A brain-computer interface neural signal classification and recognition method and system based on deep learning

Abstract

The application provides a brain-computer interface neural signal classification and recognition method and system based on deep learning, and relates to the technical field of brain-computer interface and neural engineering, and the method comprises the following steps: acquiring original electroencephalogram signals of a stroke hemiplegic patient in a continuous rehabilitation training process, the original electroencephalogram signals are used for controlling a flexible exoskeleton with a physical response delay characteristic, and the original electroencephalogram signals are divided into a plurality of continuous time window sequences; according to the time window sequence, spatial topological features and time sequence features in each time window are extracted, and an initial electroencephalogram feature sequence is obtained; based on the initial electroencephalogram feature sequence, a neural intention evolution trajectory is subjected to space-time topological mapping, and a dynamic intention feature sequence representing the continuous evolution law of the electroencephalogram signals is obtained. The application can effectively solve the problem that the non-stationary space-time evolution of neural signals and the physical lag of the mechanical response of the exoskeleton exist in the dynamic coupling mismatch between the existing brain-computer interface technology and the continuous rehabilitation training scene.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface and neural engineering technology, and in particular to a brain-computer interface neural signal classification and recognition method and system based on deep learning. Background Technology

[0002] Brain-computer interface (BCI) technology combined with a flexible exoskeleton offers an innovative, continuous rehabilitation training program for stroke patients with hemiplegia. In this scenario, the system collects EEG signals when the patient generates specific motor intentions, decodes them using a deep learning model, and then drives the flexible exoskeleton to perform corresponding assistive movements. This aims to create a closed-loop rehabilitation training system encompassing brain intention generation, signal acquisition and decoding, exoskeleton mechanical response, and proprioceptive feedback. Currently, most mainstream EEG signal classification and recognition methods employ a static concatenated mapping model that combines spatial feature extraction with temporal feature extraction.

[0003] However, existing technologies reveal the following shortcomings in continuous rehabilitation training scenarios lasting tens of minutes: there is a dynamic coupling mismatch between the non-stationary spatiotemporal evolution of neural signals and the physical lag in the mechanical response of the exoskeleton. Existing models typically treat EEG signals as stationary sequences, neglecting both the spatiotemporal topological evolution of EEG signals caused by muscle fatigue or fluctuations in attention (i.e., neurodynamic drift) and the nonlinear coupling relationship between the inherent mechanical response lag of the flexible exoskeleton (such as motor start-up delay and flexible material deformation recovery time) and the trajectory of neural intent evolution.

[0004] This can lead to a situation where, when neural signals drift while the decision boundaries of the classification model remain fixed, it can easily result in the false triggering or delayed response of the system's output instructions, causing a temporal misalignment between intention and execution in the brain-computer interface. This continuous temporal misalignment can transmit incorrect or even delayed proprioceptive feedback to the patient, thereby disrupting the neuroplasticity reconstruction process of the patient's motor cortex and even triggering feelings of frustration and abnormal compensatory movements in the patient. Summary of the Invention

[0005] This invention provides a brain-computer interface neural signal classification and recognition method and system based on deep learning, which can effectively solve the problem of dynamic coupling mismatch between the non-stationary spatiotemporal evolution of neural signals and the physical lag of exoskeleton mechanical response in continuous rehabilitation training scenarios of existing brain-computer interface technology. It can reasonably avoid the technical problems of system command mis-triggering, response delay, and further damage to the reconstruction of neural plasticity of the patient's motor cortex.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, a brain-computer interface neural signal classification and recognition method based on deep learning is provided, the method comprising: Step 1: Obtain the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics. At the same time, the raw EEG signals are divided into multiple continuous time window sequences. Step 2: Based on the time window sequence, extract the spatial topological features and temporal features within each time window to obtain the initial EEG feature sequence; Step 3: Based on the initial EEG feature sequence, perform spatiotemporal topological mapping of the neural intention evolution trajectory to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals; Step 4: Based on the dynamic intention feature sequence, decode and identify it using a pre-trained deep learning classification model to obtain the initial neural intention probability distribution reflecting the current motion intention; Step 5: Determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation for intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration on the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain the calibrated target control command; drive the flexible exoskeleton to perform the corresponding auxiliary movement based on the calibrated target control command.

[0007] Secondly, a brain-computer interface neural signal classification and recognition system based on deep learning includes: The signal acquisition and segmentation module is used to acquire the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics, and the raw EEG signals are divided into multiple continuous time window sequences. The feature extraction module is used to extract the spatial topological features and temporal features within each time window based on the time window sequence, so as to obtain the initial EEG feature sequence. The topology mapping module is used to perform spatiotemporal topology mapping of the neural intention evolution trajectory based on the initial EEG feature sequence, so as to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals. The decoding and recognition module is used to decode and recognize the dynamic intention feature sequence using a pre-trained deep learning classification model to obtain an initial neural intention probability distribution that reflects the current motion intention. The hysteresis compensation and calibration module is used to determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation of intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration of the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain a calibrated target control command; and drive the flexible exoskeleton to perform corresponding auxiliary movements based on the calibrated target control command.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By calculating the characteristic state transition relationship between adjacent time windows based on the initial EEG feature sequence, a spatiotemporal topological mapping of the neural intention evolution trajectory is constructed. Nonlinear mapping and adaptive alignment processing are then applied to it to dynamically characterize the topological evolution law of EEG signals within continuous time windows. This overcomes the defect of the traditional static serial mapping model, which ignores the neurodynamic drift caused by patient muscle fatigue or attention fluctuations, resulting in a fixed decision boundary. This improves the robustness and stability of motor intention decoding in long-term continuous rehabilitation training.

[0009] Meanwhile, by acquiring the real-time physical state parameters of the flexible exoskeleton to determine its mechanical response delay, the time deviation between intention generation and mechanical execution is calculated, and a dynamic temporal hysteresis compensation factor is generated. This actively compensates for the temporal offset of the initial neural intention probability distribution, thereby solving the core problem of dynamic coupling mismatch between the non-stationary evolution of neural signals and the hysteresis of the exoskeleton's physical response. This effectively eliminates the temporal misalignment of brain-computer interface closed-loop and the false triggering of control commands caused by response lag. Finally, through the organic synergy of neural signal spatiotemporal evolution modeling and execution-end hysteresis compensation, the temporal alignment of calibrated control commands and exoskeleton mechanical execution is achieved. This provides technical assurance for the temporal consistency of intention, action, and feedback in brain-computer interface closed-loop training and reduces the risk of false triggering of commands due to temporal misalignment. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a brain-computer interface neural signal classification and recognition method based on deep learning, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of a brain-computer interface neural signal classification and recognition system based on deep learning, provided by an embodiment of the present invention. Figure 3 This is a feature distribution evolution trajectory diagram provided by an embodiment of the present invention when adaptive alignment is not performed; Figure 4 This is a feature distribution evolution trajectory diagram with adaptive alignment added, provided by an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown, an embodiment of the present invention proposes a brain-computer interface neural signal classification and recognition method based on deep learning, the method comprising the following steps: Step 1: Obtain the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics. At the same time, the raw EEG signals are divided into multiple continuous time window sequences. Step 2: Based on the time window sequence, extract the spatial topological features and temporal features within each time window to obtain the initial EEG feature sequence; Step 3: Based on the initial EEG feature sequence, perform spatiotemporal topological mapping of the neural intention evolution trajectory to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals; Step 4: Based on the dynamic intention feature sequence, decode and identify it using a pre-trained deep learning classification model to obtain the initial neural intention probability distribution reflecting the current motion intention; Step 5: Determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation for intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration on the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain the calibrated target control command; drive the flexible exoskeleton to perform the corresponding auxiliary movement based on the calibrated target control command.

[0013] In this embodiment of the invention, by constructing a spatiotemporal topological mapping of the neural intention evolution trajectory, the dynamic intention feature sequence can track the neurodynamic drift caused by patient fatigue or attention fluctuations in real time, overcoming the limitation of fixed decision boundaries in traditional static models and improving the robustness of intention decoding in long-term continuous rehabilitation training.

[0014] By using a dynamic temporal lag compensation factor based on the mechanical response delay of the exoskeleton, the probability distribution of the initial neural intent is actively calibrated, effectively compensating for the dynamic coupling mismatch between neural signal evolution and the physical response delay of the exoskeleton, and eliminating the brain-computer closed-loop temporal misalignment and command mis-triggering caused by response lag.

[0015] By organically combining non-stationary spatiotemporal evolution modeling of EEG signals with execution-end hysteresis compensation, a dynamic adaptation is formed across the entire link from feature encoding to control commands. This ensures that the proprioceptive feedback received by the patient is synchronized with their own motor intentions in real time, which is conducive to protecting and promoting the reconstruction of neuroplasticity in the motor cortex, while reducing the patient's frustration and the risk of abnormal compensatory movements.

