Low-load and anti-fatigue asynchronous brain-controlled interface switching method

Abstract

The invention discloses a low-load and anti-fatigue asynchronous brain control interface switching method. The method comprises the following steps: S1, signal acquisition and preprocessing: performing signal acquisition by adopting a 16-bit ADC (Analog to Digital Converter) resolution and 256Hz sampling rate 8-channel dry electrode electroencephalogram cap; high-quality electroencephalogram signals are obtained through the signal collecting and preprocessing module, and then the multi-modal feature extraction module fuses the frequency domain delta beta / alpha energy ratio, the space domain C3-C4 coherence and the nonlinear gamma-band multi-scale entropy to generate robust feature vectors; after the vector is input into an asynchronous detection engine, intention state accurate awakening is achieved based on a dynamic threshold algorithm coupling false triggering history and environmental noise, meanwhile, personalized parameter calibration is completed within 5 minutes through an anti-fatigue initialization process of eye closing resting and multi-mode physiological verification guided by touch, and finally the false triggering rate lt is achieved; and the fatigue degree of the user is reduced by 60% + of the medical-level performance index, so that the use requirements are met.

Description

Technical Field

[0001] This invention relates to the field of asynchronous brain-controlled interface switching technology, and more particularly to a low-load, fatigue-resistant asynchronous brain-controlled interface switching method. Background Technology

[0002] Existing brain-controlled switches mostly adopt a synchronous paradigm (such as P300, SSVEP), which requires users to continuously focus on visual stimuli, resulting in two major drawbacks: (1) High cognitive load: It requires maintaining a state of continuous focus, and obvious fatigue occurs after an average of 20 minutes of operation. Although traditional asynchronous solutions reduce the load, they have problems with poor scene adaptability due to intention response delay (average >2s) and fixed awakening threshold; (2) High false trigger rate: Environmental noise and non-intentional EEG fluctuations can easily cause misoperation (such as alpha wave bursts causing false triggers). In view of the above, this application proposes a low-load, fatigue-resistant asynchronous brain-controlled interface switch method. Summary of the Invention

[0003] Based on the technical problems existing in the background technology, this invention proposes a low-load, fatigue-resistant asynchronous brain-controlled interface switching method.

[0004] The low-load, fatigue-resistant asynchronous brain-control interface switching method proposed in this invention includes the following steps:

[0005] S1: Signal acquisition and preprocessing: Signal acquisition was performed using an 8-channel dry electrode EEG cap with 16-bit ADC resolution and 256Hz sampling rate. The acquired EEG signals were then subjected to three levels of preprocessing, followed by signal standardization, environmental noise monitoring, segmentation, and buffering of the preprocessed data.

[0006] S2: Multimodal feature extraction: Extract three types of complementary features to form the intended feature vector F, whose complementary features include frequency domain features, spatial domain features and nonlinear features;

[0007] S3: Asynchronous Detection: Achieves zero-load in resting state and fast response in intent state through asynchronous detection engine, with detection latency ≤800ms. It adopts dynamic dual threshold mechanism and environmental adaptive compensation, maintaining a 94% intent detection rate while the false trigger rate is <0.5 times / hour under 65dB noise.

[0008] S4: Anti-fatigue initialization: Achieve physiological state calibration to adapt to individual differences through anti-fatigue initialization.

[0009] Preferably, in S1, the input impedance of the dry electrode EEG cap is >100MΩ, the common-mode rejection ratio is ≥110dB, which is used to suppress environmental interference, and the electrode layout strictly follows the international 10-20 system standard.

[0010] The specific steps for three-level preprocessing of the acquired EEG signals are as follows:

[0011] S101: An adaptive 50Hz notch filter is used to eliminate power frequency interference from the power grid. The formula used is as follows: Where w0 = 2π × 50 / 256 is the digital angular frequency, and r = 0.95 is the bandwidth control factor;

[0012] S102: Bandpass filter, using a zero-phase IIR filter for frequency band separation, high-pass cutoff of 0.5Hz to remove baseline drift, low-pass cutoff of 45Hz to suppress electromyographic noise, and retaining five key frequency bands: δ 0.5-4Hz, θ 4-8Hz, α 8-13Hz, β 13-30Hz, and γ 30-45Hz.

[0013] S103: Artifact removal is performed using an improved independent component analysis formula: X - AS + E, where X is the multi-channel observation signal matrix; A is the mixing matrix; S is the independent source signal; and E is the residual matrix.

[0014] Preferably, in step S1, two levels of standardization are required when performing signal standardization: reference electrode standardization and inter-channel energy equalization.

