Exoskeleton neuromotor synergic control method based on electroencephalogram signals and system thereof

Abstract

The application relates to an exoskeleton neuromotor cooperative control method and system based on an electroencephalogram, and relates to the technical field of intelligent rehabilitation equipment. First, multi-channel sensing features are screened based on an inter-class separability index to obtain an optimal feature index table; then, an electroencephalogram compliance index is obtained by decoding real-time electroencephalogram, and a two-dimensional impedance regulation vector is mapped with a gait phase, and a compound tracking deviation is quantified by combining neural mismatch, position and interactive force residual; then, features are extracted according to the index table to evaluate human-machine compliance indicators, and a safety correction factor is generated in combination with individualized safety boundaries; finally, the two-dimensional impedance vector is used to convert the compound deviation into an original driving torque, and the safety correction factor is scaled and constrained to generate a final servo command and issue the final servo command to an exoskeleton joint. The application effectively balances the human-machine compliance and safety of system control.

Description

Technical Field

[0001] This application relates to the field of intelligent rehabilitation equipment technology, specifically to an exoskeleton neuromotor coordination control method and system based on electroencephalogram (EEG) signals. Background Technology

[0002] Central nervous system injury often leads to lower limb motor dysfunction. While exoskeletons can provide mechanical assistance, conventional products mostly use preset trajectories or pure force control, making it difficult to provide real-time feedback of the patient's cortical neural activity to the control circuit. Training often degenerates into passive dragging, resulting in insufficient stimulation of neural plasticity. The introduction of brain-computer interfaces offers a possibility to improve this situation; however, existing solutions still have many shortcomings in terms of signal utilization depth and physical interaction.

[0003] Current brain-controlled exoskeletons generally fail to fully utilize multi-channel spatial information in signal processing. Many schemes still rely on single-channel or a few-channel temporal amplitude thresholding, failing to deeply explore the multi-channel rhythmic characteristics of cortical neural activity. In fact, fluctuations in the human-machine coupling state cause energy changes in mu and beta rhythms in the sensorimotor cortex, and this spectral characteristic is coupled with physical interaction forces and joint movement trajectories. If this neuro-physical coupling characteristic is ignored, and thresholding is based solely on raw amplitudes or simple power spectra, the accuracy of compliance assessment is difficult to improve, and it is impossible to provide a continuous compliance index reflecting the human-machine coupling state, resulting in the inability to achieve a smooth transition in control gain.

[0004] In terms of physical interaction, existing human-machine collaboration strategies often focus only on single-dimensional compensation of kinematics or contact force, rarely incorporating the coupling relationship between cortical neural activity and electromechanical response into a closed loop. Even if the exoskeleton tracks position or force accurately, mechanical behavior may become disconnected from the patient's cortical state, resulting in neurophysical asynchrony, which weakens the human-machine collaboration effect of rehabilitation training.

[0005] Furthermore, existing systems employ rather crude methods to quantify human-machine compliance, typically using a direct instantaneous ratio of joint angle increments to torque increments without effectively mining and filtering high-dimensional dynamic features from multi-channel sensing timelines. This makes the evaluation results susceptible to noise, vibration, and individual differences. When a patient experiences a sudden spasm, the system often continues to output full torque, easily exceeding the biomechanical limits of the damaged joint and causing secondary injury. Therefore, a collaborative control architecture that integrates multi-channel EEG rhythm analysis, multi-sensor feature optimization, and compliance safety monitoring is urgently needed. Summary of the Invention

[0006] This application provides a method and system for exoskeleton neuromotor coordination control based on electroencephalogram (EEG) signals, in order to at least solve the above-mentioned technical problems existing in the prior art.

[0007] According to a first aspect of this application, an exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals is provided, comprising: S1, performing feature screening and weight solidification on each channel of an offline multi-channel sensing time series based on an inter-class separability index to obtain an optimal feature index table; S2, performing compliance decoding on the acquired real-time scalp EEG signals to obtain an EEG compliance index; S3, performing bivariate impedance interpolation mapping on the EEG compliance index and normalized gait phase to obtain a two-dimensional impedance control vector, and quantifying the neural mismatch residual between the EEG compliance index and the servo execution feedback quantity through a sliding window mechanism, and combining it with joint angle position tracking residual... S4. Based on the optimized feature index table, the real-time multi-channel sensor data is simplified and feature extraction is performed. The joint angle change rate and the actuator output torque change rate are combined for weighted scoring to obtain the human-machine compliance index. S5. The human-machine compliance index is compared with the individualized safety boundary interval to obtain the safety correction factor. S6. Based on the two-dimensional impedance control vector, the composite tracking deviation is converted into time-varying impedance torque to obtain the original driving torque. The original driving torque is scaled and constrained with the safety correction factor to form the final servo drive command and sent to the exoskeleton servo joint.

[0008] According to a second aspect of this application, an exoskeleton neuromotor coordination control system based on electroencephalogram (EEG) signals is provided for executing the aforementioned exoskeleton neuromotor coordination control method based on EEG signals, comprising: a feature preprocessing module for performing feature screening and weight solidification on a channel-by-channel offline multi-channel sensing time series based on an inter-class separability index to obtain an optimal feature index table; an EEG decoding module for performing compliance decoding on the acquired real-time scalp EEG signals to obtain an EEG compliance index; and an impedance mapping module for performing bivariate impedance interpolation mapping between the EEG compliance index and the normalized gait phase to obtain a two-dimensional impedance control vector, and quantifying the neural loss between the EEG compliance index and the servo execution feedback quantity through a sliding window mechanism. The system employs a composite tracking deviation calculation, which combines the joint angle position tracking residual with the human-machine interaction force residual. A compliance evaluation module extracts simplified features from real-time multi-channel sensor data based on a preferred feature index table and performs a weighted score based on the joint angle change rate and the actuator output torque change rate to obtain a human-machine compliance index. A safety discrimination module compares the human-machine compliance index with an individualized safety boundary interval to obtain a safety correction factor. A torque output module performs time-varying impedance torque conversion on the composite tracking deviation based on a two-dimensional impedance control vector to obtain the original driving torque. The original driving torque is then scaled and constrained using the safety correction factor to form the final servo drive command, which is then sent to the exoskeleton servo joint.

