A motor imagery training system based on brain-computer interface

Abstract

The application provides a motor imagination training system based on a brain-computer interface, and relates to the technical field of intelligent control, and comprises: a rapid processing module, which is used for acquiring electroencephalogram signals when a training video is watched and performing time-frequency domain and spatial mode processing to obtain dynamic characteristics, and determining a target training video according to a mirror neuron activation probability in the dynamic characteristics; a fine processing module, which is used for acquiring multi-modal sensing signals when the target training video is watched to determine static characteristics of a brain motor cortex, the static characteristics including activation intensity of the brain motor cortex, connection intensity of inter-cortical functions and coupling intensity of nerves and muscles; and an adaptive adjustment module, which is used for adaptively adjusting at least one of an exoskeleton torque, a virtual reality task difficulty and a neural feedback sensitivity according to the activation intensity, the connection intensity and the coupling intensity. The application realizes individualized adaptation of training content and real-time dynamic closed-loop regulation and control of parameters, and improves the effect of individualized training.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to a motor imagery training system based on a brain-computer interface. Background Technology

[0002] In motor imagery training tasks, playing motion videos often stimulates the mirror neuron system of the target user, thereby promoting the plasticity of the motor cortex and the effectiveness of motor imagery. Traditional training procedures generally use pre-set video stimuli directly and then proceed to the complete training phase. However, this approach has drawbacks, such as the inability to determine the effectiveness of the training video stimuli for the target user, which may lead to the target user watching ineffective stimuli for extended periods, and the inability to adjust video stimulus parameters in real time during training.

[0003] While existing technologies attempt to build models to address the aforementioned issues, current lightweight models are fast but lack accuracy, while deep models are accurate but have high latency. They have consistently failed to provide a solution that can identify the effectiveness of video stimuli for target users before motor imagery training and can adaptively adjust in real time. Summary of the Invention

[0004] The purpose of this application is to provide a brain-computer interface-based motor imagery training system. Innovatively, before the start of motor imagery (MI) training, it uses a parallel dual-module architecture to decode EEG to quickly detect the activation state of the mirror neuron system. Based on the detection results, it selects suitable motor imagery training videos, extracts multimodal neurobiological signal coupling, and finally performs fuzzy adaptive adjustment of motor imagery training.

[0005] In some embodiments, this application provides a brain-computer interface-based motor imagery training system, the system comprising: a fast processing module, configured to acquire the electroencephalogram (EEG) signals corresponding to a target user watching a training video, and perform time-frequency domain and spatial pattern processing on the EEG signals to obtain dynamic features, and determine the target training video based on the activation probability of mirror neurons in the dynamic features; a fine processing module, configured to acquire the multimodal sensor signals corresponding to the target user watching the target training video, and determine the static features of the motor cortex of the brain based on the multimodal sensor signals, the static features including the activation intensity of the motor cortex of the brain determined based on the EEG signals, the connectivity strength of intercortical functions, and the coupling strength between nerves and muscles determined jointly by the EEG signals and electromyogram (EMG) signals in the multimodal sensor signals; and an adaptive adjustment module, configured to adaptively adjust at least one of the following: exoskeleton torque, virtual reality task difficulty, and neural feedback sensitivity based on the activation intensity of the motor cortex of the brain, the connectivity strength of intercortical functions, and the coupling strength between nerves and muscles.

[0006] In this way, the rapid processing module quickly selects target training videos that can effectively activate mirror neurons based on the dynamic characteristics of EEG signals before training, avoiding the problem of insufficient adaptation caused by the target user watching ineffective stimuli for a long time. At the same time, the fine processing module extracts multi-dimensional static features including activation intensity, connectivity intensity, and coupling intensity, providing precise neuromuscular functional representation input for the adaptive adjustment module. This enables dynamic closed-loop control of training parameters based on the target user's real-time physiological and neural state, overcoming the problem of lack of dynamic control caused by the mechanization of traditional training processes.

[0007] In some embodiments, the fast processing module further includes: a feature extraction unit, configured to acquire the electroencephalogram (EEG) signal corresponding to the target user watching the training video, extract time-frequency domain features from the EEG signal according to a preset time period to obtain time-frequency feature data; extract spatial pattern features from the time-frequency feature data to obtain spatial projection data; an activation probability calculation unit, configured to calculate the activation probability of mirror neurons based on the spatial projection data; and a target training video determination unit, configured to determine the activation state of mirror neurons based on the activation probability, and determine the target training video to be used based on the activation state.

[0008] In this way, by extracting features in both the time-frequency domain and spatial patterns by the feature extraction unit according to a preset time period, it is possible to effectively capture the event-related desynchronization and synchronization phenomena caused by action observation or imagination, and highlight the differences between the left and right motor cortices under different video stimuli. This provides the activation probability calculation unit with highly discriminative spatial projection data, making the calculation of activation probability more accurate. In turn, the target training video determination unit can match the most suitable training video for the target user based on the accurate activation state.

[0009] In some embodiments, the fast processing module further includes: a filter matrix determination unit, configured to determine a filter matrix corresponding to the category of the target user attribute in a spatial mode, project the EEG signal corresponding to the target user watching the training video onto the filter matrix to obtain a projection signal, calculate a covariance matrix using the projection signal, and normalize the variance of the projection signal according to the covariance matrix to obtain normalized feature data corresponding to the spatial projection data; wherein, the activation probability calculation unit inputs the normalized feature data into a machine learning classifier to obtain the activation probability; wherein, the target training video determination unit judges the activation state of the mirror neurons based on the activation probability, and in response to being in an activated state, determines the current training video as the target training video; in response to being in an inactive state, switches to the next training video after the current training video finishes playing, and continues to calculate the activation probability according to the next training video until it is determined to be in an activated state according to the activation probability, at which point the latest training video is determined as the target training video.

[0010] In this way, the filter matrix determination unit loads matching filter matrices for users of different attribute categories and normalizes the variance of the projected signal, eliminating the influence of individual differences and uneven frequency band energy distribution, making the normalized feature data input to the machine learning classifier more robust. At the same time, the target training video determination unit adopts a cyclical screening mechanism based on activation probability to ensure that the finally selected target training video is the first video that can effectively activate mirror neurons, significantly improving the effectiveness of training stimuli and the degree of personalized adaptation.

[0011] In some embodiments, the feature extraction unit further includes: a first category user filter selection subunit, configured to load a first initial category filter matrix in the current time window and a currently updated filter matrix in subsequent time windows in response to the target user being a first category user; wherein the first initial category filter matrix is ​​determined based on the difference in covariance matrices between the collected resting period signal and the stimulation period signal of the target user, and the resting period signal is the EEG signal of the target user in a task-free relaxation state, and the stimulation period signal is the EEG signal before the guiding stimulus induces the target user to enter the activation state; and a second category user filter selection subunit, configured to load a second initial category filter matrix in the current time window and a currently updated filter matrix in subsequent time windows in response to the target user being a second category user; wherein the second initial category filter matrix is ​​preset.

[0012] In this way, by setting a differentiated filter selection strategy for the first and second categories of users, the initial filter for the first category of users who lack prior data is calculated unsupervised by utilizing the difference in the covariance matrix between their resting and stimulated periods, thus achieving zero-sample cold start; for the second category of users who have prior conditions, the pre-trained general filter matrix is ​​directly loaded, saving initialization time. This ensures the accuracy of feature extraction while taking into account the initialization efficiency and system availability for different user groups.

[0013] In some embodiments, the fast processing module further includes: a filter matrix update unit, configured to calculate the covariance matrix of the current time window based on the collected target user's current resting period signal and stimulation period signal, recursively combine the covariance matrix of the current time window with the covariance matrix of the previous time window to obtain an updated covariance matrix, adjust the filter weights based on the updated covariance matrix, and determine the current updated filter matrix.

[0014] In this way, by using a recursive approach to combine the covariance matrix of the current time window with the historical covariance matrix through the filter matrix update unit, the periodic smooth update of the filter weights is achieved. This allows the spatial filter to gradually adapt to the drift changes of the user's EEG pattern over time, continuously improving the matching degree between the filter and individual EEG features while maintaining real-time computing performance, effectively overcoming the problem of decreased classification accuracy caused by individual differences during long-term training.

[0015] In some embodiments, the fine processing module includes: an activation intensity calculation unit, used to extract bispectral feature data of μ rhythm from the EEG signal to obtain the activation intensity of the motor cortex; a connectivity intensity calculation unit, used to filter the EEG signal, extract the inferior parietal lobule signal and the primary motor cortex signal respectively, calculate the phase lock value between the inferior parietal lobule signal and the primary motor cortex signal to obtain the connectivity intensity of intercortical functions; and a coupling intensity calculation unit, used to filter the EEG signal, calculate the time delay cross-correlation value between the EEG signal and the corresponding electromyographic signal to obtain the coupling intensity between nerves and muscles.

[0016] In this way, by activating the intensity calculation unit to extract bispectral feature data of the μ rhythm to capture the non-Gaussianity of neural oscillations, by connecting the intensity calculation unit to calculate the phase-locking value across brain regions to detect information transmission efficiency, and by coupling the intensity calculation unit to calculate the time-delay cross-correlation value of EEG and EMG to quantify the central-peripheral nerve coupling strength, the three components construct a complete static feature system from three complementary dimensions: nonlinear activation, cross-brain region functional connectivity, and neuro-muscle synergy, which greatly improves the system's ability to distinguish complex movement patterns.

