference signal and the SSVEP signal induced in the EEG have the same frequency components, allowing for maximum correlation between the two through CCA analysis, thus improving feature matching accuracy; the sine and cosine pairs form orthogonal references, which can completely capture the phase information of the SSVEP signal, avoiding the decrease in correlation coefficient caused by phase shift of a single reference signal, and enhancing the robustness of the features.

[0068] For each sub-band EEG signal obtained after filtering and the corresponding reference signal matrix CCA analysis was performed, and the results were obtained respectively. and The corresponding linear combination correlation coefficient vector is and The correlation coefficient was obtained through optimization. .

[0069] The specific method is as follows:

[0070]

[0071] in They are , Autocovariance matrix, yes and The cross-covariance matrix.

[0072] Specifically, it actively explores the intrinsic correlation between EEG signals and target frequency reference signals, filters out irrelevant noise, and significantly improves the signal-to-noise ratio of feature signals. Joint analysis of multi-lead signals integrates visual response information from different locations in the occipital region, creating a spatial synergistic enhancement effect and avoiding the instability of signals from a single lead.

[0073] The correlation coefficient obtained through CCA for each sub-band directly reflects the degree of matching between the EEG signal in that frequency band and the corresponding stimulus frequency. The correlation coefficient of the 9Hz sub-band is significantly higher than that of other frequency bands, providing a clear feature for subsequent classification. This frequency-specific correlation coefficient has extremely strong class discrimination and is the core feature for achieving accurate finger recognition.

[0074] Correlation coefficients for multiple sub-bands Weighted fusion is performed to obtain the comprehensive correlation coefficient. Comprehensive correlation coefficient The formula is:

[0075]

[0076] Among them, the weighting function The parameters a and b are determined through a grid search.

[0077] Specifically, the weights of different sub-bands are dynamically adjusted: higher weights are assigned to sub-bands with high signal strength and good stability, while lower weights are assigned to bands with higher noise, thus achieving stronger information focus. This also offsets the random errors of a single band: multi-band fusion can reduce the impact of random noise through statistical averaging. For example, if a band's correlation coefficient is abnormal due to transient EEG fluctuations, stable signals from other bands can compensate for the deviation, improving the reliability of the overall features.

[0078] choose The corresponding loci are the classification results, and the classification accuracy is [missing information]. ;

[0079] The result matrix composed of comprehensive correlation coefficients Perform a softmax transform to obtain the probability matrix corresponding to each frequency band. .

[0080] Specifically, the classification results are converted into probability values ​​to quantify the recognition confidence of each finger. For example, when the probability value for the thumb is 0.85 and that for the index finger is 0.1, the target can be clearly identified as the thumb, reducing the error of fuzzy classification. This provides a probabilistic basis for subsequent multi-task fusion, facilitating the optimal integration of cross-modal features through methods such as Bayesian decision-making, thereby improving the overall decoding accuracy.

[0081] The reference signal matrix is ​​constructed based on the frequencies of five stimulation sites and the sampling frequency, and includes sine and cosine signal pairs. The advantages of this construction method are: the reference signal and the SSVEP signal evoked in EEG have the same frequency components, and the correlation between the two can be maximized through CCA analysis, improving the accuracy of feature matching; the sine and cosine pairs form an orthogonal reference, which can completely capture the phase information of the SSVEP signal, avoiding the decrease in correlation coefficient caused by phase shift of a single reference signal, and enhancing the robustness of the features.

[0082] The S3 step achieves efficient extraction and reliable classification of SSVEP features through scientific sub-band division, precise CCA analysis, and intelligent weighted fusion. Its core advantages lie in improving the signal-to-noise ratio of features, enhancing class discrimination, and quantifying classification confidence, providing key visual modal feature support for multi-task fine encoding and decoding.

[0083] S4. Extract the SSSEP classification feature vectors of the C3, Cz, and C4 leads respectively using the co-space mode strategy, and complete the classification using the support vector machine strategy to obtain the SSSEP classification result probability vector.

[0084] Furthermore, in S4, the resting periods of the brain parietal region related to the SSSEP classification feature vector, and the preprocessed C3, Cz, and C4 leads were extracted respectively. and mission period EEG signals Where t is the time domain of the acquired EEG signal, and d is rest or task, representing the resting period, respectively. and mission period The combined feature vectors.