[0009] Compared with existing technologies, the exoskeleton neuromotor coordinated control method and system based on electroencephalogram (EEG) signals provided in this application have the following technical advantages: At the neural interface level, by performing in-depth extraction of multi-channel rhythmic features from sixteen-channel scalp EEG signals and continuous mapping of compliance assessment models, the system can output a smooth and normalized EEG compliance index. Compared with the traditional approach that relies on discrete threshold discrimination based on single-channel or a few-channel amplitudes, the resolution and temporal continuity of compliance assessment are significantly improved. As a result, the exoskeleton impedance parameters gain the ability to gradually adjust with the cortical state, eliminating torque mutations and gait discomfort caused by jumps in control gain between limited levels.

[0010] At the physical interaction level, the introduction of neural mismatch residuals enables the control loop to have the ability to explicitly perceive the degree of temporal disconnect between cortical intention and electromechanical response for the first time. The composite tracking deviation formed by the fusion of joint angle position tracking residuals and human-machine interaction force residuals simultaneously covers deviation information in three dimensions: kinematics, dynamics, and neurophysiology. Compared with the traditional impedance control scheme driven by a single physical residual, it can more accurately capture the real changes in human-machine collaborative state, thereby improving the accuracy of gait tracking and the compliance of human-machine interaction.

[0011] In terms of safety and defense, the hierarchical safety response mechanism, which is based on the simplified feature subset selected by the offline phase inter-class separability index and the safety boundary interval of individualized zero-time calibration, enables the system to release the full torque without damage to ensure training intensity when the patient is in a normal cooperative state. When stiffness resistance is detected, the driving gain is actively attenuated to avoid forced stretching damage to soft tissues. When the risk of instability and fall is identified, the active traction is quickly removed and high-damping locking is triggered to prevent falls. This effectively reduces the risk of secondary damage to the damaged joints during rehabilitation training.

[0012] At the sensor fusion level, the inter-class separability index obtained from offline data on different populations and different motor imagery tasks is used as a measure of feature importance. The most representative feature subsets of each channel are pre-selected and solidified into an index table. In the online stage, corresponding simplified features are directly extracted to construct a human-machine compliance index, which improves the problems of noise sensitivity and poor cross-individual generalization ability of the traditional instantaneous ratio method of angle increment and torque increment. At the closed-loop architecture level, by closing the neural semantic loop, motion tracking loop and physical compliance loop, the exoskeleton-assisted behavior is deeply coupled with the patient's cortical state, realizing multi-loop coordinated regulation from cortical neural activity to mechanical power output.

[0013] Furthermore, the feature screening and weight solidification completed in one step during the offline phase allow for simplified feature extraction and compliance scoring to be completed simply by looking up values ​​in the index table during the online control cycle. The entire online evaluation process involves only lightweight operations such as weighted summation and threshold comparison, which meets the computing power constraints of the embedded controller's millisecond-level real-time scheduling and ensures a low-latency closed-loop response from EEG acquisition to torque delivery. Attached Figure Description

[0014] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, wherein: in the drawings, the same or corresponding reference numerals denote the same or corresponding parts.

[0015] Figure 1 This is an overall flowchart of the exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in the exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to an embodiment of this application. Figure 3 This paper illustrates a sub-process diagram of step S2 in the exoskeleton neuromotor coordinated control method based on electroencephalogram (EEG) signals according to an embodiment of this application. Figure 4 This paper illustrates a sub-process diagram of step S3 in the exoskeleton neuromotor coordinated control method based on electroencephalogram (EEG) signals according to an embodiment of this application. Figure 5 The diagram shows a sub-flow diagram of steps S1, S4, and S5 in the exoskeleton neuromotor coordinated control method based on electroencephalogram signals according to an embodiment of this application. Figure 6 This is a block diagram of an exoskeleton neuromotor coordination control system based on electroencephalogram (EEG) signals according to an embodiment of this application. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by this application in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of this application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0017] Figure 1 The illustration shows a schematic flowchart of an exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to an embodiment of this application. Figure 1 As shown, this application provides a method for coordinated exoskeleton neuromotor control based on electroencephalogram (EEG) signals, including: S1, based on the inter-class separability index, feature filtering and weight solidification are performed channel-by-channel on the offline multi-channel sensing time series to obtain an optimal feature index table. The offline multi-channel sensing time series includes time-series data from joint encoders, torque sensors, surface electromyography (EMG), and inertial measurement units (IMU). The offline multi-channel sensing time series are collected from multiple groups of subjects of different ages and degrees of injury, and the collection tasks cover various motor imagery tasks (such as imagining the left leg lifting and the right leg stepping forward), to ensure that the subsequent feature filtering results can encode statistical information from different populations and different motor imagery tasks, enabling the solidified optimal feature index table to have cross-individual and cross-task generalization capabilities.

[0018] It is understandable that traditional human-machine compliance assessment methods typically use the instantaneous ratio of joint angle increment to torque increment as the criterion. This approach is highly sensitive to sensor noise, mechanical vibration, and inherent fluctuations in gait phase, and fails to effectively mine and filter features from high-dimensional time-series data generated by multi-source sensors. A large number of redundant or low-discrimination feature dimensions not only increase the online computational burden but also dilute truly discriminative information, resulting in poor generalization ability of the assessment results across different individuals. Therefore, this application systematically quantifies and sorts candidate features for discrimination in the offline stage, solidifying the filtering results into an index table. This allows for simplified feature extraction by simply looking up values ​​in the index table during the online control cycle, meeting the computational constraints of millisecond-level real-time scheduling for embedded controllers.

