Exoskeleton training cortex excitability fatigue phase compensation method and system

Abstract

The application discloses an exoskeleton training cortex excitability fatigue phase compensation method and system, belongs to the cross technical field of brain-computer interface and rehabilitation robot, is used for phase lag compensation, comprises the following steps: processing collected multi-modal brain signals into signals after artifact rejection, combining double-wavelength near-infrared light intensity signals to construct multi-modal feature vectors; constructing a heterogeneous graph based on the multi-modal feature vectors and generating a graph-level global embedding; outputting a phase lag index based on the graph-level global embedding; constructing a decoding acceptance gate and cumulative fatigue using the phase lag index; constructing a secondary confirmation gate; taking the transcranial magnetic stimulation pulse moment as the artifact rejection processing input, updating the stimulation intensity and limiting the amplitude. The application converts the fatigue phase into a real-time gating variable, drives the threshold value, the beat, TMS and the exoskeleton motion, optimizes the control in a closed loop, maintains the decoding reliability and prevents false assistance.

Description

Technical Field

[0001] This invention discloses a method and system for phase compensation of cortical excitability fatigue during exoskeleton training, belonging to the interdisciplinary field of brain-computer interface and rehabilitation robot technology. Background Technology

[0002] Motor imagery brain-computer interface enables active rehabilitation training of "brain-controlled legs" by decoding the ERD features of the mu / beta rhythm of the motor cortex. Combined with transcranial magnetic stimulation neuromodulation and functional near-infrared spectroscopy brain metabolism monitoring, it can theoretically form a complete neurorehabilitation closed loop.

[0003] However, during long-term training, neural fatigue leads to shallower ERD modulation depth and decreased decoding confidence. Existing solutions have fundamental flaws: using fixed thresholds cannot adapt to confidence drift caused by fatigue; fatigue is only analyzed offline and not transformed into real-time control variables; single EEG cannot detect metabolic fatigue, and fNIRS has insufficient temporal resolution; TMS cannot be closed-loop online due to electromagnetic artifacts and can only be open-loop blind stimulation.

[0004] The most serious consequence is that when the decoder misinterprets false or weak intentions as real commands in a fatigued state, the exoskeleton still performs significant assistance according to a fixed threshold. This not only weakens the training effect but may also cause injury to the patient due to incorrect movement direction or joint impact. The existing system completely lacks a safe execution judgment mechanism based on fatigue reliability. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for phase compensation of cortical excitability fatigue in exoskeleton training, in order to solve the problems in the prior art. In long-term motor imagery exoskeleton training, nerve fatigue causes a phase lag in cortical excitability relative to the fatigue rhythm. Existing methods cannot distinguish between the loss of true intention and the decline in intention caused by fatigue, resulting in a continuous decline in decoding performance. Furthermore, the exoskeleton may still perform erroneous large movements when the patient's intention weakens, causing training failure and safety hazards.

[0006] Methods for phase compensation of cortical excitability fatigue in exoskeleton training include: S1. Collect multimodal brain signals and extract the bandpass signal. Use the transcranial magnetic stimulation pulse time as the anchor point to perform artifact suppression processing to obtain the artifact-suppressed signal. Construct a cospatial pattern feature vector based on the artifact-suppressed signal and extract the desynchronized sliding window mean scalar. Extract the changes in oxyhemoglobin concentration and deoxyhemoglobin concentration based on the dual-wavelength near-infrared light intensity signal. After time alignment, concatenate it with the cospatial pattern feature vector and the sliding window mean scalar to form a multimodal feature vector. S2. Assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct node feature matrices. Construct a weighted adjacency matrix using spatial distance and functional connectivity strength to obtain a heterogeneous graph. Utilize multi-head graph attention and attention pooling to obtain graph-level global embedding. Based on graph-level global embedding, output the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. Obtain the phase lag index by weighting the phase difference between the inner circle of the sliding window and phase-locked consistency. Quantize and discount the phase lag index to obtain the confidence after fatigue modulation. S3. Calculate the decoding acceptance gate using the phase lag exponent linearly modulated intent acceptance threshold. Calculate the cumulative fatigue load using the phase lag exponent and adaptively adjust the trial interval. Insert a rest when the fatigue load reaches the micro-rest threshold. Trial a transcranial magnetic stimulation (TMS) pulse when the fatigue load reaches the TMS intervention threshold and the excitatory instantaneous phase is in the optimal phase range. Construct a secondary confirmation gate based on the decoding acceptance gate, the phase lag exponent being below the confirmation threshold, and the intent category remaining unchanged, to obtain the action amplitude gain. Use the TMS pulse timing as the artifact suppression input, compare the mean difference of the phase lag exponent before and after TMS to evaluate the intervention effect, and update the stimulation intensity and limit the amplitude based on the mean difference.

[0007] S1 includes, S1.1, acquiring multi-channel scalp EEG signals at a sampling rate of... collection The scalp potential of each electrode is used to construct a discrete primitive matrix. ;use The Butterworth bandpass filter extracts the bandpass signal from the scalp potential. , Electrode index; S1 includes S1.2, and the artifact suppression processing includes a three-level cascaded process of constructing a blinding window definition, autoregressive reconstruction, and exponential decay artifact subtraction. The blinding window definition includes the timing of synchronous output pulses using a stimulator. Defining the contamination window for each pulse : ; In the formula, For the first Timing of each transcranial magnetic stimulation pulse This is the pre-pulse blinding margin. This is the blinding margin after the pulse. Index of transcranial magnetic stimulation pulse times; Autoregressive reconstruction includes estimating the pre-pulse clean segment. AR model coefficients, forward prediction to reconstruct signals within the blinded window: ; In the formula, For the first The estimated value of the reconstructed signal of the channel within the blinding window. For AR model coefficients, For order, For discrete-time indexing, Delayed indexing for AR models ; Exponential decay artifact reduction includes modeling and subtracting the residual decay artifact after the pulse: ; ; In the formula, For the first Fitting model for residual artifacts after channel pulses. For the first Channel 1 The magnitude of each exponential component, For the first Channel 1 The time constant of each exponential component Number of components For component index, for The estimated value, For the first The signal after artifact suppression in the channel.