[0015] The formula used for its reference electrode standardization is: Where μTP9 is the mean value of the TP9 electrode within a 1-second sliding window; σTP10 is the standard deviation of the TP10 electrode within the same time window, and its function is to eliminate the systematic deviation caused by the change in the contact impedance of the reference electrode.

[0016] The formula used for energy balance between its channels is: Its function is to calculate the energy normalization factor for each channel i, so that the signals of each channel have comparable energy levels;

[0017] When monitoring environmental noise, a 0.5-10Hz bandpass filter is set in the Fz channel, and the signal variance within a 200ms time window is calculated using the following formula: And perform noise index classification, its low noise σ 2 <5μV 2 ; Medium noise 5≤σ 2 ≤15μV 2 ; High noise σ 2 >15μV 2 ;

[0018] To balance real-time performance and accuracy during segmentation and caching, an overlapping framing strategy is adopted, with a frame length of 200ms and a frame shift of 100ms. Meanwhile, the circular buffer retains the data of the most recent 10 frames.

[0019] Preferably, in S2, the frequency domain feature characterizes the energy difference ratio of the motor cortex. When a motor intention is generated, the β wave in the contralateral brain region desynchronizes and its energy decreases, while the β wave in the ipsilateral brain region synchronizes and its energy increases. The energy ratio of the left and right brain regions is used to offset the individual's absolute amplitude difference. The formula used is as follows: Where E β (C3) represents the energy of the β-band of channel C3, calculated as the sum of squares after wavelet packet decomposition from 13-30Hz; E α (C4) represents the energy of the α band of channel C4, calculated as the variance after 8-13Hz bandpass filtering; σ rest σ represents the resting-state standard deviation and the initial stage statistic. t The standard deviation of the real-time signal is calculated using a 200ms sliding window.

[0020] Spatial features are used to reflect brain region coordination and calculate motor cortical coherence. Motor cortical coherence refers to the degree of synchronization of neural oscillations in the C3 and C4 electrode regions of the left and right hemisphere motor cortexes. The physiological mechanism is that in the resting state, the left and right motor cortices coordinate highly through the corpus callosum, with coherence > 0.7; when intention is generated, the neural resources of the target brain region are locally concentrated, and the coherence decreases to 0.3-0.5. The specific calculation process is as follows:

[0021] S2011: Perform a short-time Fourier transform on the C3 / C4 channel signals and calculate the complex spectrum for each frame of the signal. The formula used is as follows:

[0022]

[0023] Where f represents the frequency point within the range of 18-22Hz; t represents the time frame index.

[0024] S2012: Calculate the spectral cross-correlation, which characterizes the phase relationship between two signals at frequency f. The formula used is: Where * denotes complex conjugate operation;

[0025] S2013: Coherence calculation employs a complex coherence algorithm to avoid spurious coherence caused by volume conduction. The formula is as follows: Where K is the average number of frames, with a default K=5, corresponding to 500ms of data;

[0026] S2014: Feature value extraction, calculation of the target frequency band mean, and dynamic normalization are performed. The formula used to calculate the target frequency band mean is as follows: The formula used for dynamic normalization is: Where μ rest σ represents the mean of the user's resting-state coherence. rest This represents the resting-state standard deviation.

[0027] S2015: Anti-interference design is implemented, limiting the frequency band to a narrow 18-22Hz band, avoiding the main frequency band of electromyographic noise >25Hz, and avoiding the θ / α frequency band, which is susceptible to electrooculography interference;

[0028] The nonlinear characteristic is the γ-band multi-scale entropy, used to capture complexity. When the human brain generates a motor intention, the firing pattern of the motor cortex neural clusters changes from a resting-state rhythmic oscillation to a non-steady-state chaotic mode. This change is particularly significant in the γ-band. The specific calculation process is as follows:

[0029] S2021: Multi-scale reconstruction, decomposing the signal into sequences with different time resolutions: Where τ is the scale factor (1,2,3), used for adjusting the temporal resolution; j is the index of the new sequence; The signal is reconstructed at the τ scale to characterize neural activity at different time windows;

[0030] S2022: Scale entropy calculation, calculating improved sample entropy for each scale sequence: sampEn Among them B τ (r) represents the number of matching m-point templates at the τ-scale; A τ (r) represents the number of matching m+1 point templates at the τ scale; r (similarity tolerance) is the dynamic matching threshold, calculated as similarity tolerance = 0.15 × σ; m (template dimension) is set to 2 to balance computational efficiency and accuracy;

[0031] S2023: Multi-scale entropy fusion Where τ 0.5 Weighting enhances large-scale contributions and improves robustness to electromyographic noise; As a scale normalization factor, it maintains the physiological range of entropy values; The goal is to achieve three-scale fusion and capture multi-level neurodynamics.