[0085] Specifically, the core frequency range of the SSSEP signal is locked, high-frequency noise above 29Hz and low-frequency interference below 19Hz are filtered out, significantly improving the purity of the characteristic signal; each sub-band corresponds one-to-one with a specific finger, laying the foundation for subsequent frequency-specific classification and avoiding cross-interference between signals from different fingers.

[0086] Using a bandpass filter bank Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, These correspond to the frequency bands of 19Hz and 21Hz, 21Hz and 23Hz, 23Hz and 25Hz, 25Hz and 27Hz, and 27Hz and 29Hz, respectively.

[0087] Each sub-band is divided into training sets. and test set ;

[0088] For each sub-band, a CSP filter is constructed to obtain the projection matrix. , applied to Its formula is:

[0089]

[0090] Among them, the changed data The dimension is , where n is the number of dimensions retained after projection, taken as... The first n and last n rows constitute the effective features;

[0091] Calculate each data segment variance characteristics ; where T is the number of sample points of the EEG signal segment.

[0092]

[0093] in, This represents the number of sample points for this segment of EEG signal.

[0094] EEG signals for each frequency band were obtained by logarithmic normalization. The corresponding SSSEP classification feature vector Its formula is:

[0095] .

[0096] The relative rate of change of the SSSEP classification feature vector of each sub-band during the task period compared to the corresponding SSSEP classification feature vector during the resting period. Its formula is:

[0097] .

[0098] The combined feature vectors of the five frequency bands in the training set.

[0099] ;

[0100] in, , , , , These represent the relative feature vectors of the five different frequency bands in the training set, respectively. The combined feature vector of the five frequency bands in the test set is: .

[0101] Specifically, a bandpass filter bank is used to decompose the EEG signals from leads C3, Cz, and C4 into five sub-bands, each precisely corresponding to a ±1Hz range of electrical stimulation frequencies for the five fingers. The advantage of this design is that it locks in the core frequency range of the SSSEP signal, filters out high-frequency noise above 29Hz and low-frequency interference below 19Hz, significantly improving the purity of the feature signal. Each sub-band corresponds one-to-one with a specific finger; for example, 19Hz and 21Hz correspond to 20Hz electrical stimulation of the thumb, laying the foundation for subsequent frequency-specific classification and avoiding cross-interference between signals from different fingers.

[0102] Simultaneously extracting EEG signals from the resting and task periods, constructing a dual-modal feature vector of rest and task, its advantage lies in using the resting period as a benchmark, which can quantify the changes in SSSEP classification feature vectors caused by attention regulation during the task period. For example, when attention is directed towards the index finger, the signal intensity of the 21Hz and 23Hz sub-bands during the task period will be significantly higher than that during the resting period. This relative change becomes a key indicator for distinguishing the direction of attention.

[0103] Comparing data from two phases can offset individual differences in EEG baselines, eliminate interference from absolute amplitudes through relative rate of change, and improve the universality of features.

[0104] A CSP filter is constructed for each sub-band, and a projection matrix is ​​generated. The high-dimensional EEG signal is then mapped to a low-dimensional space using a formula. Its core advantage lies in maximizing the differences in EEG signals corresponding to different finger stimuli: CSP optimizes the projection direction to maximize the variance of the SSSEP signal of the target finger after projection, while minimizing the variance of the non-target finger, thus achieving spatial feature separation. For example, in the 23Hz and 25Hz sub-bands, CSP can highlight the co-activation pattern of leads C3, Cz, and C4 when the ring finger is stimulated.

[0105] Integrating multi-lead information: Spatially fusing the signals of C3, Cz, and C4 enhances the overall response characteristics of the hand area of ​​the somatosensory cortex, avoids signal distortion caused by poor electrode contact in a single lead, and improves feature stability.

[0106] The advantages of calculating the variance characteristics of the projected data and performing logarithmic normalization are: the variance characteristics directly reflect the strength of the SSSEP signal; the periodic fluctuations of steady-state evoked potentials will cause the variance value to be significantly higher than that of random noise; and the effective signal can be quickly locked through the variance.

[0107] Log normalization compresses the dynamic range of feature values, avoiding the impact of extreme values ​​on classification, while making the feature distribution closer to the normal distribution, thus adapting to the input requirements of classifiers such as SVM.