[0016] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Acquire the initial acquisition signal of the stroke hemiplegic patient during continuous rehabilitation training. Perform artifact suppression and band filtering on the initial acquisition signal to obtain the original EEG signal. Specifically, this includes: acquiring the initial scalp EEG signal generated by the stroke hemiplegic patient during continuous rehabilitation training using a multi-channel EEG acquisition device; the multi-channel EEG acquisition device typically includes an EEG signal amplifier and a high-density electrode cap. Several recording electrodes are placed on the patient's scalp surface according to the international 10-20 standard lead layout. This layout uses the intersection of the longitudinal line from the root of the nose to the occipital protuberance and the transverse line connecting the bilateral preauricular points to determine the center position of the cranial vault, and then extends the electrodes outwards at certain intervals around this center to ensure that each electrode covers the main functional areas of the sensorimotor cortex while maintaining the standardized correspondence of the spatial positions between channels; during the acquisition process, all channels are recorded synchronously at a set sampling frequency. The resulting initial acquisition signal contains both neural electrical activity components related to the generation, maintenance, and switching of the patient's motor intentions, and non-neuronal noise components such as electrooculogram artifacts, electromyogram artifacts, and power line interference.

[0017] In one embodiment of this application, when performing artifact suppression processing on the initial acquired signal, an independent component analysis algorithm is used for signal decomposition and artifact removal. This algorithm is based on the assumption that the source signal components in a multi-channel EEG signal are statistically independent. By optimizing the separation matrix, the statistical independence between the output components is maximized, thereby decomposing the observed multi-channel mixed signal into several independent spatial components. After decomposition, the scalp topography distribution characteristics and temporal waveform characteristics of each independent component are calculated, and automatic screening is performed according to typical artifact component identification criteria. The specific identification criteria include: the EEG artifact component criterion is that the energy proportion of the component scalp topography in the frontal region electrodes FP1, FP2, and FPz exceeds [a certain value]. This component accounts for 60% of the total brain energy, and the energy proportion of electrodes F7 and F8 in the periocular region exceeds 30%. The criterion for EMG artifact components is that the ratio of the average power of the component power spectrum in the high-frequency band above 30Hz to the average power in the frequency band from 8Hz to 30Hz exceeds 2.0. The criterion for temporal synchronization is that the Pearson correlation coefficient between the component time process and the synchronously recorded EMG reference channel or EMG reference channel signal exceeds 0.7. Components that meet any of the above criteria are marked as artifact components and set to 0. After the identified artifact components are set to zero, they are reconstructed back to the multi-channel signal space through the inverse transformation of the separation matrix, thereby eliminating artifact interference introduced by eye movement and muscle activity and completely preserving the EEG spatial topology related to the activation state of the sensorimotor cortex.

[0018] The EEG spatial topology refers to the energy distribution pattern of neural oscillations at different frequencies on the multi-channel electrode space of the scalp surface. The typical spatial topology related to hand motor intention is manifested as local power modulation of the electrode position in the contralateral sensorimotor cortex region caused by the desynchronization of mu and beta rhythm events. This spatial topology pattern exhibits regular spatial redistribution characteristics as the type of motor intention changes and neurodynamic drift occurs. Its complete preservation is a key prerequisite for accurately extracting spatial topology features and tracking the evolution trajectory of neural intention in subsequent steps. Based on this, the signal after artifact removal is subjected to band filtering. A zero-phase-shift bandpass filter is used to filter the signal twice, forward and backward, to eliminate the phase distortion introduced by the filter and limit the signal frequency band to a specific rhythm range closely related to motor imagination and motor execution.

[0019] The specific rhythmic range mainly includes the mu rhythm band and the beta rhythm band. The mu rhythm is mainly distributed in the sensorimotor cortex and exhibits significant power suppression during motor intention generation. The beta rhythm is also distributed in the sensorimotor cortex and exhibits event-related desynchronization and subsequent beta rebound effects during motor execution and motor imagery. The power changes of neural oscillations within these two frequency bands directly reflect the synchronous activation state and desynchronization degree of motor cortex neuronal clusters, and are the core frequency domain feature source for decoding motor intention categories and capturing neurodynamic drift. During the filtering process, low-frequency baseline drift components in the signal are simultaneously filtered out, slow potential fluctuations below the mu rhythm are attenuated, and high-frequency electromagnetic interference components above the upper limit of the beta rhythm are filtered out to obtain narrowband neural oscillation signals highly correlated with motor intention. The original EEG signal obtained after artifact suppression and frequency band filtering effectively removes interference information unrelated to motor intention and retains the specific rhythmic neural oscillation components reflecting changes in the activation state of the motor cortex and their spatial topology on the scalp surface.

[0020] Step 1.2: Based on the original EEG signals, continuous segmentation processing is performed according to a preset window length and sliding step size to obtain the multiple continuous time window sequences. Specifically, this includes: based on the obtained original EEG signals, and according to the typical time-dynamic characteristics of changes in motor intention in stroke hemiplegic patients during continuous rehabilitation training and the physical hysteresis period of the mechanical response of the flexible exoskeleton, a fixed length of the time window and a sliding step size between adjacent windows are preset; the typical time-dynamic characteristics of changes in motor intention refer to the time evolution law exhibited by the neural response process from perceiving task prompts to generating stable motor intentions in stroke hemiplegic patients during rehabilitation training.

[0021] The transition from the emergence of a motor intention to a stable neural oscillation power modulation state typically requires a transition process ranging from hundreds of milliseconds to seconds. This includes the initial latency of event-related desynchronization, the time point when power decay reaches its peak, and the complete time history of neural oscillation activity recovering to the baseline level. This timescale determines that a single time window must have a sufficient time span to fully capture the entire neurodynamic process of a single motor intention from triggering to stable expression. The physical hysteresis period of the mechanical response at the actuator end of the flexible exoskeleton refers to the inherent response delay caused by physical factors such as the inflation and deflation time of the pneumatic pipeline, the viscoelastic deformation recovery time of the flexible material, and the start-stop transition process of the motor between the flexible exoskeleton receiving the control command and generating an effective auxiliary force or auxiliary displacement. The specific value of this hysteresis period depends on the driving method and structural material characteristics of the exoskeleton, and is usually in the range of tens to hundreds of milliseconds.

[0022] The window length must be set to ensure that the neural oscillation process from the emergence to stable expression of a motor intention can be completely captured within a single time window. Its value is determined based on the event-correlated desynchronization time history of the mu and beta rhythms. For example, the window length can be set to 2 to 3 times the typical transition time of the motor intention to provide sufficient temporal context information. The sliding step size must be smaller than the window length to ensure partial signal overlap between adjacent windows, and must match the physical hysteresis cycle of the exoskeleton's mechanical response to ensure the temporal resolution of the timing compensation in step 5. For example, the sliding step size can be set to a fraction of the hysteresis cycle or consistent with the sampling interval of the system's real-time control cycle. In one specific embodiment, the EEG signal sampling frequency is set to 250Hz. Based on the typical time history of event-correlated desynchronization of the mu and beta rhythms, the transition time from triggering to stable expression of a motor intention is approximately 400ms to 800ms. Therefore, the time window length is set to 1000ms. ms refers to 250 sampling points. The physical hysteresis period of the mechanical response driven by the flexible exoskeleton pneumatically is calibrated to be about 80ms by step response experiments. Based on this, the sliding step size is set to 40ms, or 10 sampling points, so that the sliding step size is about 0.5 of the hysteresis period to ensure the time resolution of the timing compensation in step 5. The overlap rate between adjacent time windows is set to 96%, or 240 overlapping sampling points. In one embodiment of this application, the first time window is intercepted from the starting time along the time axis of the original EEG signal according to the above-mentioned preset window length. The time window contains EEG data segments covering all recording channels. Subsequently, the second time window is intercepted by advancing along the time axis with a preset sliding step size. The second window shares part of the EEG data with the previous window because the sliding step size is smaller than the window length. In this way, the entire time axis of the original EEG signal is intercepted successively to form a continuous time window sequence composed of several time windows arranged in sequence and partially overlapping with each other.

[0023] This time window sequence discretizes continuous EEG signals into basic processing units that can be independently analyzed in subsequent steps using a unified time division benchmark. The overlapping structure between adjacent windows introduces information redundancy in the time dimension, enabling step 2 to effectively capture the continuous dynamic evolution information of EEG signals when extracting spatial topological features and temporal features. On this basis, the division method and window length parameter of this time window sequence also provide a unified time grid reference for the spatiotemporal topological mapping of neural intention evolution trajectory in step 3. The feature state transition calculation between adjacent windows can be directly expanded based on the topological feature matrix of each node on this time grid.