[0017] In some embodiments, within the fine processing module, the activation intensity calculation unit is specifically used to extract signals within a preset frequency band range in the motor cortex of the brain, calculate bispectral peaks, and determine the activation intensity of the motor cortex based on the positive correlation between the bispectral peaks and the activation intensity of the motor cortex; the connectivity strength calculation unit is specifically used to acquire signals within a preset frequency band range in the inferior parietal lobule and signals within a preset frequency band range in the primary motor cortex, respectively, and calculate the corresponding instantaneous phases, and calculate the phase lock value between the inferior parietal lobule and the primary motor cortex based on the instantaneous phases; determine the connectivity strength between the functions of the cortex based on the positive correlation between the phase lock value and the connectivity strength of the intercortical functions; the coupling strength calculation unit is specifically used to extract signals within a preset frequency band range in the motor cortex of the brain, calculate their time-delay cross-correlation value with the corresponding electromyographic signals within a time-delay window; determine the coupling strength between the nerve and muscle based on the positive correlation between the time-delay cross-correlation value and the coupling strength between the nerve and muscle.

[0018] In this way, by clarifying the positive correlation mapping relationship between bispectral peaks, phase-locked values, and time-delay cross-correlation values ​​and each static feature, and by defining specific frequency band ranges and time-delay windows, abstract neurophysiological concepts are transformed into quantifiable signal processing indicators. This ensures the accuracy and reproducibility of static feature extraction, enabling the features output by the fine processing module to truly reflect the specific activation patterns and central-peripheral matching of the target user's motion intentions.

[0019] In some embodiments, the adaptive adjustment module includes: a motion intention decision unit, configured to construct a hybrid architecture model, capture the nonlinear characteristics of the bispectral peaks, extract the dynamic change characteristics of the phase-locked value, analyze the synergistic patterns of nerves and muscles, and output a motion intention decision; wherein the motion intention decision includes motion intention confidence, motion pattern, and system stability; a torque control unit, configured to adjust the exoskeleton torque according to the motion intention confidence and the phase-locked value in the motion intention decision; wherein the phase-locked value is inversely proportional to the exoskeleton torque; a difficulty control unit, configured to adjust the virtual reality task difficulty according to the bispectral feature data of the μ rhythm and the classification results of the motion pattern; and a sensitivity control unit, configured to adjust the neural feedback sensitivity according to the coefficient of variation of the time delay cross-correlation value; wherein the neural feedback sensitivity is proportional to the signal interference.

[0020] In this way, the motion intention decision unit uses a hybrid architecture model to specifically process static features of different dimensions, capturing both long-term average nonlinear features and adapting to real-time dynamic changes. The output multi-dimensional decision vector comprehensively describes the motion intention. The torque control unit, difficulty control unit, and sensitivity control unit adjust the exoskeleton torque, task difficulty, and feedback sensitivity based on specific indicators in the decision vector, respectively, realizing humanoid intelligent closed-loop control in three dimensions: proportional, integral, and differential. This ensures precise dynamic adaptation of training support strength, challenge intensity, and system stability.

[0021] In some embodiments, the motion intention decision unit further includes: a motion intention classification subunit, used to identify motion intentions using the hybrid architecture model and the static features, and classify each type of motion intention; and a motion intention activation identification subunit, used to identify whether the corresponding category of motion intention is in an active state using the hybrid architecture model and the dynamic features, and in response to being in an active state, to execute a training action corresponding to the category of motion intention, and in response to being in an inactive state, not to execute a training action this time.

[0022] In this way, the motion intention classification subunit performs coarse-grained classification of motion intentions based on static features, and the motion intention activation recognition subunit performs fine-grained activation state determination of specific categories of intentions based on dynamic features. This forms a hierarchical decision-making mechanism of classification first and confirmation later, avoiding misjudgment caused by a single feature dimension. This ensures that the system only triggers training actions when valid activation is confirmed, thereby improving the reliability and safety of brain-computer interface control.

[0023] In some embodiments, the adaptive adjustment module further includes a protection mechanism unit, used to activate a protection mechanism and reduce the training intensity when the coefficient of variation of the time delay cross-correlation value is abnormal.

[0024] In this way, by monitoring the coefficient of variation of the time delay cross-correlation value through the protection mechanism unit, when it is abnormal (such as exceeding 25%), it indicates that there is serious instability or abnormal signal interference in the central-peripheral nerve coupling. At this time, the protection mechanism is immediately activated to reduce the training intensity, which effectively prevents the target user from making mistakes or secondary injuries due to high-intensity training in a state of neuromuscular coordination failure, and ensures the safety and stability of the training closed loop.

[0025] Compared with the prior art, the present invention has the following technical effects:

[0026] Before motor imagery (MI) training begins, a parallel dual-module architecture is used to decode EEG signals to quickly detect the activation state of the mirror neuron system. Based on the detection results, suitable MI training videos are selected, and then multimodal neurobiological signal coupling is extracted. Finally, fuzzy adaptive adjustments are made to the MI training. The parallel dual-module architecture includes a fast processing module and a fine processing module. The fast processing module rapidly calculates the spatial patterns of the cerebral cortex within a preset frequency band, detects the activation of the mirror neuron system, and quickly outputs preliminary motor intention recognition results. When the fast processing module does not detect significant motor intention signals, the next training video is played after the current training video has finished playing. When a significant motor intention signal is detected, the system selects the current training video and automatically triggers the fine processing module. The fine processing module extracts bispectral feature data of μ rhythm from the EEG signals to obtain the activation intensity of the motor cortex, extracts signals from the inferior parietal lobule and primary motor cortex, calculates the phase-locking value between the inferior parietal lobule and primary motor cortex signals to obtain the functional connectivity strength between cortices, and also calculates the time-delay cross-correlation value between the EEG signals and the corresponding electromyographic signals to obtain the coupling strength between nerves and muscles. Finally, the adaptive adjustment module employs a fuzzy PID algorithm to achieve closed-loop control of motor imagery training, adjusting parameters such as exoskeleton torque, virtual reality task difficulty, and neural feedback sensitivity in real time. This provides a personalized motor imagery training environment for target users, automatically recommending the most suitable training content to improve training adaptability, reduce frustration, and help them enter an effective motor imagery rhythm. Simultaneously, it dynamically adjusts training intensity to ensure each training session is just right, avoiding both insufficient intensity that affects results and overtraining that burdens the target user. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:

[0028] Figure 1 This is a schematic diagram of the overall system flow provided in one embodiment of this application;

[0029] Figure 2 This is a flowchart illustrating a fast processing module provided in one embodiment of this application;

[0030] Figure 3 This is a flowchart illustrating the fine processing module provided in one embodiment of this application;

[0031] Figure 4 This is a flowchart illustrating the adaptive adjustment module provided in one embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] The technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0034] This solution does not aim to obtain disease diagnosis results or health status. It is a system that processes the user's electroencephalogram (EEG) and electromyogram (EMG) data to achieve motor imagery training. All steps are performed by information processing methods implemented by computers and other devices.

[0035] It should be fully understood that the user information involved in this application (including but not limited to EEG signals and EMG signals) is information and data authorized by the user or fully authorized by all parties. The use of user information shall comply with the privacy policies and practices of the industry that are generally considered to meet or exceed the requirements for maintaining user privacy. The collection, use and processing of related data shall comply with relevant laws, regulations and standards, and provide corresponding operation access points for users to choose to authorize or refuse.

[0036] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0037] Example 1:

[0038] In some embodiments, such as Figures 1-4 As shown, this embodiment provides a brain-computer interface-based motor imagery training system. The system includes a fast processing module, a fine processing module, and an adaptive adjustment module. It should be understood that although the serial data flow of the three modules is shown, in other embodiments, depending on the allocation of computing resources and real-time requirements, some preprocessing steps of the fast processing module and the fine processing module can also be executed in parallel, as long as the progressive logic of filtering followed by fine extraction is satisfied.

[0039] The rapid processing module is used to acquire the EEG signal corresponding to the target user when watching the training video, and to perform time-frequency domain and spatial pattern processing on the EEG signal to obtain dynamic features, and to determine the target training video based on the activation probability of mirror neurons in the dynamic features.

[0040] Specifically, the "dynamic features" referred to in this invention refer to the feature data that reflects the instantaneous neural activity state obtained by the system rapidly extracting EEG signals within a short time window during the video pre-screening stage. Examples include time-frequency energy changes in event-related desynchronization or synchronization phenomena, and instantaneous projections of spatial distribution patterns in different brain regions. This embodiment uses a rapid processing module to quickly screen target training videos that can effectively activate mirror neurons based on the dynamic features of EEG signals before training, solving the problem of traditional systems directly using predetermined video stimuli without verifying effectiveness. In traditional motor imagery training, lightweight decision models, while having low computational latency and rapid response, have a single feature extraction dimension, often relying only on shallow features such as power spectra, leading to insufficient judgment accuracy and a tendency to misjudgment. While deep learning models offer high accuracy, their high computational complexity and significant latency make them difficult to meet the real-time requirements of rapid pre-screening before training. Therefore, the fast processing module in this embodiment adopts a dual processing mechanism of time-frequency domain and spatial mode. While ensuring low latency, it enhances the distinguishability of features through spatial mode projection, thereby achieving a balance between the speed of the lightweight model and the accuracy of the deep model. This avoids the lack of adaptability and frustration caused by the target user watching ineffective stimuli for a long time.

[0041] The fine processing module is used to acquire multimodal sensing signals corresponding to the target user when watching the target training video, and determine the static characteristics of the motor cortex of the brain based on the multimodal sensing signals. The static characteristics include the activation intensity of the motor cortex of the brain, the connectivity strength between cortical functions, and the coupling strength between nerves and muscles, determined by the combination of EEG and electromyography signals in the multimodal sensing signals.