[0019] In one of the possible implementation methods of this application, such as Figure 5 As shown, step S1 includes: S11, performing window framing and multi-dimensional feature extraction on the offline multi-channel sensing time series through a sliding window of preset length and step size to obtain an initial high-dimensional feature matrix; S12, after performing feature standardization on the initial high-dimensional feature matrix to obtain a standardized candidate feature set, using pre-labeled compliance level labels as the grouping basis, performing inter-class separability quantification evaluation on each feature dimension in the standardized candidate feature set to obtain an inter-class separability index; S13, after arranging the inter-class separability indexes in descending order of their numerical values, independently extracting the first few feature indices in each channel, and normalizing and mapping the inter-class separability indices of the extracted features into weights and pairing them with the indices to obtain a preferred feature index table.

[0020] First, sub-step S11 is executed. The acquired offline multi-channel sensor time series is framed along the time axis using a preset fixed-length and stepped sliding window. The window length is 96 sampling points and the step size is 1 to obtain multi-channel time series frame samples. Within each sliding window, time-domain statistics and frequency-domain spectra are independently calculated for each sensor channel. Time-domain statistics cover at least 14 dimensions: mean, standard deviation, maximum, minimum, peak-to-peak value, root mean square, zero-crossing rate, waveform length, absolute mean, variance, skewness, kurtosis, range, and absolute mean. Frequency-domain spectra cover at least 6 dimensions: dominant frequency, spectral centroid, spectral entropy, band energy, spectral variance, and band power ratio. Taking an 8-channel sensor as an example, the original feature dimension of each sample is 8 × 20 = 160 dimensions. The extracted statistics and spectra of all channels are then concatenated along the channel dimensions to construct an initial high-dimensional feature matrix.

[0021] Next, execute sub-step S12. For each independent feature dimension in the initial high-dimensional feature matrix, iterate through all samples to calculate the global mean and global standard deviation of that dimension. Based on the calculated mean and standard deviation, perform Z-Score standardization on each feature value in the matrix to eliminate differences in physical dimensions across sensor devices. The standardized feature value is calculated by subtracting the global mean of that dimension from the original feature value and then dividing by the global standard deviation of that dimension. This mapping generates a standardized candidate feature set.

[0022] Then, based on the pre-labeled compliance level, the samples in the standardized candidate feature set are divided into different categories of state samples. For example, the first category of state samples corresponds to the normal cooperative state, and the second category corresponds to the rigid adversarial state. For each feature dimension, the mean and variance of the first category of state samples and the mean and variance of the second category of state samples are calculated independently. Based on this, the discriminative power of this dimension is quantified: the inter-class separability index of the j-th feature is equal to the square of the difference between the means of the first and second categories of samples on this feature dimension, divided by the sum of the variances of the first and second categories of samples. Specifically, the inter-class separability index borrows the mathematical form of Fisher's Discriminant Ratio from classic Fisher discriminant analysis to measure the ability of a single feature dimension to distinguish between different categories of samples: the larger the value, the more significant the inter-class difference and the stronger the intra-class clustering of the feature across different compliance levels, i.e., the stronger the discriminative power. Its calculation formula is:

[0023] in, Represents the inter-class separability index of the j-th feature dimension; and Let represent the mean values ​​of the first type of state sample and the second type of state sample on the j-th feature dimension, respectively; and Let $\mathbf{j}$ and $\mathbf{j}$ represent the standard deviations of the first and second class state samples on the j-th feature dimension, respectively, and their squares represent the variance of the corresponding class. This calculation is performed across all channels and feature dimensions, and the results are aggregated to obtain a set of inter-class separability indices reflecting the discriminative power of each dimension.

[0024] Finally, execute sub-step S13. Within each physical sensing channel, sort the data in descending order according to the value of the inter-class separability index. Based on the sorting result, independently extract the top K feature numbers within each channel to form a simplified subset. The value of K is determined based on the actual number of sensing channels and the number of feature dimensions; for example, K=10. Map the inter-class separability indices of the extracted K features to normalized weights: the normalized weight of the k-th extracted feature is equal to the feature's own inter-class separability index divided by the sum of the inter-class separability indices of all the top K preferred features extracted from the current channel. The calculation formula is:

[0025] in, This represents the normalized weight value obtained by the k-th preferred feature extracted under the current channel; This represents the original inter-class separability index corresponding to the k-th preferred feature extracted in the current channel; The threshold value represents the set descending order truncation threshold, indicating the number of features retained in each channel. The truncated simplified feature index numbers are key-value paired and structured with the corresponding normalized weights generated by the mapping, and then stored in the control unit's flash memory to generate the final preferred feature index table.

[0026] The independent sorting design within each channel respects the physical heterogeneity of different sensor channels and avoids erroneous comparisons of feature importance across channels. Since this sorting and truncation are completed offline in a single step, online control only requires looking up values ​​in the index table, without involving any model training or iterative optimization process, making it suitable for embedded real-time deployment. The aforementioned offline feature preprocessing process only needs to be executed once; statistical information covering different populations and different motion imagery tasks is already encoded in the generated preferred feature index table, eliminating the need to recalculate the inter-class separability index during online runtime.