[0008] S1 includes, S1.3, and will Apply the cospace mode projection matrix and select the one with the largest variance after projection. The component with the smallest variance Each component is used to construct spatial filtering features. And construct the common space pattern feature vector , , It is the transpose symbol. For the projection component index, The logarithm of the common space mode components; based on Calculate the instantaneous frequency band power of the sliding window, and normalize it with the resting baseline reference power to obtain the time. Desynchronization percentage and extract sliding window mean scalar To modify Beer-Lambert's law, the changes in oxyhemoglobin concentration were calculated from dual-wavelength near-infrared light intensity signals. Changes in deoxyhemoglobin concentration .

[0009] S1 includes, S1.4, and will and Combined into a scalar blood oxygen load index : ; In the formula, This represents the peak amplitude of the response of oxygenated hemoglobin within the current sliding window to motion imagery. To train the response amplitude during the initial task period, , For two weights, To prevent small constants with a denominator of zero; Based on functional near-infrared spectral sampling rate and hemodynamic delay Constructing aligned blood oxygen load : ; Will and and Align them on the timeline and concatenate them into a multimodal feature vector. : ; In the formula, For a moment of , For a moment of .

[0010] S2 includes, S2.1, and A node feature matrix was constructed by assigning nodes to EEG nodes and functional near-infrared spectroscopy nodes. A weighted adjacency matrix was then constructed based on spatial distance and functional connectivity strength to obtain a heterogeneous graph. The EEG nodes include... and Functional near-infrared spectral nodes include ; For each node in the heterogeneous graph, calculate the multi-head attention coefficient in its neighborhood and update the node embedding; Attention pooling is used to aggregate node embeddings into graph-level global embeddings; The softmax classification head outputs the motion imagination intent category from the graph-level global embedding. With confidence probability; The regression head is used to embed the cortical excitability score from the graph-level global embedding, and the excitability instantaneous phase is extracted by slow-scale filtering and Hilbert transform. ; The neural fatigue score is output by the regression head, and after near-infrared spectral sampling rate compensation using a cubic spline upsampling function, the instantaneous phase of fatigue is extracted. .

[0011] S2 includes S2.2, calculating the phase difference between the excitatory phase and the fatigue phase using circular statistical methods. : ; In the formula, For phase winding function; Calculate the mean phase difference of the inner circle of the sliding window Consistency with phase-locked loop : ; ; In the formula, For the summation index within the sliding window, For sliding statistics windows, To reduce phase difference Represented as a unit complex number, The imaginary unit, To determine the mold length; The normalized phase lag exponent is obtained by phase-locked loop consistency weighting. : ; by Quantify the probability of discount intent and calculate the confidence level after fatigue modulation: ; In the formula, For a moment Intention confidence after fatigue modulation For the probability of intent, This is the modulation intensity coefficient.

[0012] S3 includes, S3.1, an intentional acceptance threshold using exponential linear modulation with phase lag. Calculate the decoding reception gate : ; ; In the formula, The baseline decision threshold is the threshold under fatigue-free conditions. The modulation slope of the threshold due to phase lag. For indicator functions; Calculating cumulative fatigue load using phase hysteresis index : ; In the formula, Forgetting factor; The adaptive trial interval is: ; In the formula, For a moment The adaptive trial interval, The baseline trial interval, This is the beat slowdown factor; Set micro-rest trigger threshold ,like Insert rest block The system pauses during the rest block.

[0013] S3 includes, S3.2, when Reaching the threshold for transcranial magnetic stimulation intervention and Upon entering the optimal phase interval, a transcranial magnetic stimulation pulse is delivered at the predicted optimal phase time, triggered by the following conditions: ; In the formula, For the first Triggering markers for transcranial magnetic stimulation The threshold for transcranial magnetic stimulation intervention. For the pre-calibrated optimal stimulus phase, For phase tolerance window, For logical AND operator; Phase prediction pulse timing is as follows: ; In the formula, For search indexing, Index for the current time, To obtain the minimum value.

[0014] S3 includes, S3.3, constructing a secondary confirmation gate: ; In the formula, For a moment Exoskeleton execution gate, For low fatigue phase conditions, For the purpose of timing stability verification, The phase lag confirmation threshold, For stability verification window length, For the index within the verification window; Motion amplitude gain for: ; In the formula, The base coefficient for low-amplitude gain; S3 includes S3.4, selecting the desired motion trajectory based on the intent category and uniformly scaling it using amplitude gain: ; In the formula, For a moment The assist torque, For impedance-controlled proportional gain, For the purpose of class Selected desired joint angle, This provides feedback on the actual angle of the joint. For impedance-controlled differential gain, For the purpose of class Selected desired joint angular velocity, Feedback on the actual angular velocity of the joint; S3 includes S3.5, and the difference in mean PLI values ​​before and after stimulation. : ; In the formula, For the first Short window after the second stimulation The mean, For the first Within the short window before the second stimulation mean Perform closed-loop updates of stimulus intensity: ; In the formula, For the updated stimulus intensity, For the amplitude limiting function, For the first The intensity of the secondary pulse For dose learning rate, This is the safe lower limit of stimulus intensity. This represents the safe upper limit for the intensity of the stimulus.

[0015] The exoskeleton training cortical excitability fatigue phase compensation system includes: The signal acquisition module is used to acquire multimodal brain signals; The artifact suppression module is used to perform artifact suppression processing on the acquired multimodal brain signals with the transcranial magnetic stimulation pulse time as the anchor point, and output the artifact-suppressed signal. The feature extraction module is used to construct a co-space mode feature vector from the artifact suppression signal and extract the desynchronized sliding window mean scalar, as well as extract the changes in oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration from the dual-wavelength near-infrared light intensity signal, and splice them into a multimodal feature vector after time alignment. The graph processing module is used to assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct heterogeneous graphs. Graph-level global embedding is obtained through graph attention aggregation. Based on the graph-level global embedding, the module outputs the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. The phase lag index is obtained by weighting the phase difference of the inner circle of the sliding window with phase-locked consistency. The discounted confidence probability is quantified by the phase lag index. The control response module is used to modulate the decoding threshold with phase lag exponential modulation, adapt the trial interval and insert rest, trigger transcranial magnetic stimulation pulses, and construct a secondary confirmation gate to obtain the motion amplitude gain. The closed-loop update module uses the transcranial magnetic stimulation pulse timing as input to the artifact suppression module, compares the mean difference of the phase lag index before and after transcranial magnetic stimulation to assess the intervention effect, and updates the stimulation intensity.