[0042] Specifically, the "static features" referred to in this invention are feature data that reflect the steady-state representation of neuromuscular function, extracted by the system through in-depth analysis of multimodal signals under a longer time window or steady state after confirming the target training video and entering the substantive training phase. Unlike the instantaneous dynamic features extracted by the fast processing module, static features focus on characterizing the deep nonlinear activation of the motor cortex during continuous imagination, the steady-state synchronization of cross-brain region information transmission, and the stable coupling of central commands to peripheral muscles. For example, activation intensity can be captured by extracting high-order spectral features of specific frequency bands to capture the non-Gaussianity of neural oscillations; connectivity strength can be detected by calculating the phase-locking value of cross-brain region signals to detect information transmission efficiency; and coupling strength can be quantified by calculating the time-delay cross-correlation value of EEG and EMG signals to quantify the matching of central-peripheral neural synergy. This embodiment extracts multi-dimensional static features including activation strength, connectivity strength, and coupling strength through a fine processing module, constructs a three-dimensional coupling mechanism of spatial-temporal-frequency domain features, and forms a more complete neuromuscular function representation system. This makes up for the shortcomings of single feature or shallow feature representation capabilities and provides accurate and robust input parameters for subsequent adaptive regulation.

[0043] The adaptive adjustment module is used to adaptively adjust at least one of the following based on the activation intensity of the motor cortex, the connectivity strength between cortical functions, and the coupling strength between nerves and muscles: exoskeleton torque, virtual reality task difficulty, and neural feedback sensitivity.

[0044] Specifically, the adaptive adjustment module receives the three-dimensional static features output by the fine processing module and uses them as input variables for closed-loop control. It then dynamically adapts the training parameters through humanoid intelligent control strategies (such as fuzzy logic reasoning). Specifically, the adjustment of exoskeleton torque primarily affects the physical level of auxiliary support; the adjustment of virtual reality task difficulty primarily affects the cognitive and motor level of challenge intensity; and the adjustment of neural feedback sensitivity primarily affects the system's ability to suppress abnormal signal interference and its response speed. It should be understood that although this embodiment lists exoskeleton torque, virtual reality task difficulty, and neural feedback sensitivity as three specific adjustment objects, in other embodiments, the adaptive adjustment objects can be extended to parameters such as training duration, video playback speed, and feedback screen color mapping, as long as they meet the function of dynamic closed-loop regulation based on the target user's real-time physiological and neural state. This embodiment achieves closed-loop regulation of training parameters based on the aforementioned multi-dimensional static features through the adaptive adjustment module, overcoming the shortcomings of traditional training processes that are mechanized and lack dynamic regulation capabilities. It ensures precise control of training dosage and intensity, and realizes personalized adaptation and real-time dynamic adjustment of motor imagery training.

[0045] Through the above scheme, this embodiment establishes a progressive collaborative logic of "rapid video screening - refined static feature extraction - adaptive closed-loop control". The rapid processing module solves the problem of validating video stimuli with low latency, the refined processing module solves the problem of complete representation of neuromuscular function with high precision, and the adaptive adjustment module solves the problem of dynamic adaptation of training parameters based on accurate representation. The three modules work in layers to form a defense depth, which not only avoids wasting time on ineffective stimuli, but also ensures the accuracy and safety of the training process, thereby improving the overall effect of personalized training.

[0046] like Figure 2 As shown, in some embodiments, the fast processing module further includes a feature extraction unit, an activation probability calculation unit, and a target training video determination unit. Specifically, the feature extraction unit extracts time-frequency feature data and spatial projection data, further determines whether the user category is a first-category user or a second-category user, selects the corresponding type of filter based on the user category (e.g., a first-category user filter or a second-category user filter), determines the filter matrix corresponding to the filter category, updates the filter matrix every preset time interval, and calculates the activation probability to determine the target training video. Specific embodiments are described below.

[0047] The feature extraction unit is used to acquire the EEG signal corresponding to the target user when watching the training video, extract time-frequency domain features from the EEG signal according to a preset time period to obtain time-frequency feature data, and extract spatial pattern features from the time-frequency feature data to obtain spatial projection data.

[0048] Specifically, the feature extraction unit's processing flow is divided into two progressive stages. The first stage is time-frequency domain feature extraction. The system performs wavelet transform or bandpass filtering on the preprocessed EEG signal according to a preset time period (e.g., a time window of 1 to 2 seconds, with a 50% overlap rate that can be set to ensure temporal continuity). It focuses on the event-related desynchronization and synchronization phenomena in the μ band (8-13Hz) and β band (13-30Hz), thereby generating a time-frequency dual-dimensional power distribution matrix, i.e., the time-frequency feature data. This stage aims to extract frequency domain energy change features directly related to motion observation or imagination from the original time series. The second stage is spatial pattern feature extraction. Based on the time-frequency feature data, the system uses a filter bank common spatial pattern algorithm to decompose the signal into multiple frequency sub-bands and calculates the spatial projection vector that maximizes the variance difference in each sub-band, thereby highlighting the spatial distribution differences of the left and right motor cortexes (e.g., C3 and C4 electrodes) under different video stimuli, and finally outputs spatial projection data. It should be understood that although this embodiment lists wavelet transform and FB-CSP algorithm as specific extraction methods, other time-frequency-space joint analysis strategies such as short-time Fourier transform combined with common spatial patterns (CSP) can also be used in other embodiments, as long as the function of extracting dynamic features from both time-frequency and spatial dimensions is satisfied. This embodiment, through a dual extraction mechanism of time-frequency domain and spatial pattern, enhances the discriminative power of features through spatial pattern projection under the premise of low latency, effectively capturing the degree of brain region response induced by action observation or imagination, and providing highly discriminative input for subsequent activation probability calculation.

[0049] The fast processing module also includes a filter matrix determination unit, which is used to determine the filter matrix corresponding to the category of the target user attribute in the spatial mode, project the EEG signal corresponding to the target user when watching the training video onto the filter matrix to obtain the projection signal, calculate the covariance matrix using the projection signal, and normalize the variance of the projection signal according to the covariance matrix to obtain the normalized feature data corresponding to the spatial projection data.

[0050] Specifically, the filter matrix determination unit is the core execution component for spatial pattern feature extraction. It first loads a matching filter matrix W based on the target user's attribute category, then projects the filtered EEG signal matrix X onto this matrix to obtain the projected signal Z (i.e., ...). This projection process, by highlighting key signal components and suppressing noise, helps extract features with higher discriminative power for specific brain activities. Subsequently, the system calculates the variance Var(Z) of the projected signal Z to reflect the intensity of neural activity in a specific frequency band, and further calculates the covariance matrix. To eliminate the influence of individual differences and uneven energy distribution across frequency bands, the system normalizes the variance of the projected signal based on the covariance matrix, forming a feature coefficient vector v (e.g., the normalized weighting coefficients of the i-th frequency band).

[0051]

[0052] This vector is the normalized feature data. In this embodiment, through covariance calculation and normalization, the energy differences of different frequency bands and individuals are mapped to a unified scale space, making the feature data input to the subsequent classifier more robust and comparable, and effectively preventing the classification boundary drift problem caused by differences in the amplitude of individual EEG signals.

[0053] The activation probability calculation unit inputs the normalized feature data into the machine learning classifier to obtain the activation probability.

[0054] Specifically, the activation probability calculation unit receives the normalized feature data vector v output by the filter matrix determination unit and inputs it into a machine learning classifier. This machine learning classifier can be a lightweight model such as a radial basis function kernel-based support vector machine (RBF-SVM) or random forest (RF). The classifier calculates a decision value for the current feature vector based on pre-trained parameters and maps the decision value to a posterior probability, i.e., the activation probability P(y=1|x), using a Platt-calibrated logistic function. This activation probability represents the confidence score that the EEG data within the current time window belongs to the category of "mirror neuron activation." Its physical meaning is not a direct physiological quantity, but rather a probabilistic judgment based on features and training experience. This embodiment, by inputting normalized feature data into a lightweight machine learning classifier, achieves rapid probabilistic evaluation of the activation state of mirror neurons while ensuring low-latency computation, providing a quantitative basis for real-time screening of training videos.

[0055] See also Figure 2 The target training video determination unit determines the activation state of the mirror neuron based on the activation probability. When the neuron is in an activated state, it determines the current training video as the target training video. When the neuron is in an inactive state, it switches to the next training video after the current training video finishes playing and continues to calculate the activation probability based on the next training video until it is determined to be in an activated state based on the activation probability, at which point it determines the latest training video as the target training video.

[0056] Specifically, the target training video determination unit employs a cyclical filtering mechanism, the core logic of which is "switch if inactive, lock if active." The system sets the activation probability judgment threshold range to 0.65 to 0.7, and combines this with consistency judgment over 2 to 3 consecutive time windows to improve robustness and prevent false triggering due to instantaneous noise. For example, when the system plays the first fist-clenching action video, if the activation probability calculated by the fast processing module for two consecutive windows is lower than 0.65, the mirror neuron is determined to be inactive, and the system will automatically switch to the next training video (such as a finger-extending action video) after the current video finishes playing; if the activation probability for three consecutive windows is higher than 0.7 when playing the finger-extending action video, the mirror neuron is determined to be active, the system immediately locks the finger-extending action video as the target training video, and automatically triggers the subsequent fine-tuning module. It should be understood that although this embodiment illustrates a threshold range of 0.65 to 0.7 and a consistency requirement of 2 to 3 windows, in practical applications, these parameters can be dynamically adjusted according to the severity of the target user's condition or the signal quality. For example, for target users with high signal noise, the threshold can be appropriately increased to 0.75 or the continuous window requirement can be increased to 4 to further reduce the false positive rate. This embodiment, through this cyclical screening and consistency judgment mechanism, ensures that the finally selected target training video is the first video that can effectively activate mirror neurons, avoiding the frustration and wasted time caused by the target user watching ineffective stimuli for a long time, and significantly improving the effectiveness and personalized adaptation of the training stimulus.