[0024] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: For the multi-channel EEG data within each time window, calculate the frequency domain synchronicity measure between each pair of channels within the preset target rhythm frequency band. Average the synchronicity measure values ​​at each frequency point within the preset target rhythm frequency band to obtain the average correlation strength of the channel pair in the frequency band. Arrange the average correlation strengths of all channel pairs according to the channel number to construct a symmetric correlation matrix, which serves as the spatial topological feature matrix corresponding to the time window. Specifically, this includes: Based on the obtained multiple consecutive time window sequences, let the total number of time windows be... The total number of EEG signal recording channels is For the first a time window ( ), and retrieve them one by one from the window covering all of them. The system records EEG data segments from multiple channels; it performs inter-channel spatial correlation analysis on these data segments to quantify the co-activation patterns of EEG activity in each channel in the spatial dimension. In one embodiment of this application, the average correlation strength between each pair of channels in a specific rhythmic frequency band is calculated for multi-channel EEG data within a single time window. This average correlation strength is measured using the average coherence coefficient, which reflects the degree of linear synchronization between two signals at a specific frequency. For the first... A time window, recording the channel and channels The EEG signal segments within this window are respectively and ,in For the index of time sampling points within the window, , The number of time sampling points corresponding to the preset window length in step 1.2; first calculate the channels. and channels At each frequency point within the frequency band coherence coefficient on Its definition is:

[0025] in, The value range is the specific rhythm frequency band described in step 1.1. All discrete frequency points within; and Channels and channels In the time window internal frequency The estimated power spectral density at a given location was calculated using the Welch average periodogram method, with a Hanning window used as the piecewise window function. The length of each piecewise window was equal to the number of sampling points within the window. The overlap rate between adjacent segments is 50%; For channel and channels In the time window internal frequency The cross-power spectral density estimate at point is calculated using the same method as the self-power spectral density estimate, employing the Welch average periodogram method. The piecewise window function used is the Hanning window, with each segment having a length equal to the number of sampling points within the window. The overlap rate of adjacent segments is 50%, which is calculated by performing Fourier transform on the corresponding segments of the two channels, multiplying them by their conjugates, and then taking the average. The square of the modulus of the cross-power spectral density characterizes the frequency response of the two signals. The common energy components at the location; the coherence coefficient The value of is between 0 and 1. The closer the value is to 1, the more linearly synchronized the two channel signals are at that frequency point.

[0026] Based on this, the frequency band The arithmetic mean of the coherence coefficients at all frequency points within the range is used to obtain the first... Channel within a time window and channels Average coherence coefficient in a specific rhythm frequency band The calculation formula is: In the formula For frequency band The total number of discrete frequency points within the interval; the average coherence coefficient. This comprehensively reflects the overall linear synchronization of neural oscillations between two channels within a specific rhythmic frequency band; the closer the value is to 1, the stronger the synchronization of the two channels. and channels The stronger the cooperative activation in this frequency band, the weaker it is.

[0027] Each pair of channels calculated Average coherence coefficient As the correlation strength value, arranged in order of channel number, a dimension is constructed. Symmetric Incidence Matrix , where matrix elements Due to the symmetry of the coherence coefficients, this matrix satisfies and diagonal elements Based on this, the aforementioned symmetric incidence matrix... Directly as the first The spatial topological feature matrix corresponding to each time window; the spatial topological feature matrix numerically describes the specific rhythmic spatial energy distribution pattern of multi-channel EEG signals on the scalp surface and the functional connectivity strength between different cortical regions within the current time window, retaining the EEG spatial topological structure information related to the sensorimotor cortex activation state described in step 1.1, so that the spatial topological redistribution caused by motor intention switching or neurodynamic drift can be quantitatively captured.

[0028] Step 2.2: For each pair of adjacent time windows, perform element-wise difference operations on the spatial topological feature matrices corresponding to the current time window and the previous time window to obtain a difference matrix; calculate the statistical mean of all elements in the difference matrix to obtain the global shift trend of spatial coupling strength; select channel sub-blocks corresponding to key brain regions of the sensorimotor cortex and calculate the local mean and standard deviation of the elements within the channel sub-blocks to obtain the degree of local remodeling of the key brain regions; and perform matrix decomposition on the difference matrix and extract information on the dominant change direction to obtain the structural features of spatial topological redistribution; arrange the global shift trend, the degree of local remodeling, and the structural features in a preset order to generate temporal feature vectors corresponding to adjacent time windows, specifically including: Based on the obtained spatial topological feature matrices corresponding to each time window, the variation patterns of spatial topological patterns between adjacent time windows are analyzed along the time dimension to extract the temporal dynamic variation features of EEG signals; in one embodiment of this application, for any two adjacent time windows on the time axis, i.e. the first... The first time window and the first For each of the three time windows, the corresponding spatial topological feature matrix is ​​obtained. and Perform element-wise difference operations on these two matrices: that is... Subtract the element value at each position from The element values ​​at the corresponding positions are used to generate a difference matrix Δ of the same dimension as the original spatial topological feature matrix. The element value in the i-th row and j-th column of this difference matrix directly reflects the strength of the neural functional connectivity between channel i and channel j from the i-th row and j-th column. The time window to the The net change generated during the transition of each time window is represented by a positive value, which indicates that the spatial coupling strength is enhanced during the transition, a negative value, which indicates that the spatial coupling strength is weakened during the transition, and 0 indicates that the spatial coupling strength remains unchanged.

[0029] Based on this, multi-level quantization feature extraction is performed on the above difference matrix to generate a temporal feature vector that can compactly represent the overall trend of spatial topological changes between adjacent time windows. The first level of quantization feature extraction is to calculate the statistical mean of all elements in the difference matrix. Specifically, the arithmetic mean of all M×M element values ​​in the difference matrix is ​​taken. This statistical mean reflects the overall shift magnitude and direction of functional connectivity strength between all channel pairs during the transition between adjacent windows, thereby capturing the global drift trend of the EEG signal spatial topology. The second level of quantization feature extraction is to calculate the corresponding values ​​of specific regions of interest in the difference matrix. The local statistics of the sub-blocks are specifically obtained by selecting sub-blocks corresponding to the recording channels of key brain regions of the sensorimotor cortex from the difference matrix. The recording channels corresponding to the key brain regions of the sensorimotor cortex are determined according to the international 10-20 standard lead layout as C3, C4, Cz and their four adjacent channels (front, back, left, and right), for a total of 15 channels. The mean and standard deviation of the element values ​​within the sub-block are calculated separately. These local statistics reflect the degree of local reshaping of the spatial coupling pattern between adjacent windows of the cortical regions directly related to motor intention, so that the local spatial topological redistribution caused by switching of motor intention or fluctuation of attention can be independently quantified.

[0030] The third level of quantization feature extraction is to extract the dominant change direction information of the difference matrix. Specifically, this is done by analyzing the difference matrix Δ... Perform singular value decomposition, i.e. ,in It is a left singular vector matrix. It is a singular value diagonal matrix. Given a right singular vector matrix, select the largest singular value. The corresponding left singular vector The vector has dimension M. To control the feature dimension and retain the main spatial information that dominates the direction of change, only the first part is taken. The component with the largest absolute value The value is the smaller of 5 and M divided by 6, rounded up. The filtered components are arranged in their original channel index order and used as a compressed representation of the dominant change direction feature vector, with a dimension of [dimensional value missing]. This vector characterizes the most significant spatial pattern of changes in multi-channel spatial coupling modes, capturing the structural features of spatial topological redistribution rather than merely a statistical description of the magnitude of change. All the quantitative indicators extracted from the above three levels are arranged in a preset order to form a vector corresponding to the first... For the temporal feature vectors of adjacent time windows, these vectors comprehensively include global amplitude information of spatial topological changes during the current window transition, local remodeling information of key brain regions, and information on the dominant direction of change. The above operations are performed on all adjacent window pairs one by one, and the generated temporal feature vectors are arranged sequentially along the time axis to form a feature sequence corresponding to the temporal analysis. This sequence of temporal feature vectors directly reflects the continuous dynamic evolution of the spatial topological structure of EEG signals over time, providing basic feature information of the temporal evolution dimension for constructing the initial topological representation of the neural intention evolution trajectory in step 3.

[0031] Step 2.3: Stretch the spatial topological feature matrix corresponding to each time window into a spatial feature vector. Concatenate this spatial feature vector with the temporal feature vector representing the change of this time window relative to the previous time window to form a combined feature vector. Arrange the combined feature vectors of each time window sequentially along the time axis to obtain the initial EEG feature sequence, specifically including: Based on the spatial topological feature matrix corresponding to each time window and the temporal feature vector representing the changes between adjacent windows, the two types of features are spliced ​​and fused in multiple dimensions to construct a unified feature representation that can simultaneously carry spatial static distribution information and temporal dynamic evolution information.

[0032] In one embodiment of this application, for each time window, the corresponding spatial topological feature matrix is ​​first... The matrix is ​​stretched into a spatial feature vector according to a preset unfolding order. The unfolding order can adopt a fixed rule of row-wise or column-wise unfolding, transforming the M×M dimensional symmetric matrix into a vector of dimension M. 2 A spatial feature vector of dimension, which contains a complete description of the spatial coupling strength of neural activity between all channel pairs within the current window.

[0033] Simultaneously, from the obtained time-series feature vector sequence, the current time window, i.e., the [missing information], is extracted. The time window and its previous time window, i.e., the first time window The time series feature vector calculated in the first time window directly carries the time series feature vector from the first time window. The window transitions to the first one. Information on the dynamic changes in spatial topology that occur during a window period.

[0034] The spatial feature vector and the temporal feature vector are concatenated end-to-end along the same feature dimension. That is, the components of the temporal feature vector are sequentially appended to the end of the spatial feature vector to form a combined feature vector that combines spatial state description and temporal evolution description. The first half of this combined feature vector is the spatial static attribute encoding of the current window, and the second half is the dynamic change attribute encoding of the current window relative to the previous window. The above feature concatenation operation is performed on all time windows one by one, and the combined feature vectors corresponding to each window are combined sequentially according to the arrangement order of the time windows to form an initial EEG feature sequence arranged along the time axis.