[0032] Preferably, the specific steps of S3 are as follows:

[0033] S301: State determination model: accurately identifies intention states, specifically as follows:

[0034] S3011: The state determination model determines whether a user has a control intention by quantifying the difference between real-time EEG characteristics and a personalized resting baseline. The formula used is: Where D t F is the intention difference index. t For real-time feature vectors, F1 is the energy difference ratio of the motion cortex (Δβ / α), F2 is the C3-C4 coherence, and F3 is the γ-band multi-scale entropy. rest For a personalized resting baseline, σF is the characteristic stability factor, wi For dynamic feature weights;

[0035] S3012: Dynamic weight allocation mechanism. Weight allocation is based on the reliability of features in the current environment, and the formula used is: in Where the separation degree is the difference between the mean features of the intended state and the resting state, σ noise For noise sensitivity, i represents the degree to which feature i is affected by noise, and the adjustment coefficient k = 0.3;

[0036] S3013: Dual judgment conditions, to avoid accidental triggering due to momentary interference, strict judgment logic is set: Trend monitoring requires a continuous increase in the degree of difference and exclusion of random fluctuations; time confirmation requires that the standard be met for 3 consecutive frames (300ms).

[0037] S302: Dynamic threshold algorithm: intelligently balances sensitivity, and the formula used is as follows:

[0038] Adaptive threshold formula: Where τ0 is the basic threshold, set by user calibration, with a value range of 0.8-1.2; Nerr is the false trigger count, calculated using a sliding window, with a value range ≥0; and σnoise is the real-time noise intensity, calculated through the Fz channel, with a value range of μV. 2 σbase is the reference noise, determined during initialization, with a value range of 5μV. 2 ;

[0039] Accidental trigger compensation: Compensation amount = 0.15·tanh(0.4·N) err For each false trigger detected, Nerr is incremented by 1; N_err decreases by 0.2 every 5 minutes, but is not less than 0.

[0040] Noise compensation items:

[0041] Basic threshold calibration: τ0 = 0.85 × ||F intent -F rest ||;

[0042] S303: Performs collaborative work between the state determination and dynamic threshold modules.

[0043] Preferably, the specific logical steps of S4 are as follows:

[0044] S401: When the tactile sensor detects three vibrations, the user should close their eyes and sit still.

[0045] S402: Perform multimodal monitoring, use EEG to collect brain signals, HRV monitoring and respiratory rate detection, calculate alpha wave power from the collected brain signals, analyze the high-frequency components of HRV from the HRV monitoring data, and calculate respiratory rate (RR) from the respiratory rate.

[0046] S403: The relaxation index RI is calculated based on the results of S402, using the following formula: Where Pα is the occipital alpha wave power (O1 / O2 channel); HRV is the high-frequency component of heart rate variability (HF-HRV); RR is the respiratory rate (breaths / minute);

[0047] S404: Judgment threshold: If RI>0.7, mark the frame as valid; otherwise, discard the data.

[0048] S405: Accumulated for 5 minutes, tactile alert with a 1-second long vibration;

[0049] S406: Establish F-rest baseline, monitor β events, detect β energy drop >20%, otherwise continue monitoring β events, and automatically capture 3 seconds of EEG while the user performs natural actions;

[0050] S407: Extract F-intent. If this is done three times, calculate the τ0 threshold; otherwise, continue monitoring the β event.

[0051] Preferably, in step S406, when establishing the F-rest baseline, data from the time period with RI > 0.7 are selected, and the β / α ratio characteristics of the C3 / C4 motor cortex are extracted for robust calculation. The formula used is: F rest =TrimmedMean(F t ,10%.

[0052] Preferably, in step S407, the formula used to calculate the τ0 threshold is: τ0=k×MAD(F intent -F rest ), where k = 1.48: is the Gaussian distribution calibration coefficient;

[0053] After extracting the F-intent three times, the median of the features is taken using the following formula: F i ntent = Median(F1,F2,F3).