[0057] See also Figure 2 In some embodiments, the feature extraction unit further includes a first category user filter selection subunit and a second category user filter selection subunit.

[0058] The first category user filter selection subunit is used to load a first initial category filter matrix in the current time window and a current updated filter matrix in subsequent time windows in response to the target user being a first category user. The first initial category filter matrix is ​​determined based on the difference in the covariance matrix of the target user's resting period signal and the stimulation period signal, and the resting period signal is the EEG signal of the target user in a task-free relaxation state, and the stimulation period signal is the EEG signal before the guiding stimulus induces the target user to enter the activation state.

[0059] Specifically, the "first category user" referred to in this invention refers to a user for whom there is a lack of historical prior EEG data or calibration samples in the current database of the system. A typical scenario is, for example, an early-stage stroke hemiplegic target user seeking medical attention for the first time. For such zero-sample users, the traditional approach is often to directly load a general filter matrix pre-trained based on a public dataset, or to use a random initialization strategy. However, random initialization, lacking any physiological constraints, is prone to causing a significant deviation between the initial projection space and the user's actual EEG distribution, resulting in a large number of misjudgments in the early stages of training. While directly loading a general matrix has a certain statistical basis for the population, due to the huge differences in individual EEG topology, the general matrix often fails to effectively highlight the spatial differences in the motor cortex of a specific user, resulting in insufficient feature discrimination. Therefore, this embodiment innovatively adopts an unsupervised computation strategy based on the difference in the covariance matrix between the user's resting period and the stimulation period to determine the first initial category filter matrix. Its physical meaning is that the resting period signal reflects the baseline of background neural activity of the user in a relaxed state without tasks, while the stimulation period signal reflects the transitional state of neural activity before the user enters the activation state under the induction of guided stimulation (such as playing a preparatory action video). The difference in the covariance matrices of these two components essentially characterizes the spatial energy distribution pattern of the user's brain during the transition from resting to motor preparation. Extracting this difference to construct the initial filter matrix is ​​equivalent to using the user's own intrinsic neural state transitions as a supervisory signal, achieving zero-sample cold start, even without any labeled data. This initialization method not only avoids the blindness of random initialization but also overcomes the problem of poor individual adaptability of general matrices, ensuring that the system can provide a feature projection space with basic discriminative capabilities for the first category of users at the beginning of training, significantly reducing the probability of accidental video switching in the early stages.

[0060] The second category user filter selection subunit is used to load the second initial category filter matrix in the current time window and the current update filter matrix in subsequent time windows in response to the target user being a second category user; wherein the second initial category filter matrix is ​​preset.

[0061] Specifically, the "second category user" referred to in this invention refers to users whose historical EEG data has been stored in the system or who have completed preliminary calibration training. A typical scenario is a historical rehabilitation target user entering the follow-up visit stage. For this type of user, the system database already stores the optimized and updated filter matrix parameters from the end of the previous training stage, or stores a dedicated matrix pre-trained based on their historical labeled data. Therefore, the second category user filter selection subunit directly loads the pre-set second initial category filter matrix in the current time window, without needing to perform unsupervised calculations during the resting and stimulation periods. The advantage of this strategy is that it fully utilizes historical prior information, saves the time overhead of initialization calculations, and allows the follow-up visit target user to immediately enter an efficient training state, achieving parameter inheritance and seamless connection between different visit stages. It should be understood that although this embodiment lists first-time visit target users and historical rehabilitation target users as typical scenarios, in practical applications, the criteria for classifying the first and second categories of users can be extended to other dimensions, such as dynamic determination based on signal quality scores or device wearing time, as long as the logic of differentiated selection of initialization strategies based on the presence or absence of prior data is satisfied.

[0062] See also Figure 2 The fast processing module also includes a filter matrix update unit, which calculates the covariance matrix of the current time window based on the collected target user's current resting period signal and stimulation period signal, combines the covariance matrix of the current time window with the covariance matrix of the previous time window in a recursive manner to obtain the updated covariance matrix, adjusts the filter weights based on the updated covariance matrix, and determines the current updated filter matrix.

[0063] Specifically, regardless of whether it's a first-category or second-category user, after loading the filter matrix for the initial time window, as training continues, the user's EEG signals often undergo two significant changes: first, a slow drift in individual neural activity patterns due to fatigue, attention fluctuations, or learning effects; and second, a signal distribution shift introduced by slight changes in electrode impedance or environmental noise. If the filter matrix remains statically fixed throughout the training cycle, its projection space will gradually mismatch with the user's real-time EEG distribution, leading to classification boundary drift and a continuous decrease in the accuracy of activation probability calculation. To address this issue, the filter matrix update unit employs a recursive strategy to periodically and smoothly update the covariance matrix. The system automatically collects resting and stimulation signals within the current time window every preset time interval (e.g., 15 minutes) and calculates the covariance matrix for the current time window. :

[0064]

[0065] in, It is the covariance calculated for the current time window, which is then compared with the previous covariance. Combined recursively, It is a weight constant used for smooth updates. Specifically, by introducing a smoothing weight constant... The system recursively merges the covariance matrix of the current time window with that of the previous time window. This smoothing weight constant is a small value, its physical meaning being to control the rate at which new observation data corrects the historical statistical model. If this constant is set too large, the model will excessively chase instantaneous fluctuations, leading to unstable filter weight oscillations; if it is set too small, the model's adaptability to drift will be too sluggish. Through recursive combination of formulas, the updated covariance matrix retains the stable statistical features accumulated during the historical training phase while appropriately incorporating the latest neural activity information of the current time window, achieving a smooth transition of filter weights. Finally, the system readjusts the filter weights based on this updated covariance matrix to determine the current updated filter matrix for feature extraction in the next time window. This embodiment constructs an anti-drift defense through this recursive adaptive update mechanism, enabling the spatial filter to dynamically track the slow evolution of the user's brainwave patterns. While maintaining real-time computing performance, it continuously improves the matching degree between the filter and the individual's real-time brainwave features, effectively overcoming the problem of classification accuracy decay caused by individual differences and brainwave drift during long-term training, and ensuring the long-term stability and reliability of the motor imagery training closed loop.

[0066] like Figure 3 As shown, in some embodiments, the fine processing module uses three parallel branches to calculate activation strength, connectivity strength and coupling strength respectively. These three types of features construct a defense depth from three complementary dimensions: nonlinear activation, cross-brain region functional connectivity and central-peripheral neural synergy, making up for the inadequacy of the representation ability of a single feature.

[0067] See also Figure 3 The activation intensity calculation unit is used to calculate the activation intensity of the motor cortex of the brain. Specifically, the activation intensity calculation unit extracts bispectral feature data of the μ rhythm from the electroencephalogram (EEG) signal to obtain the activation intensity of the motor cortex. Specifically, the activation intensity calculation unit first applies a bandpass filter to obtain the 8-13Hz μ rhythm signal from leads C3 / C4. Then, it calculates the bispectral... :

[0068]

[0069] Where x(n) is the signal sequence, representing the signal value sampled at the nth point, and N is the total number of sampling points of the signal. f1 and f2 represent conjugation and are frequencies. The calculated bispectrum is a two-dimensional frequency plot used to observe the cooperative oscillation characteristics between these frequencies. It should be understood that although this embodiment uses bispectral analysis to extract nonlinear interaction characteristics, traditional motion imagery analysis often relies only on shallow features such as the power spectrum to characterize activation intensity. However, the power spectrum only reflects the energy distribution of the signal and cannot capture the non-Gaussianity and phase coupling relationship of neural oscillations. Neural firing during motion imagery often has strong nonlinear coupling characteristics, which are masked by averaging in the power spectrum. By increasing the order, bispectral analysis can accurately identify specific activation patterns during motion imagery and capture the secondary phase coupling information that the power spectrum cannot reveal. Therefore, choosing the bispectral features of the μ rhythm instead of the power spectrum is to establish a more complete representational defense in the nonlinear dimension and prevent misjudgment of activation states due to the omission of shallow features.

[0070] Furthermore, the activation intensity calculation unit is used to extract signals within a preset frequency band in the motor cortex of the brain, calculate bispectral peaks, and determine the activation intensity of the motor cortex based on the positive correlation between the bispectral peaks and the activation intensity of the motor cortex. Specifically, the magnitude of the bispectral peaks directly reflects the strength of nonlinear coupling in the μ-rhythm signal, and it has a clear positive correlation with the activation intensity of the motor cortex. When the bispectral peaks are high, it indicates that the neural oscillation coordination of the motor cortex is strong and the activation intensity is large; conversely, the activation intensity is weak. In addition, this embodiment also establishes a mapping defense mechanism based on bispectral frequency distribution: when bispectral energy is detected to be concentrated in the high-frequency band (e.g., 11-13Hz), the system determines that the current task has a large cognitive and motor load on the target user. This high-frequency energy concentration is often accompanied by higher task complexity requirements. Based on this, the system can adaptively increase the complexity level of subsequent virtual reality tasks to match the high arousal state of the target user. It should be understood that this positive correlation mapping relationship and its high-frequency band mapping logic are only explanatory and not restrictive. In other embodiments, different frequency band mapping thresholds can be set according to different pathological populations, as long as the function of bispectral peaks can non-linearly characterize activation intensity and guide task difficulty adjustment is satisfied.