[0027] S2 involves decoding the acquired real-time scalp EEG signals to obtain an EEG compliance index. The scalp EEG signal is 16-channel. It is understandable that existing brain-controlled exoskeleton solutions generally fail to fully utilize multi-channel spatial information in signal processing. Most solutions remain at the level of single-channel or a few-channel temporal amplitude threshold discrimination, lacking in-depth exploration of the multi-channel rhythmic characteristics of cortical neural activity. In fact, fluctuations in the human-machine coupling state cause energy changes in mu and beta rhythms in the sensorimotor cortex, and this spectral characteristic is coupled with physical interaction forces and joint movement trajectories. If this neuro-physical coupling characteristic is ignored, and threshold discrimination is based solely on the original amplitude or simple power spectrum, the accuracy of compliance assessment is difficult to improve, and a continuous compliance index reflecting the human-machine coupling state cannot be provided, resulting in an inability to achieve a smooth transition in control gain. Unlike conventional assessment methods that rely solely on physical sensors, this step introduces cortical neural activity information so that the impedance loop can progressively adjust stiffness based on EEG compliance, avoiding abrupt changes in mechanical assistance between high and low compliance.

[0028] In one of the possible implementation methods of this application, such as Figure 3 As shown, step S2 includes: S21, using a digital bandpass filter to perform baseline drift elimination and electromyography artifact suppression on the real-time scalp EEG signal to obtain a rhythmic frequency band time-series signal; S22, performing multi-dimensional rhythmic feature extraction and splicing on the rhythmic frequency band time-series signal to obtain a high-dimensional EEG feature vector; S23, feeding the high-dimensional EEG feature vector into a pre-configured compliance assessment model to obtain a preliminary compliance score; S24, based on the upper and lower benchmark values ​​of the score, applying extreme value normalization constraints to the preliminary compliance score to obtain an EEG compliance index.

[0029] First, sub-step S21 is executed. Raw EEG time-series signals are acquired at a sampling rate of 250Hz or higher using a 16-channel silver chloride sintered electrode array attached to the patient's scalp. The raw signals are then converted from analog to digital and fed into a digital filtering unit. A 5th-order Butterworth bandpass filter is used, with a lower passband limit of 8Hz and an upper passband limit of 30Hz, to perform baseline drift elimination and EMG artifact suppression on the real-time scalp EEG signals. Through filtering calculations, DC baseline drift and high-frequency noise from EMG artifacts caused by muscle activity are eliminated from the raw signals. The mu rhythm component and beta rhythm component, rich in sensorimotor cortex state information, are extracted, and the purified rhythmic frequency band time-series signal is output.

[0030] Next, sub-step S22 is executed. The filtered 16-channel time series is organized into a feature matrix according to the channel dimension. Within a preset time window, the band energy, spectral entropy, and band power ratio between the mu (8-13Hz) and beta (14-30Hz) bands are calculated independently for each channel. Band energy reflects the total power level of the signal within that band, spectral entropy quantifies the uniformity and complexity of the spectral distribution, and the band power ratio reflects the relative energy proportion between the mu and beta bands, which can characterize the dynamic balance between the active and inhibited states of the cortex. Based on this, the instantaneous phase between adjacent channels is extracted, and the phase synchronization index between adjacent channels is quantified by calculating the absolute value of the average value of the complex exponent of the instantaneous phase difference within the time window. The formula for calculating the phase synchronization index is:

[0031] in, This represents the phase synchronization index between the m-th and n-th adjacent channels; This represents the total number of discrete sampling points within the set feature extraction time window; This represents the time step index of the discrete sampling points within the current time window; The base of the natural logarithm; Represents the imaginary unit; and These represent the instantaneous phases of the m-th and n-th channels in the rhythmic frequency band timing signal at time t.

[0032] The frequency band energy, spectral entropy, and phase synchronization index between adjacent channels of all channels are flattened and stitched together along the channel dimension in a one-dimensional manner to generate a high-dimensional EEG feature vector reflecting the spatial and frequency domain coupling characteristics of the cortex. This feature vector comprehensively reflects the spectral characteristics and spatial coordination patterns of cortical neural activity and can serve as a neurophysiological marker of human-machine coupling compliance.

[0033] Next, proceed to sub-step S23. The high-dimensional EEG feature vector is fed into the pre-configured compliance assessment model. In the offline phase, this model utilizes the mean and variance distributions of direct statistical features from multiple groups of normal cooperative samples and abnormal adversarial samples, employing a weighted scoring method as the basic assessment approach. The inter-class separability index calculated in step S1 determines the weight of each feature, and the compliance level labeled by experts serves as the basis for sample grouping. The discrimination threshold is determined through direct statistical analysis of the score distributions of the two classes of samples. This process is based solely on the direct calculation of descriptive statistics and does not involve any model training or parameter optimization.

[0034] During online execution, the model utilizes pre-calibrated neural feature weights across various dimensions to perform weighted multiplication-addition operations on each feature component of the input vector. The initial compliance score is calculated using the following formula:

[0035] in, This represents the preliminary compliance score calculated. This represents the total number of feature dimensions in a high-dimensional EEG feature vector; This represents the dimension index item in the feature vector; This represents the neural pre-defined weight coefficients for the d-th dimension feature obtained through offline pre-calibration; This represents the specific feature value of the d-th dimension in the high-dimensional EEG feature vector. The model output is a preliminary compliance score reflecting the degree of coordination and fit between the current brain motor cortex command and mechanical assistance.

[0036] Finally, extreme value normalization constraints are applied. The initial compliance score is normalized using Min-Max constraints, mapping and constraining its values ​​to the [0,1] interval. The system retrieves the lower and upper boundary baseline scores obtained during the pre-wearing zero-time calibration phase and calculates the EEG compliance index using the following formula:

[0037] in, The EEG compliance index represents the final output constraint within the range of [0,1]. This represents the initial compliance score passed from the compliance model decoding and mapping stage; This represents the lower boundary baseline score of the initial compliance score calculated and determined by the system in the relaxed baseline state during the zero-time calibration phase; This represents the upper boundary baseline score of the initial compliance score calculated and determined by the system under a concentrated imagination state during the zero-time calibration phase; and These represent the minimum and maximum value functions for the elements within the parentheses, respectively. They are used in combination here to achieve hard saturation truncation of values ​​within the [0,1] interval. This processing outputs a smooth, continuous, and standardized EEG compliance index. This index directly reflects the compliance of the human-machine coupling at the current moment. The transition band is automatically determined by zero-time calibration before wearing the device, with a default value of approximately 0.3 to 0.7. The therapist can manually fine-tune the trigger boundaries to avoid abrupt changes in control gain.