[0016] Compared with existing technologies, this invention has the following advantages: By converting neural fatigue into a real-time gating signal through the phase lag index, it realizes online quantification and dynamic tracking of fatigue decompensation, fundamentally solving the deficiency of existing solutions where fatigue is only analyzed post-eventally and cannot be intervened in real-time control; by using this index to uniformly drive decoding threshold adaptation, training rhythm regulation, transcranial magnetic stimulation phase-locking triggering, and exoskeleton secondary confirmation gating, it ensures that significant assistance is only performed when the intention is credible and fatigue is not decompensated, effectively preventing safety risks caused by erroneous assistance; at the same time, it constructs a closed-loop verification mechanism for stimulation effect and online dose update mechanism, and enables independent fatigue compensation through rest scheduling even without transcranial magnetic stimulation configuration, taking into account decoding reliability, patient safety, and clinical deployment flexibility during long-term training. Attached Figure Description

[0017] Figure 1 This is a flowchart of the technology of this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Methods for phase compensation of cortical excitability fatigue in exoskeleton training include: S1. Collect multimodal brain signals and extract the bandpass signal. Use the transcranial magnetic stimulation pulse time as the anchor point to perform artifact suppression processing to obtain the artifact-suppressed signal. Construct a cospatial pattern feature vector based on the artifact-suppressed signal and extract the desynchronized sliding window mean scalar. Extract the changes in oxyhemoglobin concentration and deoxyhemoglobin concentration based on the dual-wavelength near-infrared light intensity signal. After time alignment, concatenate it with the cospatial pattern feature vector and the sliding window mean scalar to form a multimodal feature vector. S2. Assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct node feature matrices. Construct a weighted adjacency matrix using spatial distance and functional connectivity strength to obtain a heterogeneous graph. Utilize multi-head graph attention and attention pooling to obtain graph-level global embedding. Based on graph-level global embedding, output the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. Obtain the phase lag index by weighting the phase difference between the inner circle of the sliding window and phase-locked consistency. Quantize and discount the phase lag index to obtain the confidence after fatigue modulation. S3. Calculate the decoding acceptance gate using the phase lag exponent linearly modulated intent acceptance threshold. Calculate the cumulative fatigue load using the phase lag exponent and adaptively adjust the trial interval. Insert a rest when the fatigue load reaches the micro-rest threshold. Trial a transcranial magnetic stimulation (TMS) pulse when the fatigue load reaches the TMS intervention threshold and the excitatory instantaneous phase is in the optimal phase range. Construct a secondary confirmation gate based on the decoding acceptance gate, the phase lag exponent being below the confirmation threshold, and the intent category remaining unchanged, to obtain the action amplitude gain. Use the TMS pulse timing as the artifact suppression input, compare the mean difference of the phase lag exponent before and after TMS to evaluate the intervention effect, and update the stimulation intensity and limit the amplitude based on the mean difference.

[0020] S1 includes, S1.1, acquiring multi-channel scalp EEG signals at a sampling rate of... collection The scalp potential of each electrode is used to construct a discrete primitive matrix. ;use The Butterworth bandpass filter extracts the bandpass signal from the scalp potential. , Electrode index; S1 includes S1.2, and the artifact suppression processing includes a three-level cascaded process of constructing a blinding window definition, autoregressive reconstruction, and exponential decay artifact subtraction. The blinding window definition includes the timing of synchronous output pulses using a stimulator. Defining the contamination window for each pulse : ; In the formula, For the first Timing of each transcranial magnetic stimulation pulse This is the pre-pulse blinding margin. This is the blinding margin after the pulse. Index of transcranial magnetic stimulation pulse times; Autoregressive reconstruction includes estimating the pre-pulse clean segment. AR model coefficients, forward prediction to reconstruct signals within the blinded window: ; In the formula, For the first The estimated value of the reconstructed signal of the channel within the blinding window. For AR model coefficients, For order, For discrete-time indexing, Delayed indexing for AR models ; Exponential decay artifact reduction includes modeling and subtracting the residual decay artifact after the pulse: ; ; In the formula, For the first Fitting model for residual artifacts after channel pulses. For the first Channel 1 The magnitude of each exponential component, For the first Channel 1 The time constant of each exponential component Number of components For component index, for The estimated value, For the first The signal after artifact suppression in the channel.

[0021] S1 includes, S1.3, and will Apply the cospace mode projection matrix and select the one with the largest variance after projection. The component with the smallest variance Each component is used to construct spatial filtering features. And construct the common space pattern feature vector , , It is the transpose symbol. For the projection component index, The logarithm of the common space mode components; based on Calculate the instantaneous frequency band power of the sliding window, and normalize it with the resting baseline reference power to obtain the time. Desynchronization percentage and extract sliding window mean scalar To modify Beer-Lambert's law, the changes in oxyhemoglobin concentration were calculated from dual-wavelength near-infrared light intensity signals. Changes in deoxyhemoglobin concentration .

[0022] S1 includes, S1.4, and will and Combined into a scalar blood oxygen load index : ; In the formula, This represents the peak amplitude of the response of oxygenated hemoglobin within the current sliding window to motion imagery. To train the response amplitude during the initial task period, , For two weights, To prevent small constants with a denominator of zero; Based on functional near-infrared spectral sampling rate and hemodynamic delay Constructing aligned blood oxygen load : ; Will and and Align them on the timeline and concatenate them into a multimodal feature vector. : ; In the formula, For a moment of , For a moment of .

[0023] S2 includes, S2.1, and A node feature matrix was constructed by assigning nodes to EEG nodes and functional near-infrared spectroscopy nodes. A weighted adjacency matrix was then constructed based on spatial distance and functional connectivity strength to obtain a heterogeneous graph. The EEG nodes include... and Functional near-infrared spectral nodes include ; For each node in the heterogeneous graph, calculate the multi-head attention coefficient in its neighborhood and update the node embedding; Attention pooling is used to aggregate node embeddings into graph-level global embeddings; The softmax classification head outputs the motion imagination intent category from the graph-level global embedding. With confidence probability; The regression head is used to embed the cortical excitability score from the graph-level global embedding, and the excitability instantaneous phase is extracted by slow-scale filtering and Hilbert transform. ; The neural fatigue score is output by the regression head, and after near-infrared spectral sampling rate compensation using a cubic spline upsampling function, the instantaneous phase of fatigue is extracted. .