[0054] Compared with existing technologies, the beneficial effects of this invention are:

[0055] This invention acquires high-quality EEG signals through a signal acquisition and preprocessing module, and then uses a multimodal feature extraction module to fuse the frequency domain Δβ / α energy ratio, spatial domain C3-C4 coherence, and nonlinear γ-band multi-scale entropy to generate a robust feature vector. After inputting this vector into an asynchronous detection engine, a dynamic threshold algorithm based on coupled false trigger history and environmental noise is used to achieve precise awakening of the intended state. At the same time, a fatigue-resistant initialization process involving tactile guidance for closed-eye rest and multimodal physiological verification is completed within 5 minutes to complete personalized parameter calibration. Ultimately, the invention achieves medical-grade performance indicators with a false trigger rate of <0.5 times / hour and a user fatigue reduction of 60%+, meeting the usage requirements. Attached Figure Description

[0056] Figure 1 The flowchart shows the low-load, fatigue-resistant asynchronous brain-control interface switching method proposed in this invention.

[0057] Figure 2 This is a flowchart of the fatigue-resistant initialization process in the low-load, fatigue-resistant asynchronous brain-controlled interface switching method proposed in this invention. Detailed Implementation

[0058] The present invention will be further explained below with reference to specific embodiments.

[0059] Example

[0060] Reference Figure 1-2 This embodiment proposes a low-load, fatigue-resistant asynchronous brain-controlled interface switching method, including the following steps:

[0061] S1: Signal acquisition and preprocessing: Signal acquisition was performed using an 8-channel dry electrode EEG cap with 16-bit ADC resolution and 256Hz sampling rate. The acquired EEG signals were then subjected to three levels of preprocessing, followed by signal standardization, environmental noise monitoring, segmentation, and buffering of the preprocessed data.

[0062] The dry electrode EEG cap has an input impedance >100MΩ and a common-mode rejection ratio ≥110dB to suppress environmental interference, and the electrode layout strictly follows the international 10-20 system standard.

[0063] The specific steps for three-level preprocessing of the acquired EEG signals are as follows:

[0064] S101: An adaptive 50Hz notch filter is used to eliminate power frequency interference from the power grid. The formula used is as follows: Where w0 = 2π × 50 / 256 is the digital angular frequency, and r = 0.95 is the bandwidth control factor;

[0065] S102: Bandpass filter, using a zero-phase IIR filter for frequency band separation, high-pass cutoff of 0.5Hz to remove baseline drift, low-pass cutoff of 45Hz to suppress electromyographic noise, and retaining five key frequency bands: δ 0.5-4Hz, θ 4-8Hz, α 8-13Hz, β 13-30Hz, and γ 30-45Hz.

[0066] S103: Artifact removal is performed using an improved independent component analysis, with the formula: X - AS + E, where X is the multi-channel observation signal matrix; A is the mixing matrix; S is the independent source signal; and E is the residual matrix.

[0067] Signal standardization requires two levels of standardization: reference electrode standardization and inter-channel energy equalization.

[0068] The formula used for its reference electrode standardization is: Where μTP9 is the mean value of the TP9 electrode within a 1-second sliding window; σTP10 is the standard deviation of the TP10 electrode within the same time window, and its function is to eliminate the systematic deviation caused by the change in the contact impedance of the reference electrode.

[0069] The formula used for energy balance between its channels is: Its function is to calculate the energy normalization factor for each channel i, so that the signals of each channel have comparable energy levels;

[0070] When monitoring environmental noise, a 0.5-10Hz bandpass filter is set in the Fz channel, and the signal variance within a 200ms time window is calculated using the following formula: And perform noise index classification, its low noise σ 2 <5μV 2 ; Medium noise 5≤σ 2 ≤15μV 2 ; High noise σ 2 >15μV 2 ;

[0071] To balance real-time performance and accuracy during segmentation and caching, an overlapping framing strategy is adopted, with a frame length of 200ms and a frame shift of 100ms. Meanwhile, the circular buffer retains the data of the most recent 10 frames.

[0072] S2: Multimodal feature extraction: Extract three types of complementary features to form the intended feature vector F, whose complementary features include frequency domain features, spatial domain features and nonlinear features;

[0073] The frequency domain feature characterizes the energy difference ratio of the motor cortex. When a motor intention is generated, the beta waves in the contralateral brain region desynchronize and decrease in energy, while the beta waves in the ipsilateral brain region synchronize and increase in energy. The energy ratio between the left and right brain regions is used to offset the individual's absolute amplitude difference. The formula used is as follows: Where E β (C3) represents the energy of the β-band of channel C3, calculated as the sum of squares after wavelet packet decomposition from 13-30Hz; E α (C4) represents the energy of the α band of channel C4, calculated as the variance after 8-13Hz bandpass filtering; σ rest σ represents the resting-state standard deviation and the initial stage statistic. t The standard deviation of the real-time signal is calculated using a 200ms sliding window.