[0038] S3 uses bivariate impedance interpolation to map the EEG compliance index and normalized gait phase to obtain a two-dimensional impedance control vector. A sliding window mechanism is used to quantify the neural mismatch residual between the EEG compliance index and the servo execution feedback, and the joint angle position tracking residual and human-machine interaction force residual are combined to obtain a composite tracking bias. It is understandable that traditional fixed impedance strategies struggle to simultaneously respond to cortical state fluctuations at the neural level and gait cycle evolution at the physical level; insufficient energy absorption is common during the stance phase, while excessive drag occurs during the swing phase. Meanwhile, existing human-machine collaborative control strategies typically focus only on single-dimensional compensation of kinematics or contact force, rarely incorporating the coupling relationship between cortical neural activity and electromechanical response into a closed loop. Even if the exoskeleton tracks accurately in position or force, mechanical behavior may become disconnected from the patient's cortical state, resulting in neurophysical astep loss, thus weakening the human-machine collaborative effect of rehabilitation training. Therefore, it is necessary to introduce a bivariate joint mapping of EEG compliance and gait phase in the impedance parameter generation stage, and to introduce a new dimension of neural mismatch residual in the deviation construction stage, so that the deviation signal on which the subsequent torque synthesis is based can more comprehensively reflect the real state of human-machine collaboration.

[0039] In one of the possible implementation methods of this application, such as Figure 4 As shown, step S3 includes: S31, using the EEG compliance index as the stiffness interpolation weight and the normalized gait phase as the damping interpolation variable, performing bivariate impedance parameter interpolation mapping to obtain a two-dimensional impedance control vector; S32, maintaining a preset length time window within the first-in-first-out buffer, calculating the absolute difference between the EEG compliance index and the servo execution feedback quantity at each sampling point, and taking the arithmetic mean of all absolute differences within the time window to obtain the neural mismatch residual; S33, based on the position weight, force weight, and neural mismatch weight read from the parameter library at the current rehabilitation stage, performing multi-source multi-dimensional residual weighted fusion of the joint angle position tracking residual, human-computer interaction force residual, and neural mismatch residual to obtain the composite tracking bias.

[0040] First, sub-step S31 is executed. This sub-step uses the EEG compliance index output from stage S2 and the normalized gait phase output from the gait detection unit as bivariate inputs to generate real-time stiffness coefficient and real-time damping coefficient, respectively.

[0041] In terms of stiffness mapping, the controller reads the preset minimum and maximum stiffness limits of the current joint from its non-volatile register, and performs linear interpolation to calculate the real-time stiffness coefficient using the EEG compliance index as the interpolation weight. The real-time stiffness coefficient is equal to the minimum stiffness limit plus the product of the EEG compliance index and the difference between the maximum and minimum stiffness limits.

[0042] In terms of damping mapping, the minimum and maximum damping limits are read from the same register, and linear interpolation is performed using the normalized gait phase as the interpolation variable to calculate the real-time damping coefficient. The real-time damping coefficient is equal to the minimum damping limit plus the product of the normalized gait phase and the difference between the maximum and minimum damping limits.

[0043] The calculation formula for the above bivariate impedance mapping is as follows:

[0044]

[0045] in, This represents the calculated real-time stiffness coefficient at the current moment; and These represent the minimum and maximum stiffness limits for the exoskeleton joints, respectively. The EEG compliance index, representing the output of stage S2, is constrained to a value within the range [0,1]. This represents the calculated real-time damping coefficient at the current moment. and These represent the minimum and maximum damping limits for the safety of exoskeleton joints, respectively. This represents the normalized gait phase, with values ​​constrained within the interval [0,1].

[0046] The physical significance of stiffness interpolation is as follows: when the patient's cortical activation level is low, the EEG compliance index approaches 0, and joint stiffness remains near the minimum stiffness limit, ensuring minimum anti-folding support; when the cortical activation level is high, the EEG compliance index approaches 1, and stiffness rises to the maximum stiffness limit, providing thrust for initiation or push-off. Damping automatically switches according to gait phase: after heel strike, the normalized gait phase during the support phase is lower, and damping is close to the maximum damping limit to absorb impact; after toe lift, the normalized gait phase during the swing phase is higher, and damping is close to the minimum damping limit to reduce hysteresis and promote forward swing. The specific values ​​of the minimum stiffness limit, maximum stiffness limit, minimum damping limit, and maximum damping limit are set and fixed by the therapist during initialization based on the patient's weight, muscle strength, and joint range of motion.

[0047] The calculated real-time stiffness coefficient and real-time damping coefficient are packaged into a two-dimensional impedance control vector for use in the impedance law torque synthesis stage of the subsequent S6 stage.

[0048] Then, sub-step S32 is executed. This sub-step quantifies the degree of temporal disconnect between the cortical neural intent represented by the EEG compliance index and the actual electromechanical response represented by the servo execution feedback.