[0024] S2 includes S2.2, calculating the phase difference between the excitatory phase and the fatigue phase using circular statistical methods. : ; In the formula, For phase winding function; Calculate the mean phase difference of the inner circle of the sliding window Consistency with phase-locked loop : ; ; In the formula, For the summation index within the sliding window, For sliding statistics windows, To reduce phase difference Represented as a unit complex number, The imaginary unit, To determine the mold length; The normalized phase lag exponent is obtained by phase-locked loop consistency weighting. : ; by Quantify the probability of discount intent and calculate the confidence level after fatigue modulation: ; In the formula, For a moment Intention confidence after fatigue modulation For the probability of intent, This is the modulation intensity coefficient.

[0025] S3 includes, S3.1, an intentional acceptance threshold using exponential linear modulation with phase lag. Calculate the decoding reception gate : ; ; In the formula, The baseline decision threshold is the threshold under fatigue-free conditions. The modulation slope of the threshold due to phase lag. For indicator functions; Calculating cumulative fatigue load using phase hysteresis index : ; In the formula, Forgetting factor; The adaptive trial interval is: ; In the formula, For a moment The adaptive trial interval, The baseline trial interval, This is the beat slowdown factor; Set micro-rest trigger threshold ,like Insert rest block The system pauses during the rest block.

[0026] S3 includes, S3.2, when Reaching the threshold for transcranial magnetic stimulation intervention and Upon entering the optimal phase interval, a transcranial magnetic stimulation pulse is delivered at the predicted optimal phase time, triggered by the following conditions: ; In the formula, For the first Triggering markers for transcranial magnetic stimulation The threshold for transcranial magnetic stimulation intervention. For the pre-calibrated optimal stimulus phase, For phase tolerance window, For logical AND operator; Phase prediction pulse timing is as follows: ; In the formula, For search indexing, Index for the current time, To obtain the minimum value.

[0027] S3 includes, S3.3, constructing a secondary confirmation gate: ; In the formula, For a moment Exoskeleton execution gate, For low fatigue phase conditions, For the purpose of timing stability verification, The phase lag confirmation threshold, For stability verification window length, For the index within the verification window; Motion amplitude gain for: ; In the formula, The base coefficient for low-amplitude gain; S3 includes S3.4, selecting the desired motion trajectory based on the intent category and uniformly scaling it using amplitude gain: ; In the formula, For a moment The assist torque, For impedance-controlled proportional gain, For the purpose of class Selected desired joint angle, This provides feedback on the actual angle of the joint. For impedance-controlled differential gain, For the purpose of class Selected desired joint angular velocity, Feedback on the actual angular velocity of the joint; S3 includes S3.5, and the difference in mean PLI values ​​before and after stimulation. : ; In the formula, For the first Short window after the second stimulation The mean, For the first Within the short window before the second stimulation mean Perform closed-loop updates of stimulus intensity: ; In the formula, For the updated stimulus intensity, For the amplitude limiting function, For the first The intensity of the secondary pulse For dose learning rate, This is the safe lower limit of stimulus intensity. This represents the safe upper limit for the intensity of the stimulus.

[0028] The exoskeleton training cortical excitability fatigue phase compensation system includes: The signal acquisition module is used to acquire multimodal brain signals; The artifact suppression module is used to perform artifact suppression processing on the acquired multimodal brain signals with the transcranial magnetic stimulation pulse time as the anchor point, and output the artifact-suppressed signal. The feature extraction module is used to construct a co-space mode feature vector from the artifact suppression signal and extract the desynchronized sliding window mean scalar, as well as extract the changes in oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration from the dual-wavelength near-infrared light intensity signal, and splice them into a multimodal feature vector after time alignment. The graph processing module is used to assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct heterogeneous graphs. Graph-level global embedding is obtained through graph attention aggregation. Based on the graph-level global embedding, the module outputs the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. The phase lag index is obtained by weighting the phase difference of the inner circle of the sliding window with phase-locked consistency. The discounted confidence probability is quantified by the phase lag index. The control response module is used to modulate the decoding threshold with phase lag exponential modulation, adapt the trial interval and insert rest, trigger transcranial magnetic stimulation pulses, and construct a secondary confirmation gate to obtain the motion amplitude gain. The closed-loop update module uses the transcranial magnetic stimulation pulse timing as input to the artifact suppression module, compares the mean difference of the phase lag index before and after transcranial magnetic stimulation to assess the intervention effect, and updates the stimulation intensity.

[0029] The following description, in conjunction with the accompanying drawings, further illustrates the process of this invention. Figure 1 As shown, this invention first performs multimodal acquisition (EEG+fNIRS), followed by TMS artifact suppression, then feature extraction and alignment. After graph attention aggregation, the phase lag index (PLI) is calculated, and then the PLI drives a three-level control response (threshold / beat / TMS / exoskeleton). Finally, closed-loop verification and dose update are performed. The TMS pulse timing is fed back to the TMS artifact suppression stage. These three steps constitute a complete technical solution and closed loop. The first step extracts and aligns EEG–fNIRS multimodal features under online TMS stimulation, outputting... The second step involves using a graph attention network... Jointly estimate the intention probability and the cortical excitability-fatigue phase lag index, output , , The third step is... The process involves driving adaptive thresholds, training beats, phase-locked TMS windows, and secondary confirmation with the exoskeleton. Following closed-loop verification using pre- and post-stimulation phase lag difference and dose updates, the triggered TMS timing flows back to the first step, forming a complete loop from the third step to the first, from the first step to the second, and from the second step to the third. Intermediate variables are explicitly passed step-by-step from the original matrix to the closed-loop evaluation, with a clear source. The main chain is as follows: .