[0071] See also Figure 3The connectivity strength calculation unit is used to calculate the connectivity strength between cortical functions. Specifically, the connectivity strength calculation unit filters the EEG signals, extracts the inferior parietal lobule signal and the primary motor cortex signal respectively, and calculates the phase lock value between the inferior parietal lobule signal and the primary motor cortex signal to obtain the connectivity strength between cortical functions. Specifically, the connectivity strength calculation unit uses a bandpass filter on the EEG signals of the inferior parietal lobule (IPL area) and the primary motor cortex (M1 area) to obtain the 15-25Hz β-band signal and the 30-45Hz γ-band signal respectively. The Hilbert transform is then applied to the filtered signals to obtain the instantaneous phase. :

[0072]

[0073] Where ŝ(t) is the Hilbert transform of the signal s(t), yes = , is the imaginary unit. The phase difference is then calculated:

[0074]

[0075] Finally, calculate the phase lock value (PLV):

[0076]

[0077] Where N is the number of data points within the time window. The unit complex phase factor representing the phase difference should be understood. Traditional methods often use coherence metrics when assessing the efficiency of cross-brain region information transmission. However, coherence not only includes phase synchronization information but is also strongly influenced by amplitude coupling. When the signal amplitude in a brain region increases instantaneously due to noise or muscle artifacts, coherence can be artificially inflated, leading to misjudgments of connection strength. PLV, on the other hand, only calculates the average modulus of the instantaneous phase difference, completely eliminating amplitude interference. Its value ranges from 0 to 1, where 1 represents complete phase synchronization and 0 represents no phase synchronization. Therefore, choosing PLV over coherence is to establish a pure phase defense against amplitude interference across brain regions, ensuring that the assessment of connection strength truly reflects the steady-state synchronicity of neural information transmission rather than instantaneous energy fluctuations.

[0078] Furthermore, the connectivity strength calculation unit is used to acquire signals within a preset frequency band in the inferior parietal lobule and signals within a preset frequency band in the primary motor cortex, respectively, and calculate the corresponding instantaneous phase. Based on the instantaneous phase, the phase lock value between the inferior parietal lobule and the primary motor cortex is calculated. The connectivity strength between cortical functions is determined according to the positive correlation between the phase lock value and the connectivity strength between cortical functions. Specifically, the PLV value has a clear positive correlation with the connectivity strength between cortical functions. The higher the PLV, the higher the efficiency of cross-brain region information transmission and the greater the connectivity strength. In order to provide accurate quantitative input for the subsequent adaptive adjustment module, this embodiment sets a key endpoint defense threshold for PLV: when PLV is less than 0.3, it indicates that cross-brain region information transmission is severely blocked and the transmission of the target user's motor intention is extremely weak. At this time, the system determines that it needs to provide maximum assistance to help the target user complete the action to the greatest extent. When PLV is greater than 0.7, it indicates that cross-brain region information transmission is efficient and smooth and the target user's own motor network connection is sound. At this time, the system determines that it only needs to provide minimum assistance to avoid excessive assistance that deprives the target user of their sense of active participation. These two threshold values ​​(0.3 and 0.7) are critical values ​​derived from a large amount of clinical motor imagery data, which can accurately distinguish different degrees of neurological deficits. It should be understood that although this embodiment lists 0.3 and 0.7 as endpoints, in practical applications, these endpoint values ​​can be fine-tuned for target users with different disease stages or training phases, as long as the logic of PLV being inversely proportional to the amount of assistance is satisfied.

[0079] See also Figure 3 The coupling strength calculation unit is used to calculate the coupling strength between nerves and muscles. Specifically, the coupling strength calculation unit filters the EEG signal, calculates the time-delay cross-correlation value between the EEG signal and the corresponding EMG signal, and obtains the coupling strength between nerves and muscles. Specifically, the coupling strength calculation unit acquires EMG signals from the target muscle group and performs β-band desynchronization event detection on the EEG signals. Then, it calculates the time-delay cross-correlation. :

[0080]

[0081] Where x(n) is the nth baseline EEG signal obtained from leads C3 / C4, y(n) is the nth contrast EMG signal, τ is the time delay, and N is the total number of sampling points. In this embodiment, a time delay window of ±50ms is chosen. This window setting has clear physiological and physical significance: the physiological conduction time from the central motor command to the peripheral muscles to induce contraction is typically between 30 and 50 milliseconds. Limiting the time delay window to ±50ms can both cover normal physiological conduction delays and effectively exclude noise interference or non-causal spurious correlations that exceed the physiological range. By searching for the maximum cross-correlation value within the ±50ms window, the system can quantitatively assess the coupling strength between the central and peripheral nerves and verify the matching between action execution and motor intention. This central-peripheral dimension feature extraction compensates for the deficiency of simple EEG analysis in confirming whether intention is effectively converted into action.

[0082] Furthermore, the coupling strength calculation unit is used to extract signals within a preset frequency band range from the motor cortex of the brain, calculate the time-delay cross-correlation value between the signals and the corresponding electromyographic signals within a time-delay window, and determine the coupling strength between the nerve and muscle based on the positive correlation between the time-delay cross-correlation value and the coupling strength between the nerve and muscle. Specifically, the peak value of the time-delay cross-correlation value has a clear positive correlation with the coupling strength between the nerve and muscle. The higher the cross-correlation value, the more aligned the motor command issued by the central nervous system with the response of the peripheral muscles in time and energy, and the stronger the coupling strength; conversely, it indicates that the central intention and the peripheral execution are disconnected, and the coupling strength is weak. It should be understood that although this embodiment preferably uses ±50ms as the time-delay window, in other embodiments, for different limb parts (such as the conduction delay of leg muscles may be slightly longer than that of the hand), this window can be adaptively extended according to the specific length of the neural conduction pathway, for example, extended to ±80ms, as long as it meets the function of covering the physiological conduction delay range of the target muscle.

[0083] Through detailed algorithm binding and quantization mapping of the three parallel branches, this embodiment transforms abstract neurophysiological concepts into quantifiable signal processing indicators. Bispectral peaks capture the oscillatory coordination within the cortex from a nonlinear perspective, PLV captures information transmission between networks from a cross-brain region perspective, and time-delay cross-correlation values ​​capture the spatiotemporal matching of instructions and execution from a central-peripheral perspective. These three types of features are not calculated in isolation but form a progressive association with the features of the fast processing module: the time-frequency phase coupling feature directly reuses the time-frequency decomposition results of the fast module, requiring only the addition of phase difference calculation; the higher-order spectral feature is based on the μ-rhythm filtered signal of the fast module, enhancing nonlinear discrimination capability by increasing the order; and the EMG-EEG time-delay cointegration feature jointly models the independent EEG and EMG indicators in the fast module, establishing a quantitative association between neuromuscular coupling. This three-dimensional coupling mechanism constructs a more complete neuromuscular functional representation system, providing accurate and robust multidimensional input parameters for the subsequent adaptive adjustment module.

[0084] like Figure 4 As shown, in some embodiments, the adaptive adjustment module includes a motion intention decision unit, a torque control unit, a difficulty control unit, and a sensitivity control unit.

[0085] Specifically, a hybrid architecture model is constructed through an adaptive adjustment module to output motion intention decisions, and further adjustments are made based on different data in the motion intention decisions. For example, the exoskeleton torque is adjusted based on the confidence level of the output motion intention; the task difficulty is adjusted based on the output motion pattern; and the neural feedback sensitivity or protection mechanism is adjusted based on the coefficient of variation of the latency-related value in system stability. See below for specific implementation examples.

[0086] The motion intention decision unit is used to construct a hybrid architecture model, which captures the nonlinear characteristics of the bispectral peaks, extracts the dynamic change characteristics of the phase-locked values, analyzes the coordination patterns of nerves and muscles, and outputs motion intention decisions. The motion intention decisions include motion intention confidence, motion patterns, and system stability.

[0087] Specifically, the motion intention decision unit is the core processing center of the adaptive adjustment module, employing a hybrid architecture model combining LightGBM and CNN. This hybrid architecture is not a simple model stacking, but rather a defensive triage process based on the fundamental differences in the neurophysiological representations of static and dynamic features. Static features (such as bispectral peaks, phase-locked values, and time-delay cross-correlation values) reflect the steady-state average properties of neuromuscular function over a longer time window. Their data form is typically a low-dimensional statistical vector, suitable for analysis using tree models. Therefore, this embodiment utilizes LightGBM to handle static features. To prevent the model from overfitting noise when the amount of individual data is limited, the depth of the decision tree is specifically limited to no more than 5 layers. This constraint acts as a regularization defense for the model, ensuring that it extracts coarse-grained classification rules with group generalization ability, rather than rote memorization of individual abnormal fluctuations. In contrast, dynamic time-series data (such as continuous EEG time series) contains instantaneous weak changes and rapid transition patterns in neural signals; these temporal dependencies cannot be captured by tree models. Therefore, this embodiment utilizes CNN to process dynamic temporal sequences, and specifically designs a one-dimensional temporal convolutional network with a kernel width set to 500 milliseconds. This parameter setting has a clear neurophysiological basis: typical neural micro-events (such as segments of changes in motor-related cortical potentials) in the motor cortex from the generation of intent to the issuance of execution instructions usually last for hundreds of milliseconds. A kernel width of 500 milliseconds can just cover a complete neural firing cycle, thereby effectively capturing rapid change patterns and feature representations in the signal. The static classification logic output by LightGBM and the dynamic feature representation output by CNN are fused together with attention weighting in the fusion layer, and the final output is a multi-dimensional decision vector containing the confidence of motion intent, motion pattern, and system stability. It should be understood that although this embodiment preferably uses a hybrid architecture of LightGBM and CNN, in other embodiments, a combination of random forest and long short-term memory network (LSTM) can also be used, as long as the logic of separate processing and fusion decision of static features and dynamic temporal features is satisfied.