[0049] Before calculating the neural mismatch residual, it is necessary to clarify the physical definition and normalization method of the servo execution feedback quantity. The servo execution feedback quantity is a normalized measure of the degree of performance of the servo joint's current trajectory tracking. Its physical essence is a complementary mapping value between the actual joint angle position sampled in real-time by the joint encoder and the desired reference trajectory, with values ​​constrained within the range [0,1]. Specifically, in each control cycle, the system reads the actual joint angle position output by the joint magnetic encoder and the desired reference angle position issued by the gait planner, calculates the absolute angle deviation between the two, and then compares this deviation with a preset maximum allowable angle deviation threshold. After complementary mapping and saturation truncation, the servo execution feedback quantity is obtained. The calculation formula is as follows:

[0050] in, Indicates the current time The servo execution feedback quantity is constrained to the range [0,1]. This indicates the actual joint angle position obtained in real time by the joint magnetic encoder; This represents the expected reference angle position at the current moment, as determined by the gait planner. This represents the preset maximum permissible angle deviation threshold, set by the therapist during the initialization phase based on the patient's joint range of motion and stored in the controller's non-volatile memory; for example, a value between 15° and 30°. When the actual angle position perfectly matches the desired angle position... This indicates that the servo joint has fully executed the motion command; when the angle deviation reaches or exceeds the maximum allowable threshold, This indicates that the servo joint failed to effectively execute motion commands.

[0051] It should be noted that after the above complementary mapping normalization, the servo execution feedback quantity and the EEG compliance index are uniformly mapped to the same semantic scale space: both are dimensionless scalars in the interval [0,1], and both follow a monotonic semantic direction where a higher value indicates a closer approximation to the ideal coordinated state. The EEG compliance index approaching 1 indicates that the cortical nerves are highly coordinated with the motor intention, and the servo execution feedback quantity approaching 1 indicates that the servo joints are highly compliant with the motor command. Therefore, the absolute difference between the two can directly measure the instantaneous disjointness between the degree of intention at the neural intention level and the degree of execution at the electromechanical execution level: when the patient's cortex is highly activated but the servo joint fails to keep up due to excessive load or driver delay, the EEG compliance index tends to be high while the servo execution feedback quantity tends to be low, and the difference increases, indicating a mismatch where the neural intention precedes the mechanical response; conversely, when the servo joint runs well on the preset trajectory but the patient's cortical activation level is low, the servo execution feedback quantity tends to be high while the EEG compliance index tends to be low, and the difference also increases, indicating a passive drag mismatch where the mechanical movement is separated from neural participation. The larger the difference between the two, the worse the neuro-physical synchronization of human-machine coupling.

[0052] A configurable time window, corresponding to a physical duration of 100 to 300 milliseconds, is maintained within a first-in, first-out (FIFO) buffer. For each sampling point within the window, the absolute difference between the EEG compliance index and the servo execution feedback normalized by the encoder is calculated. Then, the arithmetic mean of the absolute differences across all moments within the time window is taken to obtain the neural mismatch residual at the current moment. The calculation formula is as follows:

[0053] in, This represents the calculated neural mismatch residual at the current moment; This represents the total sampling point length configuration value within the first-in-first-out sliding time window, for example, at a sampling rate of 250Hz. This corresponds to a window duration of 200 milliseconds; Indicates the index of the local time sampling step within the sliding time window; This represents the historical value of the EEG compliance index at the kth local sampling point; This represents the historical value of the servo execution feedback quantity after range normalization at the k-th local sampling point.

[0054] A larger neural mismatch residual indicates a worse synchronization between central commands and mechanical responses, which may be caused by delayed neural signals, abrupt changes in muscle tone, or driver delay. This residual transforms the degree of disconnect between cortical neural activity and electromechanical execution from a latent state into an explicit and quantifiable deviation component, providing a neurophysiological input for subsequent multidimensional deviation fusion.

[0055] Finally, sub-step S33 is executed. This sub-step linearly weights and sums the joint angle position tracking residual, the human-computer interaction force residual, and the neural mismatch residual calculated in the previous step according to preset weights, and outputs the composite tracking bias.

[0056] Based on the current stage of rehabilitation training, the system retrieves pre-set normalized weighted coefficients from the parameter database for three types of residuals: position weight, force weight, and neural mismatch weight. In the early stages of passive training, the focus is on position tracking, with configurable position weights of 0.6, force weights of 0.3, and neural mismatch weights of 0.1. In the later stages of active rehabilitation, the focus is on neural pathways and compliant interaction, with configurable position weights of 0.3, force weights of 0.3, and neural mismatch weights of 0.4. The sum of all coefficients is equal to 1 and none is less than 0.1 to prevent any single loop from completely failing.

[0057] The formula for calculating the composite tracking error is:

[0058] in, This represents the composite tracking bias in the final fused output; This represents a predefined normalized weighting coefficient for positional deviation, retrieved from the parameter library, with a value ranging from 0.3 to 0.6. This represents the joint angle position tracking residual obtained by real-time sampling from the joint encoder; This represents a predefined normalized weighting coefficient for torque deviation, retrieved from the parameter library, with a value ranging from 0.2 to 0.4. This represents the human-computer interaction force residual obtained from real-time sampling by the force sensor; This represents a predefined normalized weighting coefficient for neural mismatch bias, retrieved from the parameter library, with a value ranging from 0.1 to 0.5.

[0059] This composite tracking bias simultaneously carries deviation information from three dimensions: kinematics, dynamics, and neurophysiology. Compared to a single physical residual, it better reflects the human-machine collaborative state and will be transmitted to the S6 stage as the bias input for time-varying impedance law calculation.

[0060] S4. Based on the optimized feature index table, simplified feature extraction is performed on real-time multi-channel sensor data, and a weighted score is obtained by combining the joint angle change rate and the actuator output torque change rate. It should be understood that traditional compliance assessment methods directly use the instantaneous ratio of joint angle increment to torque increment as the criterion. This approach is susceptible to sensor noise interference and struggles to distinguish between normal gait phase fluctuations and real human-machine interaction. By performing online simplified feature extraction based on a feature subset selected after sorting by the inter-class separability index in the offline stage, redundant dimensions are eliminated while retaining highly discriminative feature information. This ensures that the online evaluation process only involves index lookup and weighted summation, meeting the computational constraints of millisecond-level real-time scheduling of the embedded controller, while simultaneously improving the accuracy of anomaly detection and cross-individual generalization ability.