[0035] The feature vectors at each time position in this initial EEG feature sequence achieve cross-domain fusion of spatial and temporal features, so that the spatial organization state of EEG activity at the current moment and the evolutionary trend information of this moment relative to the previous moment are uniformly encoded in the same feature space, enabling subsequent steps to collaboratively analyze the spatial topological reorganization process and temporal evolution law of EEG signals under the same feature representation framework.

[0036] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: The combined feature vector of each time window in the initial EEG feature sequence is used as a state node in the neural intention feature space. For two temporally adjacent state nodes, the dimensional change vector between the two state nodes is calculated. The Euclidean distance between the two state nodes in the feature space is calculated as the overall offset. The state nodes are sequentially connected along the time axis, and the dimensional change vector and the overall offset together form a directed edge feature, constructing the initial topological representation of the neural intention evolution trajectory. Specifically, this includes: In this application, neurodynamics refers to the intrinsic law that the spatiotemporal patterns of electroencephalogram (EEG) signals related to the generation, maintenance, and switching of motor intentions in stroke hemiplegic patients undergo continuous dynamic evolution in response to internal cognitive states and external task conditions during continuous rehabilitation training. Specifically, during rehabilitation training lasting tens of minutes, the patient's EEG signals do not remain stable. The spatial distribution pattern of their neural oscillation activity, the energy intensity of rhythms in each frequency band, and the strength of functional connections between different cortical regions undergo systematic spatiotemporal topological evolution as the patient's level of attention fluctuates, muscle fatigue accumulates, and the type of motor task changes. This intrinsic evolutionary law is neurodynamics.

[0037] In this invention, the neurodynamic drift refers to the gradual shift in the statistical characteristics of EEG signals relative to the initial baseline state in the time dimension due to factors such as fluctuations in attention and accumulation of muscle fatigue during continuous rehabilitation training of stroke hemiplegic patients. This is manifested as a structural change in the overall distribution pattern of the inter-channel coherence coefficients in the spatial topological feature matrix as the training progresses. Step 3 aims to explicitly model and compensate for the impact of the aforementioned neurodynamic drift on the accuracy of intention decoding by constructing an initial topological representation of the neural intention evolution trajectory and performing adaptive alignment processing on it.

[0038] Specifically, based on the obtained initial EEG feature sequence arranged along the time axis, let the first... The combined feature vector corresponding to each time window is The dimension of this vector is D, where T represents the total number of time windows; The first half of the dimension carries the spatial feature information obtained by stretching the spatial topological feature matrix in step 2.1, while the second half carries the temporal feature information obtained by quantization of the difference matrix in step 2.2; for two temporally adjacent combined feature vectors and Each of these is considered as a feature state at two consecutive moments in the neural intent feature space, and the feature state transition relationship between them is calculated. This feature state transition relationship consists of two parts: one part is a dimension-wise changing vector. ; Each dimensional component represents the neural intent feature state on the corresponding feature dimension from the first... The time window to the The net change within each time window, with positive values ​​indicating an increase in the strength of the feature in that dimension and negative values ​​indicating a decrease; the other part is the overall offset. That is, to The sum of the squares of each component is then taken as the square root, which represents the Euclidean distance between the neural intention states at adjacent time points in the feature space. This distance is used to quantify the overall magnitude of the neurodynamic drift.

[0039] Based on this, the characteristic state transition relationship between each pair of adjacent time windows is represented as a set of... and The jointly constructed directed edge features combine the feature vectors corresponding to each time window. As state nodes, all state nodes are sequentially connected along the time axis to construct an initial topological representation of the neural intention evolution trajectory, forming a directed graph structure. This initial topological representation describes the transition path of neural intention feature states in the time dimension in the form of a graph structure, with each state node... Corresponding to the spatial and temporal fusion characteristics of EEG at a specific moment, each line from point to The directed edge carries and The encoded information on the direction and magnitude of change allows the neurodynamic drift caused by patient muscle fatigue, attention fluctuations, or switching of motor intentions to be explicitly encoded as state transition trajectories in the topological graph. This initial topological representation provides a structured input for further spatiotemporal nonlinear mapping in step 3.2, and also provides a time reference for determining the amount of time deviation between the intention generation moment and the mechanical execution moment in step 5.

[0040] Step 3.2: Input the combined feature vectors corresponding to each state node in the initial topology representation into a feedforward transform network with nonlinear mapping capability. During the forward propagation of the feedforward transform network, the combined feature vector of the current state node is weighted and fused with the combined feature vectors corresponding to its neighboring nodes in its time neighborhood to obtain the input feature vector. The input feature vector is then nonlinearly transformed through multiple stacked fully connected layers, each followed by a linear rectified activation function with leakage, to obtain the high-level semantic feature vector. For the high-level semantic feature vectors of all time windows, adaptive alignment processing is performed using iterative statistical normalization based on exponential moving average to obtain the dynamic intent feature vector. The dynamic intent feature vectors of each time window are arranged sequentially along the time axis to form a dynamic intent feature sequence, specifically including: Based on the initial topological representation of the constructed neural intent evolution trajectory, nonlinear mapping and adaptive alignment processing of spatiotemporal dimensions are performed on each state node and state transition edge to generate a high-level feature sequence that can directly serve the deep learning classification model described in step 4.

[0041] In one embodiment of this application, the feedforward transform network with nonlinear mapping capability is based on an improved multilayer perceptron architecture. An adjacent state node feature diffusion mechanism is introduced between its hidden layers, enabling the network to simultaneously utilize the features of the current window itself and the context information in its temporal neighborhood to perform high-level semantic transformation.

[0042] The feedforward transform network is constructed as follows: a backbone mapping network consisting of several fully connected layers stacked sequentially is set up. As an example rather than a limiting configuration, the backbone mapping network contains three fully connected layers, with the number of neurons in each layer being [D, 128, K], where D is the input feature dimension and K is a preset intention feature embedding dimension with a value of 64. Each fully connected layer is followed by a non-linear activation function, and the input layer receives the state node from step 3.1. The D-dimensional combined feature vectors are used to generate a high-level semantic feature vector of dimension K in the output layer. K is the preset intention feature embedding dimension; the nonlinear activation function is the LeakyReLU linear rectified function, whose expression is as follows: ; in This represents the weighted sum of neurons input to this activation function. The preset positive slope coefficient is less than 1, with a classic value of 0.01. This ensures that approximately 1% of the gradient signal is retained when the input is negative to prevent complete neuron inactivation. It can also be adjusted within the range of 0.001 to 0.1 depending on the network depth and training data size. During the forward propagation of the network, for the... A time window is added before the first hidden layer of the backbone mapping network, which sets the current window state node. The weighted fusion is performed by combining the feature vectors corresponding to its neighboring nodes in the time neighborhood. During the offline training phase, complete time series data can be used. The calculation formula for the weighted fusion is as follows: The neighboring nodes include the nodes representing the previous time window. Current window state node and the state node of the next time window During the online inference phase, since future window data is not yet available, the weighted fusion calculation formula is adjusted to... = + The neighboring nodes only include the nodes in the previous time window. and the current window state node Weighting coefficients and Similarly, the feature similarity between two nodes is adaptively determined by the softmax function, and satisfies... + =1; where , , These are weighting coefficients, adaptively determined by the feature similarity between nodes using a normalized exponential function, softmax. The feature similarity is measured using cosine similarity; for two feature vectors a and b, their similarity score is the product of their inner product divided by their respective L2 norms. Specifically, the current node's similarity is calculated first... The cosine similarity with each neighboring node is used as the similarity score. The current node's own similarity score is set to the maximum value. Furthermore, each similarity score is divided by a temperature coefficient and then normalized using the softmax function to obtain a weighted coefficient. The softmax function is applied to the ... The expression for calculating the weighting coefficients of each node is: Indicates the first The similarity score for each node. Indicates the current node With index The similarity score between the neighboring nodes it points to; During the offline training phase, the neighbor index set { is represented by} , , }, representing the set of neighbor indices during the online inference phase { , }; Represented by natural constant Exponential function with base 0; temperature coefficient This is used to control the smoothness of the weight distribution, with a classic value of 0.5. This ensures that the differences between similarity scores are moderately compressed after normalization, and the weight allocation is neither too concentrated nor too uniform. It can be adjusted between 0.1 and 1.0 according to the fluctuation of feature states in the time neighborhood. When the value is small, the weight is more concentrated on the node with the highest similarity. When the values ​​are large, the weights tend to be equal; the generated input feature vector incorporates local temporal context. The data is fed into the backbone mapping network and subjected to nonlinear transformation to obtain... .