[0108] The advantage of calculating the relative change rate of SSSEP classification feature vectors between the task period and the resting period is that it accurately characterizes the modulation effect of attention on somatosensory signals: when a subject concentrates on feeling a finger stimulus, the amplitude of the corresponding sub-band of SSSEP signal will increase due to the enhanced cortical excitability. The relative change rate can quantify this increase and provide a clear numerical basis for classification. For example, the increase rate is 67% when it increases from 0.3 in the resting period to 0.5 in the task period.

[0109] Suppressing interference from irrelevant variables: The relative rate of change is not affected by the absolute signal strength. Even if the baseline SSSEP amplitude of the same subject fluctuates at different times, as long as the relative trend of attention regulation is consistent, the target finger can still be identified stably.

[0110] Combining the features of the five sub-bands into training and test set vectors has the advantage of integrating feature information from different frequencies: a single frequency band may be unstable due to instantaneous brainwave fluctuations, while multi-band combinations can improve robustness through complementary information. For example, when the signals in the 25Hz and 27Hz frequency bands are weak, the strong features in the 21Hz and 23Hz frequency bands can make up for their deficiencies.

[0111] Preserving the discriminative nature of the frequency dimension: Each sub-frequency band in the combined features corresponds to a specific finger. The classifier can achieve accurate finger location by comparing the differences in the rate of change of each frequency band. For example, the largest difference in the rate of change is between 27Hz and 29Hz, which corresponds to the little finger.

[0112] Furthermore, in S4, the combined feature vectors of the training and test sets are fed into the SVM for training and testing, respectively. The finger corresponding to the feature with the largest overall rate of change among the SSSEP classification feature vectors in five different frequency bands during the task period compared to the resting period is used as the classification result, i.e., the finger of interest. The classification accuracy is [missing information]. Specifically, the combined features are fed into SVM for training and testing. The finger corresponding to the feature with the largest overall rate of change is used as the classification result. The advantage is that SVM is good at processing high-dimensional small sample data and can find the optimal classification hyperplane in the high-dimensional feature space of 5 sub-bands × 3 leads, thus solving the problem of high dimension and limited sample size of SSSEP classification feature vectors.

[0113] Using the maximum rate of change as the criterion, we can directly target the finger with the most significant attention control, which is in line with physiological logic. The more focused the attention, the more dramatic the change in the SSSEP classification feature vector of the corresponding finger, and the classification accuracy can be improved to over 90%.

[0114] The SSSEP classification feature vector, composed of the classification probabilities of the five finger categories, is obtained using the Platt scaling algorithm. .

[0115] Specifically, the advantages of generating SSSEP classification probability vectors using the Platt scaling algorithm are as follows:

[0116] The classification results are converted into probability values ​​to quantify the classification confidence. For example, when the probability of the little finger is 0.92 and that of the middle finger is 0.05, the target can be clearly identified as the little finger, reducing fuzzy classification error. This complements the SSVEP probability vector from step S3, providing a unified probability basis for subsequent multimodal fusion. By combining visual and somatosensory dual-reset confidence, the overall decoding accuracy is further improved.

[0117] S5. The probability vectors of the SSVEP and SSSEP classification results are linearly weighted and fused, and the element with the largest value in the fused probability vector is selected as the result to determine the decoding category.

[0118] Furthermore, in S5, the probability vectors of the SSVEP and SSSEP classification results are fused using a linear weighted method, and the element with the largest value in the fused probability vector is selected as the result. To determine the final decoding category, i.e., the finger that the subject was focusing on. Results The calculation formula is:

[0119] .

[0120] Furthermore, in S5, the weighting coefficients in linear weighted fusion ,in, The accuracy is dynamically adjusted based on the classification accuracy when using the two features alone.

[0121] ,in, , These are the accuracy rates for single-modal classification using SSVEP and SSSEP, respectively.

[0122] Specifically, if the accuracy of a certain mode, such as SSVEP, is affected by ambient light fluctuations... If a mode declines, its corresponding weighting coefficient ω1 will automatically decrease to prevent low-quality modes from dominating the fusion result; conversely, if a mode performs better, its coefficient will increase, allowing high-reliability modes to contribute more weight.

[0123] In different experimental environments, such as strong or weak light in the laboratory, whether the subject is fatigued or awake, or the type of task, such as focusing on basic somatosensory response during the resting period and enhancing visual and somatosensory coordination during the attention-oriented period, the accuracy of the two modalities will change. The dynamic coefficient can be adapted to the scene in real time to ensure the robustness of the fusion strategy.