[0061] In one possible implementation of this application, step S4 includes: S41, based on the feature numbers and dimension indexes of each channel fixed in the preferred feature index table, extracting the feature values ​​of the corresponding dimension from the real-time multi-channel sensing data channel by channel through direct memory access to obtain a simplified sensing feature subset; S42, concatenating the simplified sensing feature subset with the joint angle change rate and the actuator output torque change rate to obtain a joint description vector, and using the normalized weights in the preferred feature index table to perform a weighted summation of each component of the joint description vector to obtain the human-machine compliance index.

[0062] First, sub-step S41 is executed. The system reads the preferred feature index table, which was fixed in stage S1, from the flash memory of the control unit, and parses the Top-K feature numbers and dimension indices of each sensing channel recorded within it. Without any complex transformations, the system directly extracts the feature values ​​of the corresponding dimensions from the real-time multi-channel sensing data channel by channel through memory access, eliminating all non-critical features to obtain a simplified subset of sensing features. Taking an 8-channel sensor with 10 features extracted from each channel (i.e., K=10) as an example, the simplified subset of sensing features has a dimension of 80, which reduces the computational load by 50% compared to the original 160-dimensional feature space, while retaining the feature combinations with the highest discriminative power. Since this extraction process only involves index lookup and value reading, there is no need to re-execute the inter-class separability index calculation or any iterative optimization process, making it suitable for completion within the millisecond-level control cycle of the embedded controller.

[0063] Next, sub-step S42 is executed. The system acquires the joint angle change rate from the differential output of the joint encoder and the driver output torque change rate from the differential output of the torque sensor. These two physical change rates, reflecting transient electromechanical dynamics, are combined with the feature components in the simplified sensor feature subset extracted in the previous step, and then concatenated along the vector dimension to construct a complete joint description vector. Taking an 8-channel sensor with 10 features extracted from each channel as an example, the total dimension of the joint description vector is 82 dimensions (80 dimensions of simplified sensor features plus 1 dimension each for the joint angle change rate and the driver output torque change rate). This joint description vector simultaneously covers the static statistical features of multiple sensor sources and the transient dynamic features of joint motion and driver output, enabling a more comprehensive characterization of the human-machine interaction state within the current control cycle.

[0064] Finally, the system extracts the normalized weight sequence corresponding to each simplified sensing feature from the preferred feature index table, and performs corresponding multiplication and summation on each Z-score-normalized feature component in the joint description vector to calculate the human-machine compliance index. The calculation formula is as follows:

[0065] in, This indicates the human-machine compliance index calculated and output in this step; This represents the total number of feature dimensions contained in the concatenated joint description vector; Indicates the feature dimension index number in the joint description vector; This represents the normalized weight value corresponding to the j-th feature, which is obtained from offline solidification in stage S1. This represents the specific value of the j-th dimension feature in the joint description vector after Z-Score normalization, including the simplified sensor feature values ​​and the spliced ​​joint angle change rate and actuator output torque change rate.

[0066] The evaluation process employs a weighted scoring method as its foundation. The compliance score is obtained by weighting and summing the components of the joint descriptive vector using normalized weights (obtained from the inter-class separability index) pre-calculated and fixed in stage S1. In the offline stage, based on expert-annotated samples at levels such as normal coordination, mild aggression, rigidity / spasticity, and instability / fatigue, the discrimination threshold is determined by statistically analyzing the score distribution for each level. The entire process involves only the direct calculation of sample statistics and threshold setting, without any model training or parameter iteration. The model input dimension is determined by the number of Top-K features, the number of channels, and two physical quantities. The output is either a discrete level or a continuous score based on threshold division.

[0067] The human-machine compliance index reflects the smoothness of human-machine interaction within the current control cycle. When this index is within the normal range, it indicates that the interaction between the patient and the exoskeleton is in a safe and cooperative state; when the index deviates from the normal range, it may mean that the patient has experienced abnormal conditions such as muscle spasms, sudden changes in muscle tone, or sensor drift. This index will be passed to the S5 stage for comparison and discrimination against the individualized safety boundary interval to generate the corresponding safety correction factor.

[0068] S5 compares the human-machine compliance index with the individualized safety boundary range to determine the safety correction factor. It is understandable that existing exoskeleton systems rely on rather crude methods for monitoring human-machine compliance safety, lacking a tiered safety response mechanism based on individual differences. When a patient experiences sudden muscle spasms or abrupt changes in muscle tone during training, the system continues to output driving torque at full gain, easily exceeding the biomechanical limits of the damaged joint and causing secondary soft tissue damage. Therefore, it is necessary to establish a tiered safety discrimination mechanism based on individualized thresholds, enabling the system to automatically execute different levels of torque constraint strategies according to real-time changes in the human-machine compliance index, reducing the risk of injury under abnormal conditions while ensuring normal training intensity.

[0069] Specifically, the first step is to perform zero-time calibration of individualized safety boundary intervals. This individualized zero-time calibration procedure is a short-term online parameter calibration process performed on the current wearer before the formal start of each rehabilitation training session (i.e., at the zero point of the training timeline). Its purpose is to measure and determine, in real time, the individual's specific safety threshold based on their current physiological state and the actual wearing conditions of the sensors before training begins. The lower and upper safety thresholds are not fixed constants but are determined online through the individualized zero-time calibration procedure before wearing, eliminating threshold deviations caused by individual differences, daily state fluctuations, and sensor displacement. "Zero-time" emphasizes that this calibration is performed at the zero point before each training session begins, ensuring that the thresholds reflect the patient's true state at that time; "individualized" emphasizes that the thresholds are determined statistically by the patient's own relaxation and imagined state data, rather than using fixed empirical values ​​across groups.