[0088] The motion intention decision-making unit also includes a motion intention classification subunit and a motion intention activation recognition subunit.

[0089] The motion intent classification subunit is used to identify motion intents using the hybrid architecture model and the static features, and to classify each type of motion intent.

[0090] The motion intent activation recognition subunit is used to identify whether the corresponding category of motion intent is in an active state using the hybrid architecture model and the dynamic features. If it is in an active state, the training action corresponding to the motion intent category is executed. If it is in an inactive state, the training action is not executed this time.

[0091] Specifically, this embodiment establishes a hierarchical decision-making defense line for determining motion intentions, which involves classification followed by confirmation. Traditional single-layer decision-making often directly inputs all features into a black-box model and outputs a binary classification result (whether the action occurred or not). This approach is prone to misjudgment when faced with similar motion intentions (such as clenching a fist and extending fingers). In this embodiment, the hierarchical decision-making first uses a motion intention classification subunit to perform coarse-grained classification of motion intentions based on static features processed by LightGBM, determining which type of action the target user is currently attempting to perform (e.g., classified as a clenching fist intention). Subsequently, the motion intention activation recognition subunit uses dynamic features processed by CNN to perform fine-grained activation state confirmation for this specific category of intention, determining whether the clenching fist intention has truly entered the neural firing activation stage from the imagination stage. This decoupled design in terms of temporal logic is equivalent to determining the direction before confirming the force, effectively avoiding the risk of misjudging weak imagined fluctuations as valid execution commands. The system only executes the training action corresponding to the category when the activation recognition subunit confirms that it is in an active state; if it is determined to be in an inactive state, the training action will not be executed this time, and the system will enter a waiting or re-evaluation state, thereby ensuring the reliability and safety of brain-computer interface control and preventing exoskeleton mis-driving due to misjudgment.

[0092] A torque control unit is used to adjust the exoskeleton torque based on the motion intention confidence level and the phase lock value in the motion intention decision; wherein the phase lock value is inversely proportional to the exoskeleton torque.

[0093] Specifically, the torque control unit corresponds to the proportional term (P term) in the fuzzy PID control strategy. Its core logic is to provide appropriate support based on the target user's current motor ability. The phase lock value (PLV) reflects the efficiency of cross-brain region information transmission and is a key indicator for measuring the integrity of the target user's own motor network connections; the confidence level of motor intention reflects the system's degree of certainty in judging the current intention. This embodiment establishes a clear decoupling mapping defense line: the phase lock value is inversely proportional to the exoskeleton torque. Its physiological and physical significance is that when the PLV is high (e.g., greater than 0.7), it indicates that the motor command transmission within the target user's brain is smooth and that the user has strong active motor ability. At this time, the system should reduce the exoskeleton torque and provide minimum assistance to avoid excessive assistance that deprives the target user of their sense of active participation and opportunities for neuroplasticity training. Conversely, when the PLV is extremely low (e.g., less than 0.3), it indicates that cross-brain region information transmission is severely blocked, and the target user cannot effectively mobilize the motor network. At this time, the system must significantly increase the P term and provide maximum assistance to compensate for the lack of neural transmission with physical external force and help the target user complete the action loop. For example, in the specific setting of fuzzy logic rules, "if PLV is less than 0.3 and the confidence level is low, then significantly increase P to strengthen support" is a typical defensive rule, ensuring the strongest protection in the weakest state; "if PLV is between 0.3 and 0.6 and the confidence level is moderate, then moderately increase P"; "if PLV is higher than 0.6 and the confidence level is high, then decrease P to reduce support." It should be understood that although this embodiment lists specific PLV thresholds and fuzzy rules, in practical applications, the gain coefficient of torque control can be fine-tuned according to the mechanical characteristics of the exoskeleton, as long as the core mapping relationship of inversely proportional control between PLV and torque is satisfied.

[0094] The difficulty adjustment unit is used to adjust the difficulty of the virtual reality task based on the bispectral feature data of the μ rhythm and the classification results of the motion pattern.

[0095] Specifically, the difficulty adjustment unit corresponds to the integral term (I term) in the fuzzy PID control strategy. Its core logic is to adjust the training intensity in a gradient manner, gradually changing the difficulty of the virtual reality task to match the learning and adaptation capabilities of the target user. The role of the integral term is to accumulate the trend of training performance changes, avoiding drastic changes in difficulty caused by a single fluctuation. In this embodiment, the bispectral feature data of the μ rhythm and the classification results of the motion pattern are used as the input basis for adjusting the I term. The bispectral feature data not only reflects the activation intensity, but its frequency distribution also implies cognitive load information: when the bispectral energy is detected to be concentrated in the high frequency band, it often means that the target user is currently in a high arousal or high load state. At this time, the system should automatically increase the complexity of the task (e.g., increase the number of objects to be grasped in the virtual reality scene or reduce the size of the target) to match the target user's neural resource reserves; the classification results of the motion pattern guide the direction of difficulty adjustment. For example, for the mode classified as fine motor skills (such as pinching), its difficulty gradient should be set more gently than that of gross motor skills (such as pushing and pulling). Fuzzy logic rules include, for example: "If the cumulative error rate increases and the bispectral energy is dispersed, then decrease I to reduce the task difficulty"; "If the completion time gradually shortens and the bispectral energy is concentrated in the high-frequency band, then moderately increase I to increase the difficulty." This embodiment ensures that the target user trains under a moderate challenge through this gradient adjustment based on neural load and behavioral performance, so that they do not lose interest due to too low a difficulty level, nor feel frustrated due to too high a difficulty level.

[0096] A sensitivity adjustment unit is used to adjust the neural feedback sensitivity based on the coefficient of variation of the time delay cross-correlation value; wherein the neural feedback sensitivity is proportional to the signal interference degree.

[0097] Specifically, the sensitivity control unit corresponds to the derivative term (D term) in the fuzzy PID control strategy. Its core logic is to adjust the sensitivity of the feedback system to suppress abnormal signal interference and ensure the stability of the training closed loop. The derivative term is sensitive to the rate of change of the signal, enabling it to predict trends in advance and quickly suppress interference. This embodiment innovatively uses the coefficient of variation of the time delay cross-correlation value as the core input parameter for adjusting the D term. The coefficient of variation of the time delay cross-correlation value reflects the temporal stability of the central-peripheral nerve coupling. The larger the coefficient of variation, the more unstable the timing of nerve-muscle coordination, and the more abnormal interference or physiological tremor components in the signal. The mapping defense line established in this embodiment is: the neural feedback sensitivity is proportional to the signal interference degree, while the signal interference degree is positively correlated with the coefficient of variation of the time delay. When the coefficient of variation increases, the system determines that the interference degree is high. At this time, increasing the D term and improving the feedback sensitivity means that the system will respond and suppress weak abnormal fluctuations more quickly, preventing malfunctions caused by interference. When the coefficient of variation is small, the signal is stable, and there is little interference, decreasing the D term avoids the system from over-responding to the normal mechanical vibration or slight physiological fluctuations of the exoskeleton. For example, the fuzzy logic rules are set as follows: "If an increase in the time delay variation coefficient or a decrease in the signal-to-noise ratio is detected, increase D to improve feedback sensitivity and quickly suppress interference"; "If the signal is stable and interference is low, decrease D to avoid over-response." It should be understood that although this embodiment preferably uses the time delay variation coefficient as a quantitative indicator of interference, in other embodiments, it can also be combined with traditional indicators such as the signal-to-noise ratio for joint judgment, as long as the logic that sensitivity is proportional to interference is satisfied.

[0098] Through the above-mentioned hybrid architecture's diversion decision and fuzzy PID three-level control, this embodiment decouples and maps abstract neurophysiological indicators into physical control parameters of three dimensions: torque, difficulty, and sensitivity. This achieves closed-loop control of humanoid intelligence, ensuring the safety of training while improving the accuracy of personalized adaptation.

[0099] In some embodiments, the adaptive adjustment module further includes a protection mechanism unit, which is used to activate the protection mechanism and reduce the training intensity when the coefficient of variation of the time-delay cross-correlation value is abnormal.

[0100] Specifically, the protection mechanism unit is the final hard safety barrier in the entire closed loop of motor imagery training. During training, the target user may experience sudden muscle tremors, electrode loosening or detachment, or a sudden increase in environmental electromagnetic interference, which may cause the originally stable central-peripheral neural coupling to collapse instantly. The direct manifestation of this collapse at the signal level is an abnormal spike in the coefficient of variation of the time delay cross-correlation value. The "abnormal coefficient of variation" referred to in this invention preferably refers to the threshold value of the coefficient of variation of the time delay cross-correlation value exceeding 25%. The setting of this threshold has a solid clinical statistical basis: in normal steady-state motor imagery or action execution, the time delay jitter of the transmission of central commands to the periphery is usually controlled within a very small physiological range, and its coefficient of variation is generally less than 15%; when the coefficient of variation exceeds 25%, it means that the temporal consistency of neuromuscular coordination has been severely decoupled, and the abnormal interference components or physiological tremor components in the signal have become dominant. If the original intensity of training is maintained at this time, it is very easy to cause the exoskeleton to misdrive, resulting in joint sprains in the target user, or cause the target user to experience severe motor frustration in virtual reality tasks. Therefore, when the system detects that the coefficient of variation is greater than 25%, the protection mechanism unit will be immediately triggered, forcibly intervened and the training intensity will be reduced. For example, the auxiliary torque output of the exoskeleton will be rapidly reduced to a safe baseline value, and the difficulty of the virtual reality task will be downgraded to the most basic guided mode until the coefficient of variation returns to the normal range.