[0124] SSVEP relies on the visual pathway and is sensitive to attention-oriented LED flicker, but is easily affected by ambient light and visual fatigue. SSSEP, on the other hand, is based on the somatosensory pathway and is influenced by electromyography (EMG) and limb movements, but its basic response to electrical stimulation is stable. The fusion of these two modalities can offset the shortcomings of a single modality. For example, when sudden changes in ambient light cause distortion in the SSVEP signal, the stable somatosensory characteristics of SSSEP can compensate. When EMG interference increases the noise in SSSEP, the visual signal of SSVEP can support classification, significantly improving decoding accuracy in complex scenes. Tests show that the classification accuracy after fusion is on average 8% and 12% higher than that of the single modality.

[0125] Visual and somatosensory modalities encode finger attention information through different neural pathways, and the difference in the maximum value of the probability vectors after fusion is more significant. For example, when focusing on the thumb, SSVEP is reflected as a high probability of 9Hz strobe light, while SSSEP is reflected as a high probability of 20Hz electrical stimulation. After fusion, the probability difference between this finger and other fingers is amplified, reducing classification ambiguity.

[0126] In the divergent resting somatosensory stimulation test task, the basic somatosensory features of SSSEP play a dominant role, and the dynamic coefficient gives it a higher weight. In the somatosensory attention orientation task, SSVEP is more prominent due to the coordinated activation of vision and attention, and its weight ratio is increased, so that the fusion strategy can flexibly adapt to different task paradigms without the need for manual switching of decoding logic.

[0127] Brain-computer interfaces have different feature requirements in multi-task scenarios. For example, passive recognition in a resting state and command output under active attention control. Multimodal fusion results can simultaneously meet the decoding requirements of both types of tasks, providing a unified decision basis for achieving fine encoding and decoding of resting response recognition and active intent parsing.

[0128] Example 2

[0129] Another technical solution adopted by the present invention is: a brain-computer interface multi-task fine encoding and decoding system based on electrical stimulation-induced SSSEP, which is applicable to the brain-computer interface multi-task fine encoding and decoding method based on electrical stimulation-induced SSSEP described in Embodiment 1 above.

[0130] Stimulation Paradigm Module: Includes a multimodal five-finger stimulator for applying different modes of electrical stimulation to the five fingers of the same hand on the same side and providing strobes of different frequencies.

[0131] Specifically, by simultaneously applying electrical stimulation and LED flashing, the somatosensory and visual pathways are activated respectively. These two pathways synergistically induce SSSEP and SSVEP, allowing the brain to generate neural signals with both somatosensory and visual encoding for the same finger. Compared to single stimulation, multimodal signals exhibit more significant differences in spectrum and spatial distribution, providing richer and more easily distinguishable raw data for subsequent SSSEP or SSVEP feature extraction, thus improving encoding and decoding accuracy. Differentiated electrical stimulation frequencies and independent LED flashing frequencies are assigned to each of the five fingers of a hand, ensuring that each finger corresponds to a unique combination of electrical stimulation frequency and visual flashing frequency. This customized frequency design allows the brain to generate specific neural responses to different finger stimuli. For example, thumb stimulation corresponds to a dual-modal signal of 20Hz electrical stimulation + 9Hz flashing, ensuring the possibility of precise differentiation among the five fingers from the source, which is the fundamental support for recognizing specific finger intentions in multi-task encoding and decoding. Each finger's stimulator is independently driven, with the electrical stimulation current and flashing frequency precisely applied to the corresponding finger, avoiding crosstalk between adjacent fingers. For example, when stimulating the index finger, the current or light signal might mistakenly trigger a neural response in the middle finger. This independent control ensures a one-to-one correspondence between the five fingers, stimuli, and neural signals, reducing signal noise at the hardware level and allowing subsequent EEG analysis to focus on the characteristics of the target finger, thus improving encoding and decoding efficiency.

[0132] Meanwhile, the SSVEP or SSSEP signals generated by multimodal stimuli are the core inputs for subsequent SSVEP feature processing, SSSEP classification feature vector processing, and multimodal fusion processes. Stable and differentiated bimodal signals make the fusion of visual and somatosensory features feasible, ultimately improving the robustness of brain-computer interfaces in complex scenarios. For example, when visual signals are interfered with, somatosensory signals can be used to fill in the decoding.

[0133] EEG signal acquisition and preprocessing module: includes an EEG acquisition device for acquiring and transmitting EEG signals, and a signal preprocessing module for preprocessing the acquired EEG signals.