[0043] The training process of the feedforward transform network is jointly performed with the pre-trained deep learning classification model described in step 4. Specifically, in the offline training phase before system deployment, EEG signal data of multiple stroke hemiplegic patients during continuous rehabilitation training are collected, and the corresponding true motor intention category at each moment is labeled. In the offline training phase, EEG data of no less than 20 stroke hemiplegic patients during continuous rehabilitation training are collected. The patients are aged 40 to 75 years and Brunnstrom stage 3 to 5. No less than 5 training sessions are collected for each patient, and each session lasts about 30 minutes. Each session includes four types of movements: hand opening, hand clenching, wrist flexion, and wrist extension. Each motor intention is tested at least 50 times. Each test is triggered by a screen prompt, causing the patient to generate a specified motor intention. The prompt presentation time is used as the time reference for the actual motor intention label. When the model is deployed to a new patient, the parameters of the pre-trained model are fine-tuned using data from the patient's previous three training sessions. The fine-tuning strategy is to freeze the parameters of the first two fully connected layers of the feedforward transform network, updating only the parameters of the third fully connected layer and the intention decoding branch network. The fine-tuning learning rate is set to 0.1 of the initial training learning rate, and the number of fine-tuning iterations does not exceed 50 epochs. The raw EEG signals are processed using steps 1 to 2 to obtain the initial EEG feature sequence, and then the initial topological representation is constructed using the method in step 3.1. Further, each state node... The high-level semantic feature vector is obtained by inputting the above feedforward transform network. ,Will The input continues into an intent decoding branch network consisting of several fully connected layers and a classification output layer, which outputs the predicted probability of each preset motion intent category. The entire training process aims to minimize the cross-entropy loss between the classification prediction probability and the real motion intent label. The parameters of the feedforward transform network and the intent decoding branch network are updated simultaneously through error backpropagation. After training convergence, in this invention, the entire feedforward transform network and the intent decoding branch network, which have been jointly trained, are collectively referred to as the pre-trained deep learning classification model. Among them, some or all layers of the intent decoding branch network serve as the classification output, while the feedforward transform network provides nonlinear and context-aware feature mapping capabilities for step 3.2. During the training process, through mixed training on different rehabilitation training periods and different patient data, the feedforward transform network implicitly learns the ability to adaptively adjust to the neurodynamic changes of different drift speeds and drift amplitudes.

[0044] After the above nonlinear mapping is completed in the online inference phase, the high-level semantic feature vectors for all time windows are... An adaptive alignment process is performed. To meet the requirements of real-time online inference, this adaptive alignment process adopts an iterative statistical normalization operation based on exponential moving average, specifically: maintaining a global mean vector that is updated with the feature vector of each new time window. and global standard deviation vector For the current number High-level semantic feature vectors of each time window The update formula for the global mean vector is: The update formula for the global variance vector is: ,in The preset smoothing coefficient takes a value between 0 and 1, with a classic value of 0.01, which allows the statistic to slowly track the drift trend of the neural signal distribution. and These are the global mean vector and global variance vector after the previous time window update, respectively; Depend on The square root of each component is used to obtain the result; further, real-time updates are utilized. and right After normalization, dynamic intent feature vectors are obtained. Its calculation formula is ,in For a very small positive number, the classic value is This is to avoid calculation errors caused by a denominator of 0.

[0045] This maps the dynamic intent feature vectors of each time window to a uniform distribution scale, ensuring that the fixed decision boundary of the pre-trained deep learning classification model in step 4 maintains accurate discrimination capability when faced with continuously evolving neural signals. Based on this, experiments are conducted to visually verify the effectiveness of the adaptive alignment strategy in suppressing neurodynamic drift. Figure 3 and Figure 4 As shown in the figure, the dynamic intention feature vectors of the first to fifth training periods (periods 1 to 5) in the continuous rehabilitation training process are selected for dimensionality reduction visualization. The horizontal and vertical axes of the figure uniformly adopt the first two principal components (Principal Component 1, PC1 and Principal Component 2) of the dynamic intention features. These principal components are composed of high-level semantic feature vectors of all periods. Principal component analysis (PCA) projection is used to obtain the variance information in the feature space that is related to the distinction of motion intention categories to the greatest extent.

[0046] in, Figure 3 This represents the feature distribution evolution trajectory without adaptive alignment. Figure 4To incorporate the feature distribution evolution trajectory after adaptive alignment, the horizontal and vertical coordinates of the two images are defined in exactly the same way, only the processing methods are different; Figure 3 The horizontal axis represents PC1 (principal component 1), and the vertical axis represents PC2 (principal component 2). Different colors correspond to different training periods. Period 1 represents the initial training stage, where patients are focused and the feature distribution is concentrated in the lower left corner of the coordinate system, highly consistent with the distribution reference system during offline training. Period 2 represents about 5 minutes of training, where the feature distribution begins to shift slightly to the upper right, reflecting a global drift caused by the overall enhancement of neural oscillation power. Period 3 represents about 10 minutes of training, where the feature distribution further shifts to the upper right. Period 4 represents about 15 minutes of training, where the feature distribution shift intensifies, and some sample points have almost no overlap with the distribution in Period 1, indicating that the neurodynamic drift has significantly altered the feature statistical properties. Period 5 represents the later stage of training (about 20 minutes), where the feature distribution completely deviates from the initial reference system, and the center of the elliptical region shifts by more than 2 standard deviations compared to Period 1. If the original decision boundary is still used for decoding, a large number of misjudgments will occur.

[0047] Figure 4 (With adaptive alignment) The horizontal axis remains PC1 (principal component 1), and the vertical axis remains PC2 (principal component 2). The color definition is the same as... Figure 3 Consistent: Time period 1 and Figure 3 Consistent, the feature distribution is concentrated in the central region of the initial reference frame; period 2 shows that although the original features have slightly drifted, the global mean, updated in real time through exponential moving average, remains consistent. Compared with global standard deviation The feature distribution was immediately rescaled to return to a uniform scale; in period 3, the adaptive alignment operation continuously tracked the drift trend, and the feature distribution was pulled back to a coordinate range similar to that of period 1; in period 4, even with the accumulation of patient fatigue and decreased attention, the feature distribution remained stable within a uniform reference frame and maintained a high degree of overlap with the distribution of period 1; in period 5, the later stage of training, the feature distribution still did not show any systematic deviation, and the data points of all periods fluctuated around the same central region.

[0048] By using adaptive alignment processing based on exponential moving average, we can effectively track and compensate for neurodynamic drift in real time, and stably constrain the dynamic intent feature vector under a unified feature distribution scale. This ensures that the pre-trained deep learning classification model in step 4 can maintain high decoding stability and classification accuracy at different training periods, and minimizes the possibility of false intent triggering or response delay caused by feature distribution drift.

[0049] After the above nonlinear mapping and adaptive alignment processing, the normalized high-level semantic feature vectors corresponding to each time window are arranged sequentially along the time axis to form a dynamic intention feature sequence that represents the continuous evolution law of EEG signals. This dynamic intention feature sequence not only completely preserves the evolution trajectory information of neural intentions in continuous time windows, but also improves the distinguishability of features for motor intention categories through nonlinear mapping, and enhances the robustness to neurodynamic drift through adaptive alignment. It provides high-quality feature input for the deep learning classification model in step 4 to output an initial neural intention probability distribution that accurately reflects the current motor intention.

[0050] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: The dynamic intention feature vectors corresponding to each time window in the dynamic intention feature sequence are sequentially input into a pre-trained deep learning classification model. The classification model consists of several fully connected hidden layers and a classification output layer stacked sequentially. Each fully connected hidden layer performs a linear affine transformation on the input vector and then maps it through a non-linear activation function to obtain a hidden feature vector. The classification output layer performs a linear affine transformation on the hidden feature vector output by the last hidden layer to obtain an initial prediction score vector with a dimension equal to the total number of preset motion intention categories. Each component of the initial prediction score vector represents the original discrimination score of the EEG signal in the current time window belonging to each preset motion intention category, specifically including: Specifically, based on the obtained dynamic intent feature sequence arranged along the time axis, for the first... Normalized high-level semantic feature vectors corresponding to each time window Its dimensions are The input is then fed into the intent decoding branch network obtained after joint training and convergence with the feedforward transform network in step 3.2. This intent decoding branch network is the pre-trained deep learning classification model mentioned in this step. This deep learning classification model consists of several fully connected hidden layers and a classification output layer stacked sequentially. Let the total number of layers in the model be... As an exemplary and not restrictive configuration, this classification model contains two fully connected hidden layers with the number of neurons in the ranges [128, 64], and the classification output layer is the [128, 64]th layer. The number of neurons is C, which is the total number of preset motion intention categories. Each fully connected hidden layer is followed by the leaky linear rectified function. As a non-linear activation function, Dropout regularization is applied after each hidden layer, with a dropout rate set to 0.5 during training. The classification output layer does not have an activation function set to directly output a linear score.

[0051] The feature vector As the input to the first layer, forward propagation calculations are performed layer by layer, and at the [missing information]... layer( The operation involves first performing a linear affine transformation on the input vector of this layer, and then performing a nonlinear mapping. The calculation expression is as follows:

[0052] in , For the first The weight matrix of the fully connected hidden layer. For the first Layer bias vector, For the first The hidden feature vectors output by the first layer; the last layer, i.e., the classification output layer, is the hidden feature vector of the first layer. The layer output is subjected to a linear affine transformation to obtain the initial prediction score vector. Its calculation expression is:

[0053] in The weight matrix for the classification output layer has dimensions of . , The bias vector for the classification output layer has dimensions of . , The preset total number of motor intention categories; the preset motor intention categories include basic hand motor intention types used in rehabilitation training for stroke hemiplegic patients, such as hand opening, hand clenching, wrist flexion, and wrist extension; initial prediction score vector. The Each component This indicates that the EEG signal within this time window belongs to the first... The original discrimination score of a preset motion intention category, This score reflects the relative support of the deep learning classification model for each possible motion intention at the current moment without probability normalization.