[0074] Spatial features are used to reflect brain region coordination and calculate motor cortical coherence. Motor cortical coherence refers to the degree of synchronization of neural oscillations in the C3 and C4 electrode regions of the left and right hemisphere motor cortexes. The physiological mechanism is that in the resting state, the left and right motor cortices coordinate highly through the corpus callosum, with coherence > 0.7; when intention is generated, the neural resources of the target brain region are locally concentrated, and the coherence decreases to 0.3-0.5. The specific calculation process is as follows:

[0075] S2011: Perform a short-time Fourier transform on the C3 / C4 channel signals and calculate the complex spectrum for each frame of the signal. The formula used is as follows:

[0076]

[0077] Where f represents the frequency point within the range of 18-22Hz; t represents the time frame index.

[0078] S2012: Calculate the spectral cross-correlation, which characterizes the phase relationship between two signals at frequency f. The formula used is: Where * denotes complex conjugate operation;

[0079] S2013: Coherence calculation employs a complex coherence algorithm to avoid spurious coherence caused by volume conduction. The formula is as follows: Where K is the average number of frames, with a default K=5, corresponding to 500ms of data;

[0080] S2014: Feature value extraction, calculation of the target frequency band mean, and dynamic normalization are performed. The formula used to calculate the target frequency band mean is as follows: The formula used for dynamic normalization is: Where μ rest σ represents the mean of the user's resting-state coherence. rest This represents the resting-state standard deviation.

[0081] S2015: Anti-interference design is implemented, limiting the frequency band to a narrow 18-22Hz band, avoiding the main frequency band of electromyographic noise >25Hz, and avoiding the θ / α frequency band, which is susceptible to electrooculography interference;

[0082] The nonlinear characteristic is the γ-band multi-scale entropy, used to capture complexity. When the human brain generates a motor intention, the firing pattern of the motor cortex neural clusters changes from a resting-state rhythmic oscillation to a non-steady-state chaotic mode. This change is particularly significant in the γ-band. The specific calculation process is as follows:

[0083] S2021: Multi-scale reconstruction, decomposing the signal into sequences with different time resolutions: Where τ is the scale factor (1,2,3), used for adjusting the temporal resolution; j is the index of the new sequence; The signal is reconstructed at the τ scale to characterize neural activity at different time windows;

[0084] S2022: Scale entropy calculation, calculating improved sample entropy for each scale sequence: sampEn Among them B τ (r) represents the number of matching m-point templates at the τ-scale; A τ (r) represents the number of matching m+1 point templates at the τ scale; r (similarity tolerance) is the dynamic matching threshold, calculated as similarity tolerance = 0.15 × σ; m (template dimension) is set to 2 to balance computational efficiency and accuracy;

[0085] S2023: Multi-scale entropy fusion Where τ 0.5 Weighting enhances large-scale contributions and improves robustness to electromyographic noise; As a scale normalization factor, it maintains the physiological range of entropy values; The goal is to achieve three-scale fusion and capture multi-level neurodynamics;

[0086] S3: Asynchronous Detection: Achieves zero-load in resting state and fast response in intent state through asynchronous detection engine, with detection latency ≤800ms. It adopts dynamic dual threshold mechanism and environmental adaptive compensation, maintaining a 94% intent detection rate while the false trigger rate is <0.5 times / hour under 65dB noise.

[0087] The specific steps are as follows:

[0088] S301: State determination model: accurately identifies intention states, specifically as follows:

[0089] S3011: The state determination model determines whether a user has an intention to control by quantifying the difference between real-time EEG characteristics and a personalized resting baseline. The formula used is: D t = Where D t F is the intention difference index. t For real-time feature vectors, F1 is the energy difference ratio of the motion cortex (Δβ / α), F2 is the C3-C4 coherence, and F3 is the γ-band multi-scale entropy.rest For a personalized resting baseline, σF is the characteristic stability factor, w i For dynamic feature weights;

[0090] S3012: Dynamic weight allocation mechanism. Weight allocation is based on the reliability of features in the current environment, and the formula used is: in Where the separation degree is the difference between the mean features of the intended state and the resting state, σ noise For noise sensitivity, i represents the degree to which feature i is affected by noise, and the adjustment coefficient k = 0.3;

[0091] S3013: Dual judgment conditions, to avoid accidental triggering due to momentary interference, strict judgment logic is set: Trend monitoring requires a continuous increase in the degree of difference and exclusion of random fluctuations; time confirmation requires that the standard be met for 3 consecutive frames (300ms).