[0030] The following is a further explanation with reference to the embodiments. The embodiments of the present invention first perform multi-channel bioelectrical synchronous sampling: ; ; In the formula, For the first Continuous scalp potentials were collected by one electrode, which was arranged in a 10–20 system and focused on covering the primary sensorimotor areas (C3, C2, C4 and their surroundings) because event-related desynchronization (ERD) of left and right limb motor imagery mainly occurs in this area. For the sampling rate, this scheme takes... Such a high sampling rate is not for ERD (ERD itself only requires hundreds of hertz), but to capture the microsecond-millisecond transients generated by TMS pulses in time and to model them accurately. The number of electrodes (32 or 64). The number of samples within a single sliding analysis window determines the temporal resolution of all subsequent features. This is the original multichannel matrix output in this step.

[0031] conduct First-order Butterworth bandpass filter (IIR) (amplitude-frequency response and difference equation implementation): ; ; In the formula, Angular frequency, The cutoff angular frequency is taken as the passband frequency. , This corresponds to the mu rhythm (8–13 Hz) and the beta rhythm (13–30 Hz). The filter order is 4. The Butterworth type has the flattest amplitude-frequency response in the passband, which can avoid introducing additional ripple distortion to the ERD amplitude. These are the numerator / denominator coefficients obtained by discretizing the passband boundary.

[0032] The process involves a three-stage cascaded process: blinding window definition, autoregressive reconstruction, and exponential decay artifact subtraction. The first stage is the blinding window (using the pulse timing synchronously provided by the stimulator). ): ; The second level is based on fitting the pre-pulse clean segment. An AR model is used to forward predict and reconstruct the signal within the blinded window: ; The third stage involves modeling and subtracting the residual attenuation artifacts after the pulse: ; ; In the formula, For the first The timing of each TMS pulse is directly given by the synchronous triggering of the stimulator (TTL / optical coupler), which is an external input with a clearly defined source in this step, without relying on reverse derivation from the EEG, thus avoiding the introduction of circular dependencies. This is a blinding margin before / after the pulse (taken as 2ms / 10ms), covering the transient main body and amplifier recovery period. The AR model coefficients are estimated by least squares from the clean signal preceding the pulse. The order is taken as 15 to 30. A sufficiently high order is needed to characterize the narrowband oscillation structure of mu / beta. The first The amplitude of each exponential component and the time constant, The number of components (ranging from 2 to 3) is obtained by least-squares fitting of the post-pulse residuals. .

[0033] Perform co-space mode spatial filtering: ; ; In the formula, The CSP projection matrix is ​​obtained by solving two types of covariance generalized eigenvalue problems during the offline training phase. For artifact suppression multichannel matrix; For the first One spatial projection component; The selected CSP logarithm is 3 (i.e., using the top 3 with the largest variance and the bottom 3 with the smallest variance, for a total of 6 components). Normalization makes the features approximate a normal distribution, which is beneficial for subsequent classification.

[0034] The frequency band energy changes related to motion imagery are estimated using the sliding window power spectrum, and the event-related desynchronization depth features are obtained after baseline normalization. ; ; In the formula, This refers to the instantaneous bandwidth power of the sliding window. The power smoothing window length is taken as the number of samples corresponding to 0.5s. The longer the smoothing window, the more stable it is, but the lower the time resolution. A compromise needs to be made. The resting baseline reference power (the average value approximately 1 second before the task) is used to normalize the absolute power of different subjects and different electrode impedances to relative changes.

[0035] The changes in oxygenated and deoxygenated hemoglobin concentrations were calculated from dual-wavelength near-infrared light intensity signals by modifying the Beer-Lamber law. ; In the formula, for The solution can be obtained by combining the above equations for the two wavelengths. and . The light intensity detected by the motor cortex. Baseline light intensity; The molar extinction coefficient of oxygenated / deoxygenated hemoglobin at this wavelength; Source – Exploring Distance; It is a differential path factor that compensates for the scattering and extension of light in the tissue.

[0036] The changes in HbO / HbR concentrations are integrated into a scalar blood oxygen load index that monotonically increases with fatigue, including two dimensions: oxygen exchange status and response amplitude decay. ; In the formula, This represents changes in oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration. The first term represents oxygen exchange status: during normal activation, HbO increases and HbR decreases; the ratio of these two values ​​characterizes the oxygen supply and demand match. During fatigue, this match shifts. The second term represents the attenuation of response amplitude. The peak amplitude of HbO response to motion imagination within the current sliding window. The smaller the ratio, the weaker the hemodynamic response and the deeper the fatigue, indicating the initial (fresh state) task period of training. Two weights (which can be calibrated according to the subject). To prevent small constants with a denominator of zero.

[0037] After subtracting the inherent hemodynamic delay, the blood oxygen load and EEG features are aligned on the time axis and concatenated into a unified multimodal feature vector: ; ; In the formula, This is the hemodynamic delay (typically about 5 seconds), which is the inherent time lag between neural activation and the peak of the blood oxygen response; The fNIRS sampling rate. By... Move forward on the timeline This aligns the second-level blood oxygenation response with the same neural event (millisecond-level EEG) that triggered it in time, resulting in the aligned blood oxygenation load. .

[0038] EEG electrodes and fNIRS channels are used as heterogeneous nodes to construct a graph structure weighted by spatial distance and functional connectivity. The node feature matrix and adjacency matrix are defined, and the graph is defined as follows. Total number of nodes ( EEG electrode nodes and (Number of fNIRS channel nodes). The node feature matrix consists of... The distribution to each node consists of: ; Weighted adjacency matrix (integrating anatomical proximity and functional coupling): ; In the formula, For the first Each node at time... Features: EEG nodes load from CSP components With ERD scalar (Excitatory information), fNIRS node loading aligned blood oxygen load (Fatigue information); It represents the feature dimension of a single node. For nodes Spatial distance on the scalp As a distance-scale parameter, this Gaussian term enables stronger connections between spatially adjacent electrodes / channels. To represent the functional connection strength between nodes, the magnitude-squared coherence of the mu / beta band is used to characterize the statistical coupling of neural activity between the two nodes.