[0054] Step 4.2: Input the initial predicted score vector into a preset probability normalization function. The preset probability normalization function performs an exponential operation on each component with the natural constant as the base, and calculates the ratio of the exponent value of each component to the sum of the exponent values ​​of all components to obtain the probability value corresponding to each preset motion intention category. Arrange the probability values ​​of all categories in order of category to form an initial neural intention probability distribution vector reflecting the current motion intention, specifically including: Based on the obtained first The initial prediction score vector corresponding to each time window The probability distribution of the neural intent is obtained by nonlinearly mapping the components using a preset probability normalization function to transform each component into a probability value between 0 and 1, with a sum of 1. The preset probability normalization function is the softmax function. The preset motion intention categories are mapped using the following formula:

[0055] in, Indicates the first EEG signals within the first time window were identified as the first The probability value of a preset motion intention category; This is the initial predicted score for the corresponding category output in step 4.1; Represented by natural constant Exponential function with base 0; denominator For all The index scores of each preset motion intention category are summed and used as a normalization factor to ensure that the sum of the probabilities is 1. To sum and iterate through the indices, the range of values ​​is... Same, indicating all preset motion intent categories; will all Probability values ​​of each category Arranged according to category order, forming the [number]th Initial neural intent probability distribution vector for each time window The initial neural intent probability distribution directly reflects the confidence estimate of the deep learning classification model for the patient's various possible movement intentions at the current moment, providing a probability-based decision-making basis for combining the temporal hysteresis compensation and generating the final control command in step 5.

[0056] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Real-time acquisition of the physical state parameters of the exoskeleton, including the current displacement, current velocity, current output torque of the end effector, and current pressure value in the drive cavity. When the flexible exoskeleton is driven by pneumatic artificial muscles, the absolute value of the pressure difference between the target pressure value required for the exoskeleton to perform the target auxiliary action and the real-time acquired current pressure value in the drive cavity is used, combined with the internal volume of the pneumatic drive cavity, the mass flow rate of the air supply pipeline, the specific gas constant of the working gas, and the absolute temperature of the working gas, to calculate the hysteresis time of the inflation and deflation process. The inflation and deflation hysteresis time is added to the pre-calibrated deformation recovery inherent response time of the flexible material to obtain the mechanical response delay time of the flexible exoskeleton, specifically including: During continuous rehabilitation training, displacement sensors, torque sensors, and air pressure sensors deployed at the actuator end of the flexible exoskeleton collect real-time physical state data of the exoskeleton. These real-time physical state parameters include the current displacement, current movement speed, current output torque of the exoskeleton's end effector, and the current air pressure value within the pneumatic drive chamber. Based on these real-time physical state parameters, the mechanical response delay time is determined according to the driving method of the flexible exoskeleton. In one embodiment of this application, when the flexible exoskeleton is driven by pneumatic artificial muscles, the mechanical response delay time is determined by the hysteresis duration of the pneumatic pipeline inflation and deflation process. The calculation formula for this hysteresis duration is as follows:

[0057] in, The inflation / deflation delay time, in seconds; The target absolute air pressure value required to drive the exoskeleton to perform the target assisted action, measured in Pascals; The pressure value inside the pneumatic drive chamber is collected in real time by the pressure sensor and converted into absolute pressure, in Pascals. It should be noted that if the pressure sensor outputs a gauge pressure value, it must be superimposed with the local standard atmospheric pressure to be converted into absolute pressure before it can be substituted into the formula for calculation, otherwise it will lead to time estimation deviation. The internal volume of the pneumatic drive cavity is expressed in cubic meters. This parameter is determined by the pneumatic structure design of the exoskeleton. The mass flow rate of the air supply pipeline for the pneumatic system, expressed in kilograms per second, is calibrated based on the rated flow rate of the air source and the flow conductance characteristics of the pipeline. The specific gas constant of the working gas used, in joules per kilogram per Kelvin, for dry air. Approximately 287 joules per kilogram per Kelvin. If other working gases such as carbon dioxide are used, they must be replaced with the corresponding specific gas constant value. The absolute temperature of the working gas, in Kelvin, can be obtained by adding 273.15 to the Celsius temperature value collected by the temperature sensor.

[0058] Based on this, the formula for calculating the mechanical response delay is: in This refers to the inherent mechanical response time of the flexible material's deformation recovery. This inherent mechanical response time is determined by the viscoelastic properties of the exoskeleton's flexible structural material and is obtained through pre-calibration via step response experiments. It should be noted that the above hysteresis time calculation formula is based on mass flow rate... The simplified assumption that it is a constant means that the mass flow rate will decrease slightly as the back pressure increases in the actual pneumatic system. However, for the rehabilitation training scenario targeted by this invention, where the air pressure difference is usually small and a stable air source is provided, this simplified model has sufficient engineering accuracy to meet the real-time and accuracy requirements for estimating the mechanical response delay time. The calculated mechanical response delay duration is expressed in time. When the flexible exoskeleton is driven by a shape memory alloy, the mechanical response delay duration is determined by the thermodynamic hysteresis time from the material's heating phase transition to its deformation response. It is calculated based on the difference between the current temperature and the material's phase transition temperature threshold, as well as the heating power parameters. The physical meaning of the mechanical response delay duration is the time interval from the moment the control command is sent until the moment the exoskeleton's end effector begins to generate effective displacement or effective torque output. The determination of this mechanical response delay duration provides a physical delay reference for the execution end of the calculation of the time deviation between the intention generation and mechanical execution in step 5.2.

[0059] Step 5.2: Using the start timestamp of the current time window as the benchmark for neural intention generation, the sum of the system processing time from EEG data input to output of the initial neural intention probability distribution and the mechanical response delay is determined as the time deviation between intention generation and mechanical execution, specifically including: Based on the obtained first Initial neural intent probability distribution vector for each time window And the determined mechanical response delay duration of the flexible exoskeleton The time deviation between the moment when the neural intent is generated and the moment when the flexible exoskeleton's mechanical execution takes effect is calculated.

[0060] In one embodiment of this application, the neural intent generation time is the first of the time window sequences in step 1.2. Based on the window start timestamp of each time window, this timestamp is denoted as... This moment represents the time when the patient's EEG signal was acquired and used to generate the motor intent decoding result, and its value is equal to the value obtained in step 1.2 by sliding step size starting from the initial moment of the original EEG signal. After each step of the process The start time coordinates of each window.

[0061] The effective time of mechanical execution of a flexible exoskeleton is defined as the time from the issuance of the control command to the mechanical response delay. The subsequent time point, where the control command is issued at the moment immediately after the output of the initial neural intent probability distribution in step 4.2, is used to generate the command. This time point is related to... The difference between them only includes the processing time consumed by feature extraction, topological mapping, and decoding calculations in steps 2 to 4, which is denoted as . The processing time The determination method is as follows: During the system initialization phase, a complete processing flow from steps 1 to 4 is run once without load, and the total time taken from the input of EEG data in the current time window to the output of the initial neural intent probability distribution is measured and recorded. This value is then calibrated as... Because the processing time mentioned above is much shorter in real-time systems than... Furthermore, it can be considered a fixed system delay and included in the calculation. The time deviation is compensated together; therefore, the formula for calculating the time deviation is as follows: ;in For the first The time deviation corresponding to the first time window, the physical meaning of which is the first time deviation. The neural intent decoding results for each time window need to be sent to the flexible exoskeleton in advance to offset the total duration of its mechanical response lag. The value is determined by the physical properties of the flexible exoskeleton and the current physical state. When the air pressure difference in the exoskeleton's pneumatic tubing is large... Enlargement leads to The pressure difference increases accordingly when the exoskeleton is initially at rest and the air pressure difference is small. Reduced The corresponding reduction; this time deviation is the core input parameter for generating the dynamic timing hysteresis compensation factor in step 5.3, and its calculation accuracy directly affects the accuracy of subsequent timing calibration.

[0062] Step 5.3: Based on the time deviation and the preset time window sliding step size, calculate the integer number of steps in units of the sliding step size, which serves as the dynamic timing hysteresis compensation factor for timing calibration. Specifically, this includes: based on the calculated time deviation... Combined with the preset time window sliding step size A dynamic temporal hysteresis compensation factor is calculated and generated for temporal calibration of the initial neural intent probability distribution; in one embodiment of this application, The value is set in step 1.2 based on the physical hysteresis period of the flexible exoskeleton's mechanical response. Its value is less than the preset window length in step 1.2 to ensure information overlap between adjacent windows. The value of also determines the time resolution of timing compensation; the formula for calculating the dynamic timing hysteresis compensation factor is as follows: in The function is called the floor function, and its function expression is: , indicating that the value is not less than The smallest integer of , where Input as a real number. The reason for using rounding up instead of rounding down or rounding to the nearest integer is to ensure that the number of compensation steps covers at least the complete hysteresis duration of the exoskeleton's mechanical response at the time window granularity, and to avoid timing misalignment caused by insufficient compensation due to rounding. The calculated hysteresis compensation steps are expressed in units of time window sliding steps. The integer value represents the number of windows that need to be shifted forward on the time window sequence to achieve the desired decoding result, after which the shifted window... The calibration probability distribution used in each time window The corresponding original decoding time and the first The actual mechanical execution times of the exoskeleton within each time window are aligned in time.