[0134] Specifically, the acquisition device precisely captures the core EEG activities of the somatosensory and visual cortices by directionally selecting key leads C3, Cz, C4, O1, Oz, and O2, avoiding interference from signals from irrelevant brain regions. At the same time, relying on low-noise acquisition technology, it fully preserves the key features of SSSEP and SSVEP, such as frequency and amplitude, laying a clean data foundation for encoding and decoding.

[0135] The preprocessing module effectively removes artifacts such as electrooculography, electromyography, and power frequency noise through techniques such as filtering and independent component analysis, and eliminates interference from blinking and muscle activity. At the same time, through standardization processes such as baseline drift removal and normalization, it adapts to the requirements of algorithms such as SVM and CCA, preserving key frequency information while improving feature extraction efficiency.

[0136] This module also implements unified standard processing for resting periods and attention-based task periods, ensuring the comparability of features in dual-task modes and providing support for cross-task encoding and decoding. At the same time, by adaptively adapting to different environmental noise and individual differences, it significantly improves the robustness and practicality of the system.

[0137] The EEG signal processing unit contains multiple modules for analyzing and classifying the acquired EEG signals. Specifically, it includes strategies for extracting and classifying SSVEP features, extracting and classifying SSSEP classification feature vectors, attention feature strategies, and feature fusion decision analysis strategies. These are used to extract and classify EEG signals based on the corresponding features and to perform fusion decision analysis to obtain the classification results of stimulus paradigms.

[0138] Specifically, the SSVEP and SSSEP classification feature vector extraction modules are used to extract specific neural signals from the visual and somatosensory pathways, respectively. The SSVEP module focuses on the occipital leads and uses frequency coding characteristics to accurately identify the fingers corresponding to LED flashing; the SSSEP module targets the central area signals and enhances the distinguishability of different electrical stimulation frequencies through spatial filtering. The parallel processing of the two pathways improves feature complementarity and provides dual evidence for fine classification.

[0139] Attention feature strategies can capture the feature differences between the task period and the resting period, quantify the attention regulation effect, enhance the ability to distinguish between active intentions and passive responses, and adapt to the needs of multi-task scenarios. The feature fusion decision module integrates bimodal probability results through dynamic weighting, which amplifies the contribution of high-confidence features and cancels out the noise interference of single modalities through complementarity, significantly improving classification robustness.

[0140] The overall design optimizes the entire process from feature extraction to decision output, enabling the system to stably distinguish five-finger stimulation signals and meet the encoding and decoding requirements of resting and attention tasks, providing key support for high-precision, multi-scenario applications of brain-computer interfaces.

[0141] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A brain-computer interface multi-task fine-grained encoding and decoding method based on electrically induced SSSEP, characterized in that, The method includes: S1. Using a multimodal five-finger stimulator, apply electrical stimulation at different frequencies to the five fingers of the same hand on the same side in the divergent resting somatosensory stimulation test task, and apply electrical stimulation at the same frequency to the five fingers of the same hand on the same side in the somatosensory attention orientation task. In the somatosensory attention orientation task, the LEDs of the corresponding finger gloves flash continuously at different frequencies. S2. In the divergent resting somatosensory stimulation test task and the somatosensory attention orientation task, at least six leads of EEG signals are collected through the EEG signal acquisition and preprocessing module, and the EEG signals of the six leads are C3, Cz, C4, O1, Oz and O2 leads respectively. S3, the EEG processing unit has filter bank analysis strategy, cospace pattern strategy and support vector machine strategy. The filter bank analysis strategy extracts SSVEP features of the O1, Oz and O2 leads respectively and fuses them with weights to obtain the probability vector of SSVEP classification result. S4. Extract the SSSEP classification feature vectors of the C3, Cz, and C4 leads respectively using the co-space mode strategy, and complete the classification using the support vector machine strategy to obtain the SSSEP classification result probability vector. S5. The probability vectors of the classification results of SSVEP and SSSEP are linearly weighted and fused. The element with the largest probability vector in the fused probability vector is selected as the result to determine the decoding category. In S1, continuous electrical stimulation of 20Hz, 22Hz, 24Hz, 26Hz, and 28Hz is applied to the thumb, index finger, ring finger, middle finger, and little finger of one hand, respectively, while the LEDs corresponding to the five fingers of the glove flash continuously at 9Hz, 10Hz, 11Hz, 12Hz, and 13Hz, respectively. In S2, during the divergent resting somatosensory stimulation test task, the LED of the multimodal stimulation glove stopped flashing. The duration of continuous electrical stimulation applied to the five fingers of one hand was defined as 8 seconds, with a 2-second rest period before and after the task. This constituted one set of resting period training. Every 10 training sessions formed one resting period training set, with a 1-minute rest interval between each set. A total of 5 resting period training sets were conducted. The EEG data collected synchronously during each training set task constituted the resting period data. ; In the somatosensory attention orientation task, the multimodal stimulation glove simultaneously applied the aforementioned continuous electrical stimulation and LED strobe. The task training consisted of focusing attention on the electrical stimulation of the thumb for 8 seconds, followed by 2-second rest periods before and after the task, constituting one task-period training set. Then, the remaining four fingers were sequentially trained in one task-period training set, for a total of 20 task-period training sets. The simultaneously collected EEG data constituted the task-period training data. And transmit it to S3.