[0039] For each node in the heterogeneous graph, multi-head attention coefficients are calculated within its neighborhood. Neighbor information is aggregated in an adaptive manner to update the node embedding; attention scoring is then performed. (Represents vector concatenation) ; Normalized attention coefficients within the neighborhood: ; Multi-head aggregation yields the first Layer node embedding: ; In the formula, It is a learnable linear transformation matrix; The attention vector maps the concatenated high-dimensional features to a scalar score. , representing a node For nodes The degree of importance; For nodes The set of neighbors. The attention weights are normalized. . The number of attention heads is set to 4–8, with each head using independent parameters. To capture different coupling modes, It is a non-linear activation (ELU). After... Output after layer stacking .

[0040] Attention pooling is used to aggregate node embeddings into a graph-level global embedding, which is shared by the three task branches. The node-level attention pooling coefficients and the graph-level global embedding are as follows: ; ; In the formula, For learnable pooled query vectors, For nodes Weights when aggregated into a graph-level representation ( ); For a moment The graph-level global embedding is then simultaneously fed into three task heads: the intent head, the excitability head, and the fatigue head, all three sharing the same [framework / framework]. .

[0041] The softmax classification head outputs the motion intention category and confidence probability from the global embedding: ; ; ; In the formula, These are intent header parameters; For each MI category (e.g., left hand / right hand / both feet / resting), there is the posterior probability. The determined intention category (which determines the direction of exoskeleton movement). The corresponding maximum posterior probability is used as a scalar for the confidence level of the intent.

[0042] The cortical excitability score is extracted from the global embedding using a regression head, and then subjected to slow-scale filtering and Hilbert transform to extract the instantaneous phase of excitability; excitability score (regression head): ; Its slow-scale fluctuations (around the slow band) The instantaneous phase is obtained by taking the Hilbert transform of the bandpass filter. ; ; In the formula, For excitatory head parameters; For cortical excitability scoring, the ERD depth at the EEG end is used during training. As a weakly supervised label, the amplitude of motion evoked potentials (MEPs) is used as a high-confidence calibration label at the moment when TMS is applied; The Discrete Hilbert Transform converts a real signal into an analytic signal. Take the complex argument to obtain the instantaneous phase The significance of slow-scale bandpass is that the slow fluctuations in excitability on a trial-by-trial scale are comparable to the time scale of fatigue.

[0043] The neural fatigue score is output from the regression head. After compensating for the low sampling rate of fNIRS with triple spline upsampling, the instantaneous phase of fatigue is extracted, and the fatigue score (regression head) is obtained. ; The low sampling rate signal is upsampled to the control rate via cubic spline interpolation and then bandpassed. Then take the analytical signal: ; ; In the formula, For fatigue head parameters; For assessing nervous fatigue, the blood oxygen load index was used during training. The trend and the decline in decoding accuracy are used as monitoring. Since the fNIRS sampling rate (approximately 5–10 Hz) is much lower than the EEG / control rate, direct phase estimation will be distorted due to insufficient time resolution; therefore, cubic splines are first used to... Interpolation to Consistent rate, then bandpass and Hilbert transform .

[0044] The phase difference between the excitatory phase and the fatigue phase is calculated using circular statistical methods, and the normalized phase lag exponent is obtained by phase-locked consistency weighting; point-by-point winding phase difference: ; Consistency between the average phase difference within the sliding window and phase-locked loop: ; ; Phase lag exponent (amplitude × consistency, normalized to) ): ; In the formula, The instantaneous quantity in which the excitatory phase leads or lags the fatigue phase. Fold it into the principal value range to avoid Jump. Set the sliding statistical window length and set it to patient-adaptive: when fatigue changes rapidly ( (Large) time shortened To improve tracking speed, and conversely to increase stability, the length is increased, which directly addresses the second difficulty. The mean phase difference within the window (the correct averaging method for phase under circular statistics). For phase-locked loop consistency (the more concentrated the phase difference) The closer to 1, the more random it is, and the closer to 0. Ultimately... When the lag is large and the direction is consistent, the higher value is taken; when the lag fluctuates randomly, the higher value is taken. Suppress it.

[0045] The discounting probability is quantified using the phase lag exponent, thus suppressing the confidence of intent during fatigue decompensation: ; In the formula, The modulation intensity coefficient controls the suppression of confidence level by phase lag. When (No significant lag) The confidence level remains unchanged; when As the exponent increases, the exponent term decays rapidly. It was suppressed.

[0046] The intention to receive threshold is modulated by exponential linear modulation of phase lag, so that the deeper the fatigue, the higher the threshold for receiving intention. Dynamic threshold: ; Decode the receiving gate: ; In the formula, The baseline decision threshold for the fresh (fatigue-free) state is determined by offline cross-validation; The modulation slope of the threshold due to phase lag determines "how high the threshold will be raised once fatigue lags". This is an indicator function; it takes a value of 1 if the condition is true (accept the intention), and 0 otherwise (reject / consider it as silent). This is the decoder gate output in this step.

[0047] The cumulative fatigue load is obtained by integrating the phase lag exponent using leakage. Based on this, the training pace is adaptively slowed down, and a rest is inserted when the threshold is reached. The leakage integral of the phase lag (cumulative fatigue load) is: ; Trial interval (beat) adaptation and micro-rest triggering: ; ; In the formula, The accumulated fatigue load is obtained by integrating the phase hysteresis exponent with the first-order leakage integral. It acts as a forgetting factor (the closer to 1, the longer the memory), smoothing out momentary jitters and accumulating long-term trends. The baseline trial interval, As a rate slowdown factor, the deeper the fatigue, the longer the interval between adjacent motion visualization tasks and the slower the rhythm. This is the threshold for triggering micro-rests; exceeding this threshold will insert a short rest block into the training process.

[0048] When cumulative fatigue reaches the TMS intervention threshold and the excitatory phase enters the optimal window, a TMS pulse is delivered at the predicted optimal phase time. Triggering conditions: (cumulative fatigue reaches the threshold and the excitatory phase enters the optimal window). ; Phase predictive pulse timing (retrieves the next phase hit time): ; In the formula, The threshold for TMS intervention (above the rest threshold, reflecting graded escalation). The optimal stimulation phase is pre-calibrated—applying TMS at this phase is most effective in increasing excitability and shortening hysteresis; This is the phase tolerance window. For logical AND. The first step determined in this process The moment of TMS pulse delivery, which is synchronously triggered by the stimulator, is precisely the external input for artifact suppression. Stimulation intensity Use low doses.