[0101] It should be understood that although this embodiment preferably uses 25% as the threshold for determining abnormal coefficient of variation, this threshold is not absolutely fixed in practical applications. For target users with advanced Parkinson's disease whose condition is extremely unstable or who have severe muscle tremors, the system can appropriately relax the threshold to 30% to avoid frequent interruptions to training; while for target users in the early stage of fine rehabilitation who require extremely high stability, the threshold can be tightened to 20% to achieve more sensitive safety protection. The goal is simply to fulfill the function of dynamically triggering the protection mechanism based on the degree of abnormality in the coefficient of variation.

[0102] To more clearly illustrate the irreplaceable nature of this invention's choice of the coefficient of variation of the time delay cross-correlation value, rather than other indicators, as the protection triggering parameter, a set of comparative examples is provided below to explain the mechanism:

[0103] In traditional brain-computer interface (BCI) safety strategies, a decrease in the connectivity strength between cortical functions (i.e., phase-locked value, PLV) is often used as an indicator to trigger protection. However, a decrease in PLV presents a false danger scenario that can easily lead to misjudgment—normal physiological fatigue. When a target user experiences moderate fatigue after prolonged training, the efficiency of cross-brain region information transmission naturally decreases, causing the PLV value to slowly decline. This decline is a normal process of neural resource dissipation during motor imagery training and does not indicate that the target user is in a dangerous, out-of-control state. If the protection mechanism is mistakenly triggered to forcibly reduce the training intensity at this time, it will not only interrupt the normal fatigue adaptation and neural remodeling process but also confuse and frustrate the target user. Conversely, the coefficient of variation of the latency cross-correlation value can accurately distinguish between "fatigue" and "out of control." In a normal state of fatigue, although the PLV decreases, the transmission timing between central commands and peripheral muscles remains stable, with minimal latency jitter and a low coefficient of variation. Only when there is true coupling out of control or strong external interference will the transmission timing fluctuate violently, and the coefficient of variation will spike abnormally. Therefore, the coefficient of variation of time delay instead of PLV is chosen as the protection trigger indicator because it directly reflects the structural stability of the central-peripheral coupling in the time dimension. It can filter out normal fatigue decay signals and more accurately capture the critical state of abnormal interference and coupling loss of control, thus proving the irreplaceable role of this parameter in preventing false triggers and ensuring training continuity.

[0104] This embodiment, through this hard protection line based on the time delay coefficient of variation, forms a complementary safety architecture with the torque control (soft assistance) based on PLV mentioned above: PLV control solves the normal assistance problem of "insufficient capability," while the coefficient of variation protection solves the abnormal risk avoidance problem of "state out of control." The two work together to ensure both personalized adaptation of the training loop within normal fluctuation ranges and physical safety under extreme abnormal conditions, achieving a balance between the safety and effectiveness of the motion imagery training system.

[0105] This embodiment obtains dynamic features based on the time-frequency domain and spatial pattern processing of EEG signals before training through a rapid processing module. It then selects target training videos based on the activation probability of mirror neurons, solving the problem of traditional systems directly using predetermined video stimuli without verifying effectiveness. This avoids patients watching ineffective stimuli for extended periods and achieves personalized adaptation of training content. A fine processing module extracts multi-dimensional static features including the activation intensity of the motor cortex, the strength of intercortical functional connections, and the coupling strength between nerves and muscles, constructing a three-dimensional coupling mechanism of spatial-temporal-frequency domain features. This forms a more complete neuromuscular function representation system, compensating for the shortcomings of single-feature representation capabilities. An adaptive adjustment module, based on the aforementioned multi-dimensional static features, outputs precise movement intention decisions using a hybrid architecture model. A fuzzy PID algorithm is employed to separately control the exoskeleton auxiliary torque, virtual reality task difficulty, and neural feedback sensitivity, achieving parameter adaptation, rapid response, and online optimization. This overcomes the shortcomings of traditional training processes, which are mechanized and lack dynamic control capabilities, ensuring precise control of training dosage and intensity, improving the effectiveness of personalized training, and providing a new technological paradigm for the application of brain-computer interfaces in precision medicine.

[0106] Example 2:

[0107] To more clearly illustrate the personalized adaptation and dynamic safety control capabilities of the system of this invention, the complete operational loop of the above-mentioned modules is described in detail below using a clinical rehabilitation application scenario for target users with early-stage stroke and hemiplegia as an example. It should be understood that this embodiment is merely illustrative and not restrictive, intended to demonstrate the practical application of the technical solution in a real clinical environment, and does not constitute a limitation on medical advice.

[0108] The typical target user is defined as someone with early-stage stroke and hemiplegia, belonging to the first category of users, meaning those without prior historical EEG data or calibration samples in the system. Due to the early stage of the disease, these users have extremely weak voluntary motor function on the affected side and are highly susceptible to sudden muscle tremors or spasms during training. Therefore, extremely high demands are placed on the system's zero-sample cold start capability, low-latency video screening capability, and safety protection capabilities against abnormal states. For example... Figure 1 As shown, the specific steps are as follows:

[0109] S101, Signal Acquisition and System Cold Start. The target user wears a 64-channel EEG signal acquisition device and an 8-channel wireless surface electromyography (EMG) acquisition system, with electrodes primarily covering motor-related cortical areas such as C3 and C4, as well as the muscle belly of the target muscle on the affected side. Since the target user is a first-class user, the first-class user filter selection subunit of the fast processing module responds to this attribute by loading the first initial class filter matrix within the current time window. Specifically, the system automatically acquires the target user's resting period signal (unattended, relaxed state) and stimulation period signal (before the guided stimulus induces an activation state) for the first 30 to 60 seconds of video stimulation, and unsupervisedly calculates the first initial class filter matrix based on the difference in their covariance matrices. The physical significance of this cold start strategy lies in using the target user's own endogenous neural state differences during the transition from resting to motor preparation as a supervisory signal, avoiding the blindness of random initialization. This ensures that the system can provide the zero-sample target user with a feature projection space with basic discriminative capabilities from the initial training stage, significantly reducing the probability of video switching errors caused by filter mismatch in the early stages.

[0110] S102, Training Video Pre-screening. The system starts playing the first candidate training video, such as a fist-clenching video. The feature extraction unit of the fast processing module performs time-frequency domain and spatial pattern processing on the acquired EEG signals according to a preset time period to obtain dynamic features. The filter matrix determination unit projects the signal onto the first initial category filter matrix and performs normalization processing. The activation probability calculation unit inputs the normalized feature data into the machine learning classifier to obtain the activation probability of the mirror neurons. Since the visual stimulation of the fist-clenching action may not effectively awaken the damaged motor network of early hemiplegic target users, the calculated activation probability is less than 0.65. In response to being in an inactive state, the target training video determination unit will switch to the next training video after the current fist-clenching video finishes playing.

[0111] S103, Target Video Locking and Fine-tuning Module Triggering. The system automatically switches to the finger-extending motion video. The visual representation of the finger-extending motion is simpler and more direct, often more likely to induce resonance of mirror neurons in early target users. At this time, the continuous window activation probability calculated by the fast processing module is greater than 0.65, satisfying the consistency judgment logic. The target training video determination unit responds to being in an active state, determines that the current finger-extending video is used as the target training video, and automatically triggers the fine-tuning module. This cyclical screening mechanism ensures that target users do not watch ineffective fist-clenching stimuli for a long time, but immediately lock onto the truly effective finger-extending stimuli, significantly reducing frustration and wasted time.

[0112] S104, Fine-tuning Module 3D Static Feature Extraction. After confirming the target training video, the system enters the substantive training phase, and the fine-tuning module initiates 3D static feature extraction. The activation intensity calculation unit extracts signals in the 8-13Hz frequency band of the motor cortex, calculates the bispectral peak value, and determines the current cortical activation intensity based on the positive correlation between the bispectral peak value and activation intensity. The connectivity strength calculation unit acquires signals in the 15-25Hz frequency band of the inferior parietal lobule and the 30-45Hz frequency band of the primary motor cortex, calculates the instantaneous phase and phase lock value, and evaluates the efficiency of cross-brain region information transmission. The coupling strength calculation unit extracts signals in the motor cortex within a preset frequency band, calculates the time delay cross-correlation value between the signals and the electromyographic signals of the affected side within a ±50ms time delay window, and quantifies the matching of central commands to peripheral transmission. Because the target user is in an early hemiplegic state, their cross-brain region information transmission is severely blocked, and the calculated phase lock value is less than 0.3, indicating extremely weak connectivity strength.

[0113] S105, Hybrid Model Decision Making and Fuzzy PID High-Assist Control. The motion intention decision unit of the adaptive adjustment module constructs a hybrid architecture model, using LightGBM to process the aforementioned static features (limiting the depth to ≤5 layers to prevent overfitting) and CNN to process dynamic temporal sequences (with a kernel width of 500ms to capture neural micro-events), fusing the outputs to determine the motion intention. Due to the extremely low phase lock value and weak electromyographic response on the affected side, the hybrid model outputs a low-confidence motion intention. Based on this decision, the fuzzy PID controller activates a high-assistance mode: the torque control unit, based on the low confidence and extremely low phase lock value (less than 0.3), determines that the target user's own motor network connection is extremely weak, significantly increasing the proportional term P and adjusting the exoskeleton torque to the maximum assistance level, using physical external force to compensate for the lack of neural conduction; the difficulty control unit, based on the bispectral features and the low-confidence classification results, reduces the integral term I, downgrading the virtual reality task difficulty to the most basic guidance mode, avoiding anxiety caused by high-difficulty tasks; the sensitivity control unit, based on the coefficient of variation of the time delay cross-correlation value, moderately adjusts the differential term D to suppress the interference of unstable weak signals on the affected side. The mechanism of this regulatory logic is that, for target users with severe damage in the early stages, the system does not mechanically require them to complete actions, but rather constructs a safe motor closed loop through maximum physical compensation and minimum cognitive load, first ensuring the physical completion of actions, and then gradually reshaping endogenous control capabilities through neural plasticity.