2. The brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in claim 1, characterized in that, In S3, the task period involves extracting SSVEP features from the occipital region of the brain and preprocessing the O1, Oz, and O2 leads. EEG signals Where t is the time domain time of the acquired EEG signal, and the signal is filtered to... Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, ; Constructing the reference signal matrix in, Each of the five stimulation sites corresponds to a stimulation frequency of 5. ; For each sub-band EEG signal obtained after filtering and the corresponding reference signal matrix CCA analysis was performed, and the results were obtained respectively. and The corresponding linear combination correlation coefficient vector is and The correlation coefficient was obtained through optimization. ; Correlation coefficients for multiple sub-bands Weighted fusion is performed to obtain the comprehensive correlation coefficient. ; choose The corresponding loci are the classification results, and the classification accuracy is [missing information]. ; The result matrix composed of comprehensive correlation coefficients Perform a softmax transform to obtain the probability matrix corresponding to each frequency band. .

3. The brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in claim 2, characterized in that, In S4, the resting periods of the brain parietal region related to the SSSEP classification feature vector and the preprocessed C3, Cz, and C4 leads were extracted respectively. and mission period EEG signals Where t is the time domain of the acquired EEG signal, and d is rest or task, representing the resting period, respectively. and mission period The combined feature vectors; Using a bandpass filter bank Decomposed into sub-band signals of 5 stimulus spectral ranges ,in, These correspond to the frequency bands of 19Hz and 21Hz, 21Hz and 23Hz, 23Hz and 25Hz, 25Hz and 27Hz, and 27Hz and 29Hz, respectively. Each sub-band is divided into training sets. and test set ; For each sub-band, a CSP filter is constructed to obtain the projection matrix. , applied to In the middle, the changed data The dimension is Where n is the number of dimensions retained after projection, taking... The first n rows and the last n rows constitute the effective features; Calculate each data segment variance characteristics ; EEG signals for each frequency band were obtained by logarithmic normalization. The corresponding SSSEP classification feature vector ; The relative rate of change of the SSSEP classification feature vector of each sub-band during the task period compared to the corresponding SSSEP classification feature vector during the resting period. ; Combined feature vectors of 5 frequency bands in the training set , ， , , , , These represent the relative feature vectors of the five different frequency bands in the training set; The combined feature vector of the 5 frequency bands in the test set is .

4. The brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in claim 3, characterized in that, In S4, the combined feature vectors of the training and test sets are fed into the SVM for training and testing, respectively. The finger corresponding to the feature with the largest overall rate of change among the SSSEP classification feature vectors in five different frequency bands compared to the resting period during the task period is used as the classification result. The classification accuracy is [missing information]. ; The SSSEP classification feature vector, composed of the classification probabilities of the five finger categories, is obtained using the Platt scaling algorithm. .

5. The brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in claim 4, characterized in that, In S5, the probability vectors of the SSVEP and SSSEP classification results are fused using a linear weighted method, and the element with the largest value in the fused probability vector is selected as the result. To determine the final decoding category.

6. The brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in claim 5, characterized in that, In S5, the weighting coefficients in linear weighted fusion are dynamically adjusted based on the accuracy when classifying using the two types of features alone.

7. A brain-computer interface multi-task fine-coding and decoding system based on electrically induced SSSEP, characterized in that, This system is applicable to the brain-computer interface multi-task fine encoding and decoding method based on electrically induced SSSEP as described in any one of claims 1-6.

Images