[0092] S302: Dynamic threshold algorithm: intelligently balances sensitivity, and the formula used is as follows:

[0093] Adaptive threshold formula: Where τ0 is the basic threshold, set by user calibration, with a value range of 0.8-1.2; Nerr is the false trigger count, calculated using a sliding window, with a value range ≥0; and σnoise is the real-time noise intensity, calculated through the Fz channel, with a value range of μV. 2 σbase is the reference noise, determined during initialization, with a value range of 5μV. 2 ;

[0094] Accidental trigger compensation: Compensation amount = 0.15·tanh(0.4·N) err For each false trigger detected, Nerr is incremented by 1; N_err decreases by 0.2 every 5 minutes, but is not less than 0.

[0095] Noise compensation items:

[0096] Basic threshold calibration: τ0 = 0.85 × ||F intent -F rest ||;

[0097] S303: Performs collaborative operation of the state determination and dynamic threshold modules;

[0098] S4: Anti-fatigue initialization: Achieve physiological state calibration to adapt to individual differences through anti-fatigue initialization;

[0099] The specific logical steps are as follows:

[0100] S401: When the tactile sensor detects three vibrations, the user should close their eyes and sit still.

[0101] S402: Perform multimodal monitoring, use EEG to collect brain signals, HRV monitoring and respiratory rate detection, calculate alpha wave power from the collected brain signals, analyze the high-frequency components of HRV from the HRV monitoring data, and calculate respiratory rate (RR) from the respiratory rate.

[0102] S403: The relaxation index RI is calculated based on the results of S402, using the following formula: Where Pα is the occipital alpha wave power (O1 / O2 channel); HRV is the high-frequency component of heart rate variability (HF-HRV); RR is the respiratory rate (breaths / minute);

[0103] S404: Judgment threshold: If RI>0.7, mark the frame as valid; otherwise, discard the data.

[0104] S405: Accumulated for 5 minutes, tactile alert with a 1-second long vibration;

[0105] S406: Establish the F-rest baseline, monitor β events, and detect a β energy drop >20%; otherwise, continue monitoring β events. If a β event occurs, automatically capture a 3-second EEG while the user performs a natural action. When establishing the F-rest baseline, select data from periods with RI > 0.7, extract the β / α ratio features of the C3 / C4 motor cortex, and perform robust calculations using the following formula: F rest =TrimmedMean(F t ,10%);

[0106] S407: Extract F-intent, complete three times, then calculate the τ0 threshold; otherwise, continue monitoring β events. The formula used to calculate the τ0 threshold is: τ0=k×MAD(F intent -F rest ), where k = 1.48: is the Gaussian distribution calibration coefficient;

[0107] After extracting the F-intent three times, the median of the features is taken using the following formula: F i ntent = Median(F1,F2,F3).

[0108] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A low-load, fatigue-resistant asynchronous brain-controlled interface switching method, characterized in that, Includes the following steps: S1: Signal acquisition and preprocessing: Signal acquisition was performed using an 8-channel dry electrode EEG cap with 16-bit ADC resolution and 256Hz sampling rate. The acquired EEG signals were then subjected to three levels of preprocessing, followed by signal standardization, environmental noise monitoring, segmentation, and buffering of the preprocessed data. S2: Multimodal feature extraction: Extract three types of complementary features to form the intended feature vector F, whose complementary features include frequency domain features, spatial domain features and nonlinear features; S3: Asynchronous Detection: Achieves zero-load in resting state and fast response in intent state through asynchronous detection engine, with detection latency ≤800ms. It adopts dynamic dual threshold mechanism and environmental adaptive compensation, maintaining a 94% intent detection rate while the false trigger rate is <0.5 times / hour under 65dB noise. S4: Anti-fatigue initialization: Achieve physiological state calibration to adapt to individual differences through anti-fatigue initialization.

2. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 1, characterized in that, In S1, the input impedance of the dry electrode EEG cap is >100MΩ and the common-mode rejection ratio is ≥110dB, which is used to suppress environmental interference, and the electrode layout strictly follows the international 10-20 system standard. The specific steps for three-level preprocessing of the acquired EEG signals are as follows: S101: An adaptive 50Hz notch filter is used to eliminate power frequency interference from the power grid. The formula used is as follows: Where w0 = 2π × 50 / 256 is the digital angular frequency, and r = 0.95 is the bandwidth control factor; S102: Bandpass filter, using a zero-phase IIR filter for frequency band separation, high-pass cutoff of 0.5Hz to remove baseline drift, low-pass cutoff of 45Hz to suppress electromyographic noise, and retaining five key frequency bands: δ 0.5-4Hz, θ 4-8Hz, α 8-13Hz, β 13-30Hz, and γ 30-45Hz. S103: Artifact removal is performed using an improved independent component analysis formula: X - AS + E, where X is the multi-channel observation signal matrix; A is the mixing matrix; S is the independent source signal; and E is the residual matrix.

3. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 1, characterized in that, In S1, two levels of standardization are required when performing signal standardization: reference electrode standardization and inter-channel energy equalization. The formula used for its reference electrode standardization is: Where μTP9 is the mean value of the TP9 electrode within a 1-second sliding window; σTP10 is the standard deviation of the TP10 electrode within the same time window, and its function is to eliminate the systematic deviation caused by the change in the contact impedance of the reference electrode. The formula used for energy balance between its channels is: Its function is to calculate the energy normalization factor for each channel i, so that the signals of each channel have comparable energy levels; When monitoring environmental noise, a 0.5-10Hz bandpass filter is set in the Fz channel, and the signal variance within a 200ms time window is calculated using the following formula: And perform noise index classification, its low noise σ 2 <5μV 2 ; Medium noise 5≤σ 2 ≤15μV 2 ; High noise σ 2 >15μV 2 ; To balance real-time performance and accuracy during segmentation and caching, an overlapping framing strategy is adopted, with a frame length of 200ms and a frame shift of 100ms. Meanwhile, the circular buffer retains the data of the most recent 10 frames.

4. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 1, characterized in that, In S2, the frequency domain feature characterizes the energy difference ratio of the motor cortex. When a motor intention is generated, the beta wave in the contralateral brain region desynchronizes and its energy decreases, while the beta wave in the ipsilateral brain region synchronizes and its energy increases. The energy ratio of the left and right brain regions is used to offset the individual's absolute amplitude difference. The formula used is as follows: Where E β (C3) represents the energy of the β-band of channel C3, calculated as the sum of squares after wavelet packet decomposition from 13-30Hz; E α (C4) represents the energy of the α band of channel C4, calculated as the variance after 8-13Hz bandpass filtering; σ rest σ represents the resting-state standard deviation and the initial stage statistic. t The standard deviation of the real-time signal is calculated using a 200ms sliding window. Spatial features are used to reflect brain region coordination and calculate motor cortical coherence. Motor cortical coherence refers to the degree of synchronization of neural oscillations in the C3 and C4 electrode regions of the left and right hemisphere motor cortexes. The physiological mechanism is as follows: in the resting state, the left and right motor cortices coordinate highly through the corpus callosum, with coherence > 0.7; when intention is generated, the neural resources of the target brain region are locally concentrated, and the coherence decreases to 0.3-0.

5. The specific calculation process is as follows: S2011: Perform a short-time Fourier transform on the C3 / C4 channel signals and calculate the complex spectrum for each frame of the signal. The formula used is as follows: Where f represents the frequency point within the range of 18-22Hz; t represents the time frame index. S2012: Calculate the spectral cross-correlation, which characterizes the phase relationship between two signals at frequency f. The formula used is: Where * denotes complex conjugate operation; S2013: Coherence calculation employs a complex coherence algorithm to avoid spurious coherence caused by volume conduction. The formula is as follows: Where K is the average number of frames, with a default K=5, corresponding to 500ms of data; S2014: Feature value extraction, calculation of the target frequency band mean, and dynamic normalization are performed. The formula used to calculate the target frequency band mean is as follows: The formula used for dynamic normalization is: Where μ rest σ represents the mean of the user's resting-state coherence. rest This represents the resting-state standard deviation. S2015: Anti-interference design is implemented, limiting the frequency band to a narrow 18-22Hz band, avoiding the main frequency band of electromyographic noise >25Hz, and avoiding the θ / α frequency band, which is susceptible to electrooculography interference; The nonlinear characteristic is the γ-band multi-scale entropy, used to capture complexity. When the human brain generates a motor intention, the firing pattern of the motor cortex neural clusters changes from a resting-state rhythmic oscillation to a non-steady-state chaotic mode. This change is particularly significant in the γ-band. The specific calculation process is as follows: S2021: Multi-scale reconstruction, decomposing the signal into sequences with different time resolutions: Where τ is the scale factor (1,2,3), used for adjusting the temporal resolution; j is the index of the new sequence; The signal is reconstructed at the τ scale to characterize neural activity at different time windows; S2022: Scale entropy calculation, calculating improved sample entropy for each scale sequence: Among them B τ (r) represents the number of matching m-point templates at the τ-scale; A τ (r) represents the number of matching m+1 point templates at the τ scale; r is the dynamic matching threshold, calculated as similarity tolerance = 0.15 × σ; m is taken as 2 to balance computational efficiency and accuracy; S2023: Multi-scale entropy fusion Where τ 0.5 Weighting enhances large-scale contributions and improves robustness to electromyographic noise; As a scale normalization factor, it maintains the physiological range of entropy values; The goal is to achieve three-scale fusion and capture multi-level neurodynamics.

5. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 1, characterized in that, The specific steps of S3 are as follows: S301: State determination model: accurately identifies intention states, specifically as follows: S3011: The state determination model determines whether a user has a control intention by quantifying the difference between real-time EEG characteristics and a personalized resting baseline. The formula used is: Where D t F is the intention difference index. t For real-time feature vectors, F1 is the energy difference ratio of the motion cortex (Δβ / α), F2 is the C3-C4 coherence, and F3 is the γ-band multi-scale entropy. rest For a personalized resting baseline, σF is the characteristic stability factor, w i For dynamic feature weights; S3012: Dynamic weight allocation mechanism. Weight allocation is based on the reliability of features in the current environment, and the formula used is: in Where the separation degree is the difference between the mean features of the intended state and the resting state, σ noise For noise sensitivity, i represents the degree to which feature i is affected by noise, and the adjustment coefficient k = 0.3; S3013: Dual judgment conditions, to avoid accidental triggering due to momentary interference, strict judgment logic is set: Trend monitoring requires a continuous increase in the degree of difference and exclusion of random fluctuations; time confirmation requires that the standard be met for 3 consecutive frames (300ms). S302: Dynamic threshold algorithm: intelligently balances sensitivity, and the formula used is as follows: Adaptive threshold formula: Where τ0 is the basic threshold, set by user calibration, with a value range of 0.8-1.2; Nerr is the false trigger count, calculated using a sliding window, with a value range ≥0; and σnoise is the real-time noise intensity, calculated through the Fz channel, with a value range of μV. 2 σbase is the reference noise, determined during initialization, with a value range of 5μV. 2 ; Accidental trigger compensation: Compensation amount = 0.15·tanh(0.4·N) err For each false trigger detected, Nerr is incremented by 1; N_err decreases by 0.2 every 5 minutes, but is not less than 0. Noise compensation items: Basic threshold calibration: τ0 = 0.85 × ||F intent -F rest ||; S303: Performs collaborative work between the state determination and dynamic threshold modules.

6. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 1, characterized in that, The specific logical steps of S4 are as follows: S401: When the tactile sensor detects three vibrations, the user should close their eyes and sit still. S402: Perform multimodal monitoring, use EEG to collect brain signals, HRV monitoring and respiratory rate detection, calculate alpha wave power from the collected brain signals, analyze the high-frequency components of HRV from the HRV monitoring data, and calculate respiratory rate (RR) from the respiratory rate. S403: The relaxation index RI is calculated based on the results of S402, using the following formula: Where Pα is the occipital α wave power; HRV is the high-frequency component of heart rate variability; RR is the respiratory rate; S404: Judgment threshold: If RI>0.7, mark the frame as valid; otherwise, discard the data. S405: Accumulated for 5 minutes, tactile alert with a 1-second long vibration; S406: Establish F-rest baseline, monitor β events, detect β energy drop >20%, otherwise continue monitoring β events, and automatically capture 3 seconds of EEG while the user performs natural actions; S407: Extract F-intent. If this is done three times, calculate the τ0 threshold; otherwise, continue monitoring the β event.

7. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 6, characterized in that, In S406, when establishing the F-rest baseline, data from the time period with RI > 0.7 are selected, and the β / α ratio characteristics of the C3 / C4 motor cortex are extracted for robust calculation. The formula used is: F rest =TrimmedMean(F t ,10%.

8. The low-load, fatigue-resistant asynchronous brain-controlled interface switching method according to claim 6, characterized in that, In S407, the formula used to calculate the τ0 threshold is: τ0=k×MAD(F intent -F rest ), where k = 1.48: is the Gaussian distribution calibration coefficient; After extracting the F-intent three times, the median of the features is taken using the following formula: F i ntent = Median(F1,F2,F3).

Images