[0114] S106, Emergency Intervention for Sudden Anomalies. After a period of training, the target user experiences a sudden muscle tremor on the affected side due to fatigue or pathological reasons. This tremor causes a momentary breakdown in the temporal consistency between central motor commands and peripheral muscle responses. The coupling strength calculation unit detects an abnormal spike in the coefficient of variation of the time-delay cross-correlation value, exceeding the preset endpoint threshold of 25%. Responding to this abnormal coefficient of variation, the protection mechanism unit immediately activates the protection mechanism and reduces the training intensity, rapidly lowering the exoskeleton's auxiliary torque to a safe baseline value, while simultaneously suspending the virtual reality task. It should be understood that if the system only uses a decrease in the phase-lock value as the protection trigger, it might misjudge normal fatigue-induced connection attenuation as a dangerous state, frequently interrupting training. The time-delay coefficient of variation accurately captures the temporal decoupling and abnormal interference caused by tremors, filtering out normal fatigue signals, thereby achieving precise risk avoidance in critical situations and protecting the target user from secondary injury.

[0115] S107, Recursive Adaptive Update of the Filter Matrix. After the protection mechanism is lifted and the target user's state stabilizes, training continues. Every 15 minutes, the filter matrix update unit calculates the covariance matrix of the current time window based on the collected resting and stimulation signals of the target user. This covariance matrix is ​​then recursively combined with the covariance matrix of the previous time window, incorporating a smoothing weight constant to determine the current updated filter matrix. This recursive mechanism allows the spatial filter to dynamically track the slow evolution of the target user's EEG pattern driven by the learning effect, overcoming individual differences and EEG drift during long-term training, and ensuring that the classification accuracy of subsequent video selection and intent determination does not decay over time.

[0116] Through the aforementioned complete clinical closed loop, this embodiment clearly demonstrates how the system of the present invention starts from a zero-sample cold start, rapidly pre-screens the most suitable finger-extending videos for the target user, accurately diagnoses the weak state of cross-brain region connectivity through fine feature extraction, provides highly auxiliary regulation with maximum physical compensation and minimum cognitive load through fuzzy PID, achieves hard safety protection by keenly capturing the time delay variation coefficient during sudden muscle tremors, and finally continuously adapts to the target user's neural remodeling process through recursive updates. This process fully confirms the defensive depth constructed by the present invention to address the problems of insufficient personalization and lack of dynamic regulation in motor imagery training, greatly enhancing the feasibility and commercial persuasiveness of the solution in real clinical rehabilitation scenarios.

[0117] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0118] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A brain-computer interface-based motor imagery training system, characterized in that, The system includes: A rapid processing module is used to acquire the EEG signals corresponding to the target user watching training videos, and to perform time-frequency domain and spatial pattern processing on the EEG signals to obtain dynamic features. The target training video is determined based on the activation probability of mirror neurons in the dynamic features. The rapid processing module includes: a feature extraction unit, used to acquire the EEG signals corresponding to the target user watching training videos, and to perform time-frequency domain feature extraction on the EEG signals according to a preset time period to obtain time-frequency feature data; and to perform spatial pattern feature extraction on the time-frequency feature data to obtain spatial projection data; an activation probability calculation unit, used to calculate the activation probability of mirror neurons based on the spatial projection data; and a target training video determination unit, used to determine the activation state of mirror neurons based on the activation probability, and to determine the target training video to be used based on the activation state. A fine processing module is used to acquire multimodal sensor signals corresponding to the target user watching the target training video, and determine the static features of the motor cortex of the brain based on the multimodal sensor signals. The static features include the activation intensity of the motor cortex of the brain determined by the electroencephalogram (EEG) signals, the connectivity strength between cortical functions, and the coupling strength between nerves and muscles determined by the combined EEG and electromyography (EMG) signals in the multimodal sensor signals. The fine processing module includes: an activation intensity calculation unit, used to extract bispectral feature data of μ rhythm from the EEG signals to obtain the activation intensity of the motor cortex of the brain; a connectivity strength calculation unit, used to filter the EEG signals, extract the inferior parietal lobule signal and the primary motor cortex signal respectively, calculate the phase lock value between the inferior parietal lobule signal and the primary motor cortex signal to obtain the connectivity strength between cortical functions; and a coupling strength calculation unit, used to filter the EEG signals, calculate the time delay cross-correlation value between the EEG signals and the corresponding EMG signals to obtain the coupling strength between nerves and muscles. An adaptive adjustment module is used to adaptively adjust at least one of the following based on the activation intensity of the motor cortex of the brain, the connectivity strength between cortical functions, and the coupling strength between nerves and muscles: exoskeleton torque, virtual reality task difficulty, and neural feedback sensitivity.

2. The system according to claim 1, characterized in that, The rapid processing module also includes: The filter matrix determination unit is used to determine the filter matrix corresponding to the category of the target user attribute in the spatial mode, project the EEG signal corresponding to the target user when watching the training video onto the filter matrix to obtain the projection signal, calculate the covariance matrix using the projection signal, and normalize the variance of the projection signal according to the covariance matrix to obtain the normalized feature data corresponding to the spatial projection data. The activation probability calculation unit inputs the normalized feature data into the machine learning classifier to obtain the activation probability. The target training video determination unit determines the activation state of the mirror neurons based on the activation probability. When the neurons are in an activated state, the unit determines the current training video as the target training video. When the neurons are in an inactive state, the unit switches to the next training video after the current training video finishes playing and continues to calculate the activation probability based on the next training video until the activation probability determines that the neurons are in an activated state, at which point the latest training video is determined as the target training video.

3. The system according to claim 2, characterized in that, The feature extraction unit further includes: The first category user filter selection subunit is used to load a first initial category filter matrix in the current time window and load a current updated filter matrix in subsequent time windows in response to the target user being a first category user; wherein, the first initial category filter matrix is ​​determined based on the difference in the covariance matrix of the collected target user's resting period signal and stimulation period signal, and the resting period signal is the EEG signal of the target user in a task-free relaxation state, and the stimulation period signal is the EEG signal before the guiding stimulus induces the target user to enter the activation state; The second category user filter selection subunit is used to load the second initial category filter matrix in the current time window and the current update filter matrix in subsequent time windows in response to the target user being a second category user; wherein the second initial category filter matrix is ​​preset.

4. The system according to claim 3, characterized in that, The rapid processing module also includes: The filter matrix update unit is used to calculate the covariance matrix of the current time window based on the collected target user's current resting period signal and stimulation period signal. The covariance matrix of the current time window is recursively combined with the covariance matrix of the previous time window to obtain the updated covariance matrix. Based on the updated covariance matrix, the filter weights are adjusted to determine the current updated filter matrix.

5. The system according to claim 1, characterized in that, In the fine processing module, The activation intensity calculation unit is specifically used to extract signals within a preset frequency band range in the motor cortex of the brain, calculate bispectral peaks, and determine the activation intensity of the motor cortex of the brain based on the positive correlation between the bispectral peaks and the activation intensity of the motor cortex of the brain. The connection strength calculation unit is specifically used to acquire signals within a preset frequency band in the inferior lobule and signals within a preset frequency band in the primary motor cortex, respectively, and calculate the corresponding instantaneous phase. Based on the instantaneous phase, the phase lock value between the inferior lobule and the primary motor cortex is calculated. The connection strength of intercortical functions is determined based on the positive correlation between the phase-locked value and the connection strength of intercortical functions; The coupling strength calculation unit is specifically used to extract signals within a preset frequency band range in the motor cortex of the brain, calculate the time delay cross-correlation value between the signals and the corresponding electromyographic signals within a time delay window, and determine the coupling strength between the nerves and muscles based on the positive correlation between the time delay cross-correlation value and the coupling strength between the nerves and muscles.

6. The system according to claim 5, characterized in that, The adaptive adjustment module includes: A motion intention decision unit is used to construct a hybrid architecture model, which captures the nonlinear characteristics of the bispectral peaks, extracts the dynamic change characteristics of the phase-locked values, analyzes the coordination patterns of nerves and muscles, and outputs motion intention decisions; wherein, the motion intention decisions include motion intention confidence, motion pattern, and system stability; A torque control unit is used to adjust the exoskeleton torque based on the motion intention confidence level and the phase lock value in the motion intention decision; wherein the phase lock value is inversely proportional to the exoskeleton torque; The difficulty adjustment unit is used to adjust the difficulty of the virtual reality task based on the bispectral feature data of the μ rhythm and the classification results of the motion pattern. A sensitivity adjustment unit is used to adjust the neural feedback sensitivity based on the coefficient of variation of the time delay cross-correlation value; wherein the neural feedback sensitivity is proportional to the signal interference degree.

7. The system according to claim 6, characterized in that, The motion intention decision unit also includes: A motion intent classification subunit is used to identify motion intents using the hybrid architecture model and the static features, and to classify each type of motion intent. The motion intent activation recognition subunit is used to identify whether the corresponding category of motion intent is in an active state using the hybrid architecture model and the dynamic features. If it is in an active state, the training action corresponding to the motion intent category is executed. If it is in an inactive state, the training action is not executed this time.

8. The system according to claim 1, characterized in that, The adaptive adjustment module further includes: A protection mechanism unit is used to activate the protection mechanism and reduce the training intensity when the coefficient of variation of the time delay cross-correlation value is abnormal.

Images