[0049] The three-condition logic of "decoding acceptance, low fatigue phase, and stable intent timing" serves as a hard condition for significantly enhancing exoskeleton execution; otherwise, the performance is reduced. A secondary confirmation gate (decoding acceptance, low fatigue phase, and stable intent timing) is used. ; Motion amplitude gain: ; In the formula, For low fatigue phase conditions, The phase lag confirmation threshold is used; only when the lag is sufficiently small (excitability and fatigue rhythms are essentially in phase and there is no decompensation) is the process allowed. The multiplication term is used to verify the temporal stability of the intent, requiring the most recent... They were classified as belonging to the same category at that time. To avoid triggering actions due to sudden jumps. All three conditions must be met. The exoskeleton provides significant assistance in this direction, increasing the gain. If any condition is not met, the amplitude will decrease to low or no action will be taken, and the low-amplitude gain will increase accordingly. The larger it is, the smaller it becomes.

[0050] The desired motion trajectory is selected based on the intent category, joint assist torque is generated using impedance control, and uniformly scaled using amplitude gain. ; In the formula, This refers to the assist torque applied to the exoskeleton joints; For the purpose of class Selected desired joint angle and angular velocity trajectory (e.g., reference trajectory for flexion and extension of the left / right lower limb); Feedback on the actual angle and angular velocity of the joint; For impedance stiffness and damping coefficient; It represents the gain in motion amplitude and serves as a scaling factor for the overall assist.

[0051] The intervention effect was assessed by comparing the changes in phase hysteresis index before and after TMS stimulation, and the stimulation dose was updated online accordingly and fed back to the stimulation triggering stage. The difference in the window mean of the phase lag index before and after the second stimulus: ; Stimulus intensity closed-loop update (including safety limits): ; In the formula, The first Within the short window before and after the TMS The mean; where The calculation depends on the recovery of the EEG after stimulation; otherwise, the excitatory phase after stimulation cannot be measured. This indicates that the lag has been shortened (the stimulus is effective); Indicates invalidity or deterioration. For dose learning rate, Limiting the intensity to a safe range—stimulation is ineffective. If the intensity is too high, it should be adjusted accordingly; if it is effective, it should be maintained or lowered, but it should never exceed the safe upper limit.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for phase compensation of cortical excitability fatigue during exoskeleton training, characterized in that, include: S1. Collect multimodal brain signals and extract the bandpass signal. Use the transcranial magnetic stimulation pulse time as the anchor point to perform artifact suppression processing to obtain the artifact-suppressed signal. Construct a cospatial pattern feature vector based on the artifact-suppressed signal and extract the desynchronized sliding window mean scalar. Extract the changes in oxyhemoglobin concentration and deoxyhemoglobin concentration based on the dual-wavelength near-infrared light intensity signal. After time alignment, concatenate it with the cospatial pattern feature vector and the sliding window mean scalar to form a multimodal feature vector. S2. Assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct node feature matrices. Construct a weighted adjacency matrix using spatial distance and functional connectivity strength to obtain a heterogeneous graph. Utilize multi-head graph attention and attention pooling to obtain graph-level global embedding. Based on graph-level global embedding, output the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. Obtain the phase lag index by weighting the phase difference between the inner circle of the sliding window and phase-locked consistency. Quantize and discount the phase lag index to obtain the confidence after fatigue modulation. S3. Calculate the decoding acceptance gate using the phase lag exponent linearly modulated intent acceptance threshold. Calculate the cumulative fatigue load using the phase lag exponent and adaptively adjust the trial interval. Insert a rest when the fatigue load reaches the micro-rest threshold. Trial a transcranial magnetic stimulation (TMS) pulse when the fatigue load reaches the TMS intervention threshold and the excitatory instantaneous phase is in the optimal phase range. Construct a secondary confirmation gate based on the decoding acceptance gate, the phase lag exponent being below the confirmation threshold, and the intent category remaining unchanged, to obtain the action amplitude gain. Use the TMS pulse timing as the artifact suppression input, compare the mean difference of the phase lag exponent before and after TMS to evaluate the intervention effect, and update the stimulation intensity and limit the amplitude based on the mean difference.

2. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 1, characterized in that, S1 includes, S1.1, acquiring multi-channel scalp EEG signals at a sampling rate of... collection The scalp potential of each electrode is used to construct a discrete primitive matrix. ;use The Butterworth bandpass filter extracts the bandpass signal from the scalp potential. , Electrode index; S1 includes S1.2, and the artifact suppression processing includes a three-level cascaded process of constructing a blinding window definition, autoregressive reconstruction, and exponential decay artifact subtraction. The blinding window definition includes the timing of synchronous output pulses using a stimulator. Defining the contamination window for each pulse : ； In the formula, For the first Timing of each transcranial magnetic stimulation pulse This is the pre-pulse blinding margin. This is the blinding margin after the pulse. Index of transcranial magnetic stimulation pulse times; Autoregressive reconstruction includes estimating the pre-pulse clean segment. AR model coefficients, forward prediction to reconstruct signals within the blinded window: ； In the formula, For the first The estimated value of the reconstructed signal of the channel within the blinding window. For AR model coefficients, For order, For discrete-time indexing, Delayed indexing for AR models ; Exponential decay artifact reduction includes modeling and subtracting the residual decay artifact after the pulse: ； ； In the formula, For the first Fitting model for residual artifacts after channel pulses. For the first Channel 1 The magnitude of each exponential component, For the first Channel 1 The time constant of each exponential component Number of components For component index, for The estimated value, For the first The signal after artifact suppression in the channel.

3. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 2, characterized in that, S1 includes, S1.3, and will Apply the cospace mode projection matrix and select the one with the largest variance after projection. The component with the smallest variance Each component is used to construct spatial filtering features. And construct the common space pattern feature vector , , It is the transpose symbol. For the projection component index, The logarithm of the common space mode components; based on Calculate the instantaneous frequency band power of the sliding window, and normalize it with the resting baseline reference power to obtain the time. Desynchronization percentage and extract sliding window mean scalar To modify Beer-Lambert's law, the changes in oxyhemoglobin concentration were calculated from dual-wavelength near-infrared light intensity signals. Changes in deoxyhemoglobin concentration .

4. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 3, characterized in that, S1 includes, S1.4, and will and Combined into a scalar blood oxygen load index : ； In the formula, This represents the peak amplitude of the response of oxygenated hemoglobin within the current sliding window to motion imagery. To train the response amplitude during the initial task period, , For two weights, To prevent small constants with a denominator of zero; Based on functional near-infrared spectral sampling rate and hemodynamic delay Constructing aligned blood oxygen load : ； Will and and Align them on the timeline and concatenate them into a multimodal feature vector. : ； In the formula, For a moment of , For a moment of .

5. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 4, characterized in that, S2 includes, S2.1, and A node feature matrix was constructed by assigning nodes to EEG nodes and functional near-infrared spectroscopy nodes. A weighted adjacency matrix was then constructed based on spatial distance and functional connectivity strength to obtain a heterogeneous graph. The EEG nodes include... and Functional near-infrared spectral nodes include ; For each node in the heterogeneous graph, calculate the multi-head attention coefficient in its neighborhood and update the node embedding; Attention pooling is used to aggregate node embeddings into graph-level global embeddings; The softmax classification head outputs the motion imagination intent category from the graph-level global embedding. With confidence probability; The regression head is used to embed the cortical excitability score from the graph-level global embedding, and the excitability instantaneous phase is extracted by slow-scale filtering and Hilbert transform. ; The neural fatigue score is output by the regression head, and after near-infrared spectral sampling rate compensation using a cubic spline upsampling function, the instantaneous phase of fatigue is extracted. .

6. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 5, characterized in that, S2 includes S2.2, calculating the phase difference between the excitatory phase and the fatigue phase using circular statistical methods. : ； In the formula, For phase winding function; Calculate the mean phase difference of the inner circle of the sliding window Consistency with phase-locked loop : ； ； In the formula, For the summation index within the sliding window, For sliding statistics windows, To reduce phase difference Represented as a unit complex number, The imaginary unit, To determine the mold length; The normalized phase lag exponent is obtained by phase-locked loop consistency weighting. : ； by Quantify the probability of discount intent and calculate the confidence level after fatigue modulation: ； In the formula, For a moment Intention confidence after fatigue modulation For the probability of intent, This is the modulation intensity coefficient.

7. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 6, characterized in that, S3 includes, S3.1, an intentional acceptance threshold using exponential linear modulation with phase lag. Calculate the decoding reception gate : ； ； In the formula, The baseline decision threshold is the threshold under fatigue-free conditions. The modulation slope of the threshold due to phase lag. For indicator functions; Calculating cumulative fatigue load using phase hysteresis index : ； In the formula, Forgetting factor; The adaptive trial interval is: ； In the formula, For a moment The adaptive trial interval, The baseline trial interval, This is the beat slowdown factor; Set micro-rest trigger threshold ,like Insert rest block The system pauses during the rest block.

8. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 7, characterized in that, S3 includes, S3.2, when Reaching the threshold for transcranial magnetic stimulation intervention and Upon entering the optimal phase interval, a transcranial magnetic stimulation pulse is delivered at the predicted optimal phase time, triggered by the following conditions: ； In the formula, For the first Triggering markers for transcranial magnetic stimulation The threshold for transcranial magnetic stimulation intervention. For the pre-calibrated optimal stimulus phase, For phase tolerance window, For logical AND operator; Phase prediction pulse timing is as follows: ； In the formula, For search indexing, Index for the current time, To obtain the minimum value.

9. The method for phase compensation of cortical excitability fatigue in exoskeleton training according to claim 8, characterized in that, S3 includes, S3.3, constructing a secondary confirmation gate: ； In the formula, For a moment Exoskeleton execution gate, For low fatigue phase conditions, For the purpose of timing stability verification, The phase lag confirmation threshold, For stability verification window length, For the index within the verification window; Motion amplitude gain for: ； In the formula, The base coefficient for low-amplitude gain; S3 includes S3.4, selecting the desired motion trajectory based on the intent category and uniformly scaling it using amplitude gain: ； In the formula, For a moment The assist torque, For impedance-controlled proportional gain, For the purpose of class Selected desired joint angle, This provides feedback on the actual angle of the joint. For impedance-controlled differential gain, For the purpose of class Selected desired joint angular velocity, Feedback on the actual angular velocity of the joint; S3 includes S3.5, and the difference in mean PLI values ​​before and after stimulation. : ； In the formula, For the first Short window after the second stimulation The mean, For the first Within the short window before the second stimulation mean Perform closed-loop updates of stimulus intensity: ； In the formula, For the updated stimulus intensity, For the amplitude limiting function, For the first The intensity of the secondary pulse For dose learning rate, This is the safe lower limit of stimulus intensity. This represents the safe upper limit for the intensity of the stimulus.

10. An exoskeleton training cortical excitability fatigue phase compensation system, characterized in that, include: The signal acquisition module is used to acquire multimodal brain signals; The artifact suppression module is used to perform artifact suppression processing on the acquired multimodal brain signals with the transcranial magnetic stimulation pulse time as the anchor point, and output the artifact-suppressed signal. The feature extraction module is used to construct a co-space mode feature vector from the artifact suppression signal and extract the desynchronized sliding window mean scalar, as well as extract the changes in oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration from the dual-wavelength near-infrared light intensity signal, and splice them into a multimodal feature vector after time alignment. The graph processing module is used to assign multimodal feature vectors to EEG nodes and functional near-infrared spectroscopy nodes to construct heterogeneous graphs. Graph-level global embedding is obtained through graph attention aggregation. Based on the graph-level global embedding, the module outputs the motor imagery intention category, confidence probability, excitatory instantaneous phase, and fatigue instantaneous phase. The phase lag index is obtained by weighting the phase difference of the inner circle of the sliding window with phase-locked consistency. The discounted confidence probability is quantified by the phase lag index. The control response module is used to modulate the decoding threshold with phase lag exponential modulation, adapt the trial interval and insert rest, trigger transcranial magnetic stimulation pulses, and construct a secondary confirmation gate to obtain the motion amplitude gain. The closed-loop update module uses the transcranial magnetic stimulation pulse timing as input to the artifact suppression module, compares the mean difference of the phase lag index before and after transcranial magnetic stimulation to assess the intervention effect, and updates the stimulation intensity.

Images