[0063] Will As a dynamic temporal hysteresis compensation factor, the value of this compensation factor is dynamically updated according to the changes in the real-time physical state parameters of the flexible exoskeleton. When the difference between the target air pressure and the current air pressure is small... Reduced The pressure difference decreases accordingly when the pressure difference is large. Enlargement leads to The corresponding increase allows for adaptive matching of the exoskeleton's mechanical response hysteresis; during rehabilitation training, as the exoskeleton initiates movement from a static state to continuously perform auxiliary movements, The value may gradually transition from an initial small value to a relatively constant value under stable motion conditions. This dynamic change characteristic ensures that the timing offset compensation in step 5.4 is always synchronized with the current physical state.

[0064] It should be noted that, in this application, the target intent probability distribution and the target intent probability distribution vector refer to the same mathematical object. The two equivalent expressions are not contradictory; the probability distribution focuses on its statistical function, that is, it describes the probability values ​​of C motion intention categories; the vector focuses on its data structure, that is, it is represented as a one-dimensional ordered array during the calculation process.

[0065] In a preferred embodiment of the present invention, step 5 above may further include: Step 5.4: During operation, a cache queue of initial neural intent probability distributions covering the current time and several previous time windows is continuously maintained. The cache queue stores the initial neural intent probability distribution vectors corresponding to each time window in order of time window number. When generating the calibration instruction for the current time window, the initial neural intent probability distribution vector corresponding to the number of windows shifted in the historical direction is extracted from the cache queue and used as the target intent probability distribution vector after the current time window is time-calibrated. When the window number shifted in the historical direction exceeds the range of the cache queue, the probability distribution vector at the initial moment is used as the default value, specifically including: In one embodiment of this application, the temporal offset compensation processing is implemented by continuously maintaining an initial neural intent probability distribution cache queue covering the current time and several previous time windows during the system operation. This queue stores the probability distribution vectors corresponding to each time window in ascending order of time window number; during the generation of the... When a calibration instruction for a given time window is received, the first one is retrieved from the cache queue. The initial neural intent probability distribution vector corresponding to each time window This is directly assigned the calibration probability distribution at the current time, i.e. ;in For the first The target intent probability distribution vector after time-window timing calibration.

[0066] The physical meaning of this assignment is to use the pre-decoded neural intent information, which matches the exoskeleton's current actual execution time, as the control basis, so that the intent decoding result of the control command sent to the flexible exoskeleton at the current moment exactly matches the exoskeleton's actual execution time after mechanical response hysteresis; when When the value is less than 1, meaning there are not enough historical probability distributions before the current time window, the probability distribution vector at the initial time is used. This is processed as the default value; the result after time-series offset compensation is... This refers to the target intent probability distribution after time-series calibration. This calibration operation shifts the neural intent probability distribution output by the decoder forward along the time axis, aligning the timing of the intent information reaching the exoskeleton execution end with the timing of the exoskeleton's mechanical response. This fundamentally and effectively eliminates the time-series misalignment problem between brain-computer closed-loop intent and execution described in the background technology.

[0067] Step 5.5: In the time-calibrated target intent probability distribution vector, determine the motion intent category corresponding to the component with the highest probability value, and determine whether the highest probability value meets the preset confidence threshold. When the highest probability value is greater than or equal to the confidence threshold, extract the corresponding target control command from the pre-established mapping relationship between motion intent categories and control commands, and send it to the control execution unit of the flexible exoskeleton to drive the exoskeleton to perform assisted movements that are time-synchronized with the patient's current motion intent. When the highest probability value is less than the confidence threshold, maintain and continue to execute the target control command from the previous time window, specifically including: Based on the obtained time-calibrated target intent probability distribution vector The highest probability motion intention category is determined from the C probability components of the vector (here and in step 4, it refers to the total number of preset motion intention categories). At the same time, it is checked whether the highest probability value meets the preset confidence threshold. In one embodiment of this application, the preset confidence threshold is set based on a comprehensive consideration of the classification accuracy of the deep learning classification model in offline testing and the tolerance for false triggering and missed triggering risks during rehabilitation training. The threshold is set to a value between 0.5 and 1.0, for example, 0.7. When the highest probability value is greater than or equal to 0.7, it indicates that the current decoding result has a high confidence level and can be safely converted into a control command. When the highest probability value is less than 0.7, it indicates that the model's judgment of the current intention is relatively ambiguous. At this time, no new control command is generated. The flexible exoskeleton maintains and continues to execute the target control command of the previous time window to maintain the current posture or assist in movement until a new highest probability motion intention category that meets the confidence threshold is generated in a subsequent time window. Thus, training interruption caused by missed triggering is avoided to the greatest extent while ensuring training safety.

[0068] When the highest probability value is greater than or equal to the preset confidence threshold, the motion intention category corresponding to the highest probability is determined as the valid motion intention at the current moment. The pre-established mapping relationship between motion intention categories and control commands is specifically that each preset motion intention category corresponds to a set of pre-calibrated exoskeleton execution parameters. Among them, the target control command mapped to the motion intention category of hand opening includes the target angle sequence and upper limit value of the extension torque for driving the extension joints of each finger of the exoskeleton; the target control command mapped to the motion intention category of hand clenching includes the target angle sequence and upper limit value of the flexion torque for driving the flexion joints of each finger of the exoskeleton; the target control command mapped to the motion intention category of wrist flexion includes the target angle and torque parameters for driving the wrist flexion degree of freedom of the exoskeleton; and the target control command mapped to the motion intention category of wrist extension includes the target angle and torque parameters for driving the wrist extension degree of freedom of the exoskeleton.

[0069] Based on the category of effective motor intention, the corresponding target control command is extracted from the mapping relationship and then sent to the control execution unit of the flexible exoskeleton. This drives the exoskeleton to perform auxiliary movements that are synchronous with and consistent with the patient's current motor intention, completing a closed-loop rehabilitation training process that spans the entire chain, from EEG signal acquisition, neural intention decoding, temporal delay compensation to exoskeleton mechanical execution. This process ensures that the proprioceptive feedback felt by the patient is highly synchronized with the motor intention generated by the patient in terms of time and content through the execution of accurate commands after temporal calibration. This provides a closed-loop training environment that conforms to physiological laws for the reconstruction of neuroplasticity of the motor cortex, effectively avoiding frustration and abnormal compensatory movements caused by temporal misalignment.

[0070] like Figure 2 As shown, embodiments of the present invention also provide a brain-computer interface neural signal classification and recognition system based on deep learning, comprising: The signal acquisition and segmentation module is used to acquire the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics, and the raw EEG signals are divided into multiple continuous time window sequences. The feature extraction module is used to extract the spatial topological features and temporal features within each time window based on the time window sequence, so as to obtain the initial EEG feature sequence. The topology mapping module is used to perform spatiotemporal topology mapping of the neural intention evolution trajectory based on the initial EEG feature sequence, so as to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals. The decoding and recognition module is used to decode and recognize the dynamic intention feature sequence using a pre-trained deep learning classification model to obtain an initial neural intention probability distribution that reflects the current motion intention. The hysteresis compensation and calibration module is used to determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation of intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration of the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain a calibrated target control command; and drive the flexible exoskeleton to perform corresponding auxiliary movements based on the calibrated target control command.

[0071] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0072] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A brain-computer interface neural signal classification and recognition method based on deep learning, characterized in that, The method includes: Step 1: Obtain the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics. At the same time, the raw EEG signals are divided into multiple continuous time window sequences. Step 2: Based on the time window sequence, extract the spatial topological features and temporal features within each time window to obtain the initial EEG feature sequence; Step 3: Based on the initial EEG feature sequence, perform spatiotemporal topological mapping of the neural intention evolution trajectory to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals; Step 4: Based on the dynamic intention feature sequence, decode and identify it using a pre-trained deep learning classification model to obtain the initial neural intention probability distribution reflecting the current motion intention; Step 5: Determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation for intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration on the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain the calibrated target control command; drive the flexible exoskeleton to perform the corresponding auxiliary movement based on the calibrated target control command.

2. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 1, characterized in that, Raw electroencephalogram (EEG) signals of stroke hemiplegic patients during continuous rehabilitation training were acquired. These raw EEG signals were used to control a flexible exoskeleton with physical response delay characteristics. The raw EEG signals were divided into multiple consecutive time window sequences, including: The initial acquisition signal of a stroke hemiplegic patient during continuous rehabilitation training is obtained, and the initial acquisition signal is processed by artifact suppression and frequency band filtering to obtain the original EEG signal. Based on the original EEG signal, continuous segmentation is performed according to the preset window length and sliding step size to obtain the multiple consecutive time window sequences.

3. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 2, characterized in that, The initial EEG feature sequence includes: For multi-channel EEG data within each time window, the frequency domain synchronization measure between each pair of channels is calculated within the preset target rhythm frequency band. The synchronization measure values ​​at each frequency point within the preset target rhythm frequency band are averaged to obtain the average correlation strength of the pair of channels in the frequency band. The average correlation strength of all channel pairs is arranged according to the channel number to construct a symmetric correlation matrix, which serves as the spatial topological feature matrix corresponding to the time window. For each pair of adjacent time windows, element-wise difference operations are performed on the spatial topological feature matrices corresponding to the current time window and the previous time window to obtain a difference matrix. The statistical mean of all elements in the difference matrix is ​​calculated to obtain the global shift trend of spatial coupling strength. Channel sub-blocks corresponding to key brain regions of the sensorimotor cortex are selected, and the local mean and standard deviation of the elements within the channel sub-blocks are calculated to obtain the degree of local remodeling of the key brain regions. Matrix decomposition is performed on the difference matrix, and information on the dominant change direction is extracted to obtain the structural features of spatial topological redistribution. The global shift trend, the degree of local remodeling, and the structural features are arranged in a preset order to generate temporal feature vectors corresponding to adjacent time windows. The spatial topological feature matrix corresponding to each time window is stretched into a spatial feature vector. This spatial feature vector is then concatenated with the temporal feature vector representing the change of this time window relative to the previous time window to form a combined feature vector. The combined feature vectors of each time window are arranged sequentially along the time axis to obtain the initial EEG feature sequence.

4. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 3, characterized in that, The dynamic intent feature sequence includes: The combined feature vector of each time window in the initial EEG feature sequence is used as a state node in the neural intention feature space. For two temporally adjacent state nodes, the dimensional change vector between the two state nodes is calculated. The Euclidean distance between the two state nodes in the feature space is calculated as the overall offset. The state nodes are sequentially connected along the time axis, and the dimensional change vector and the overall offset are used to form a directed edge feature to construct the initial topological representation of the neural intention evolution trajectory. The combined feature vectors corresponding to each state node in the initial topology representation are input into a feedforward transform network with nonlinear mapping capability. During the forward propagation of the feedforward transform network, the combined feature vectors of the current state node and the combined feature vectors corresponding to its neighboring nodes in the time neighborhood are weighted and fused to obtain the input feature vector. The input feature vector is nonlinearly transformed through multiple stacked fully connected layers, with each fully connected layer followed by a leaky linear rectified activation function to obtain a high-level semantic feature vector. The high-level semantic feature vectors of all time windows are adaptively aligned using iterative statistical normalization based on exponential moving average to obtain dynamic intent feature vectors. The dynamic intent feature vectors of each time window are arranged sequentially along the time axis to form a dynamic intent feature sequence.

5. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 4, characterized in that, Based on the dynamic intent feature sequence, a pre-trained deep learning classification model is used for decoding and recognition to obtain an initial neural intent probability distribution reflecting the current motion intent, including: The dynamic intention feature vectors corresponding to each time window in the dynamic intention feature sequence are sequentially input into a pre-trained deep learning classification model. The classification model consists of several fully connected hidden layers and a classification output layer stacked sequentially. Each fully connected hidden layer performs a linear affine transformation on the input vector and then maps it through a nonlinear activation function to obtain a hidden feature vector. The classification output layer performs a linear affine transformation on the hidden feature vector output by the last hidden layer to obtain an initial prediction score vector with a dimension equal to the total number of preset motion intention categories. Each component of the initial prediction score vector represents the original discrimination score of the EEG signal in the current time window belonging to each preset motion intention category. The initial predicted score vector is input into a preset probability normalization function. The preset probability normalization function performs an exponential operation on each component with the natural constant as the base, and calculates the ratio of the exponential value of each component to the sum of the exponential values ​​of all components to obtain the probability value corresponding to each preset motion intention category. The probability values ​​of all categories are arranged in order of category to form an initial neural intention probability distribution vector reflecting the current motion intention.

6. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 5, characterized in that, Determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; based on the initial neural intent probability distribution and the mechanical response delay duration, perform spatiotemporal coupling hysteresis compensation between intent and execution to obtain a dynamic temporal hysteresis compensation factor, including: The physical state parameters of the exoskeleton are collected in real time, including the current displacement, current motion speed, current output torque of the end effector, and current pressure value in the drive cavity. When the flexible exoskeleton is driven by pneumatic artificial muscles, the absolute value of the pressure difference between the target pressure value required for the exoskeleton to perform the target auxiliary action and the current pressure value in the drive cavity collected in real time is used to calculate the hysteresis time of the inflation and deflation process, based on the internal volume of the pneumatic drive cavity, the mass flow rate of the air supply pipeline, the specific gas constant of the working gas, and the absolute temperature of the working gas. The inflation and deflation hysteresis time is added to the pre-calibrated deformation recovery inherent response time of the flexible material to obtain the mechanical response delay time of the flexible exoskeleton. Using the start timestamp of the current time window as the benchmark for the moment of neural intent generation, the sum of the system processing time from inputting EEG data to outputting the initial neural intent probability distribution and the mechanical response delay is determined as the time deviation between intent generation and mechanical execution. Based on the time deviation and the preset time window sliding step size, the number of integer steps in units of the sliding step size is calculated and used as the dynamic timing hysteresis compensation factor for timing calibration.

7. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 6, characterized in that, Driving the flexible exoskeleton to perform corresponding auxiliary movements includes: During operation, an initial neural intent probability distribution cache queue covering the current time window and several previous time windows is continuously maintained. The cache queue stores the initial neural intent probability distribution vector corresponding to each time window in order of time window number. When generating the calibration instruction for the current time window, the initial neural intent probability distribution vector after shifting the temporal hysteresis compensation factor in the historical direction for a number of windows is extracted from the cache queue and used as the target intent probability distribution vector after the temporal calibration of the current time window. When the window number after shifting in the historical direction exceeds the range of the cache queue, the probability distribution vector at the initial moment is used as the default value. In the time-calibrated target intent probability distribution vector, the motion intent category corresponding to the component with the highest probability value is determined, and it is determined whether the highest probability value meets the preset confidence threshold. When the highest probability value is greater than or equal to the confidence threshold, the corresponding target control command is extracted from the mapping relationship according to the pre-established mapping relationship between motion intent category and control command, and sent to the control execution unit of the flexible exoskeleton to drive the exoskeleton to perform auxiliary motion that is time-synchronized with the patient's current motion intent. When the highest probability value is less than the confidence threshold, the target control command of the previous time window is maintained and continues to be executed.

8. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 7, characterized in that, The calculation process of the dynamic time-series hysteresis compensation factor is as follows: Calculate the ratio of the time deviation to the preset time window sliding step size, and round the ratio up to obtain the integer compensation step number in units of sliding step size. This integer compensation step number is the dynamic timing hysteresis compensation factor.

9. The brain-computer interface neural signal classification and recognition method based on deep learning according to claim 8, characterized in that, Based on the obtained initial EEG feature sequence arranged along the time axis, let the first... The combined feature vector corresponding to each time window is Combined feature vectors The dimension is D, where T represents the total number of time windows; for two temporally adjacent combined feature vectors, they are treated as feature states at two consecutive moments in the neural intent feature space, and the feature state transition relationship between them is calculated; wherein, the feature state transition relationship includes: Dimensionally changing vector ;in, Each dimensional component represents the neural intent feature state on the corresponding feature dimension from the first... The time window to the The net change over each time window, with positive values ​​indicating an increase in the strength of that dimension's feature and negative values ​​indicating a decrease, where... Indicates the first The combined feature vectors corresponding to each time window; Overall offset That is, to The sum of the squares of each component is then taken as the square root, which represents the Euclidean distance between the neural intention states at adjacent time points in the feature space. This distance is used to quantify the overall magnitude of the neurodynamic drift.

10. A brain-computer interface neural signal classification and recognition system based on deep learning, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The signal acquisition and segmentation module is used to acquire the raw EEG signals of stroke hemiplegic patients during continuous rehabilitation training. The raw EEG signals are used to control a flexible exoskeleton with physical response delay characteristics, and the raw EEG signals are divided into multiple continuous time window sequences. The feature extraction module is used to extract the spatial topological features and temporal features within each time window based on the time window sequence, so as to obtain the initial EEG feature sequence. The topology mapping module is used to perform spatiotemporal topology mapping of the neural intention evolution trajectory based on the initial EEG feature sequence, so as to obtain a dynamic intention feature sequence that characterizes the continuous evolution law of EEG signals. The decoding and recognition module is used to decode and recognize the dynamic intention feature sequence using a pre-trained deep learning classification model to obtain an initial neural intention probability distribution that reflects the current motion intention. The hysteresis compensation and calibration module is used to determine the mechanical response delay duration of the flexible exoskeleton with physical response delay characteristics; perform spatiotemporal coupling hysteresis compensation of intention and execution based on the initial neural intention probability distribution and the mechanical response delay duration to obtain a dynamic temporal hysteresis compensation factor; perform temporal calibration of the initial neural intention probability distribution based on the dynamic temporal hysteresis compensation factor to obtain a calibrated target control command; and drive the flexible exoskeleton to perform corresponding auxiliary movements based on the calibrated target control command.

Images