[0070] The first stage of the calibration process is baseline acquisition in a relaxed state. After donning the exoskeleton, the trainee maintains a seated or standing relaxed position while the system continuously acquires 16-channel EEG signals and multi-sensor data for 60 to 120 seconds. A simplified subset of sensor features is extracted according to the preferred feature index table determined in stage S1. Then, following the same weighted scoring process as stage S4, a comprehensive compliance score for each frame is calculated. The compliance index sequence for all frames is recorded, and the average relaxed state value is calculated. Standard deviation of relaxed state .

[0071] The second stage of the calibration process is the collection of focused imagery. The therapist guides the trainee through a display screen or voice commands to perform a motor imagery task consistent with subsequent rehabilitation training (such as imagining the left leg lifting, the right leg stepping forward, etc.). Each imagery session lasts 4 to 6 seconds, with a 2 to 3-second rest interval, repeated 10 to 20 times. All sensor data are collected simultaneously, and the mean of the imagery is calculated using the same procedure. Standard deviation of imagined state .

[0072] The third stage of the calibration process is threshold calculation. The initial discrimination threshold is automatically calculated using the two-state Gaussian error rate criterion. The derivation of this criterion is based on the following assumption: the relaxed-state compliance index follows a certain pattern. For the mean, The imagined compliance index follows a Gaussian distribution with standard deviation of 100%. For the mean, The standard deviation is a Gaussian distribution. Let the two-state probability density functions be equal at the threshold. After taking the logarithm and ignoring the quadratic term, the initial discrimination threshold is obtained. Its calculation formula is:

[0073] in, Indicates the initial discrimination threshold; This represents the mean of the relaxation compliance index series; This represents the mean of the imagined compliance index series; This represents the standard deviation of the relaxation compliance index series; This represents the standard deviation of the imagined compliance index sequence. The physical meaning of this formula is to make the false positive rate (misclassification of the relaxed state as abnormal) approximately equal to the false negative rate (missed judgment in the imagined state), thereby minimizing the total discrimination error rate.

[0074] The pooled standard deviation is then calculated, and the lower and upper safety thresholds are set using the pooled standard deviation as a scaling factor. The formula for calculating the pooled standard deviation is:

[0075] in, This represents the pooled standard deviation; and represents the standard deviation of the compliance index series for relaxed state and imagined state, respectively.

[0076] The formulas for calculating the lower safety threshold and the upper safety threshold are as follows:

[0077]

[0078] in, Indicates the lower safety threshold; Indicates the upper safety boundary threshold; Indicates the initial discrimination threshold; and These represent the configurable safety margin coefficients, with the default value being... The therapist can adjust the value within the range of 0.5 to 3.0 based on the patient's risk level.

[0079] The fourth stage of the calibration process is manual fine-tuning. The system visualizes the distribution of compliance indices for relaxed and imagined states using a bar chart on the human-computer interface, and uses vertical lines to indicate the automatically calculated lower and upper safe thresholds. Therapists can apply offsets to the lower safe threshold by touching sliders or inputting numbers. Apply an offset to the upper safety boundary threshold. Personalized fine-tuning is performed, and the final threshold is then stored in the controller's non-volatile memory. This zero-time calibration process takes approximately 3 to 5 minutes and is automatically prompted to be executed each time the device is worn.

[0080] Then, the human-machine compliance index is compared with the lower and upper boundary thresholds of the individualized safety boundary interval. A corresponding gain scaling value is assigned based on three criteria: falling within the interval, below the lower boundary, or above the upper boundary, to obtain a safety correction factor. Within each online control cycle, the system reads the safety lower boundary threshold determined by the individualized zero-time calibration procedure from the controller's non-volatile memory. With the upper boundary threshold of safety The real-time human-machine compliance index output in the S4 stage will be used to... and safety boundary interval Numerical comparisons are performed, and corresponding security correction factors are assigned based on the three discrimination results. The segmented discrimination logic is as follows:

[0081] in, This represents the safety correction factor output after the decision-making process. This represents the human-machine compliance index transmitted from stage S4; The lower boundary threshold representing the individualized safety boundary interval; The upper boundary threshold of the individualized safety boundary interval is represented by the safety threshold. This represents the stiffness attenuation coefficient preset in the system's internal non-volatile memory for use in rigid resistance states, with a value ranging from 0.4 to 0.7, for example, 0.5; This represents the minimum sustaining value used for unstable drop states. It takes the value of 0 or a positive minimum close to 0, for example, 0.05.

[0082] When the human-machine compliance index falls within the safe boundary range, it is determined to be in a safe cooperative state, and the safety correction factor is set to 1.0, allowing the original driving torque to proceed without attenuation. When the human-machine compliance index is below the lower safe boundary threshold, it is determined to be in a rigid resistance state, indicating that the unit torque is insufficient to drive effective angular displacement, and the patient's affected side may be rigid or strongly resisting. The safety correction factor is set to a preset stiffness attenuation coefficient (e.g., 0.5), actively reducing mechanical gain to prevent forced stretching damage to soft tissue. When the human-machine compliance index is above the upper safe boundary threshold, it is determined to be in an unstable fall state, indicating that the angular displacement under torque is abnormally large, possibly due to loss of muscle tone, tremor, or sensor drift. The safety correction factor is set to a minimum maintenance value (e.g., 0 or 0.05), and an emergency locking command is sent to the servo actuator, causing the joint to enter a high-damping braking mode, providing rigid support for a short time to prevent falls.

[0083] The safety correction factor will be passed to stage S6 to scale the original driving torque in a multiplicative manner, forming a hard boundary constraint on the theoretical torque.

[0084] S6, based on a two-dimensional impedance modulation vector, performs time-varying impedance torque conversion on the composite tracking deviation to obtain the original driving torque. Then, a safety correction factor is used to scale and constrain the original driving torque to form the final servo drive command, which is then sent to the exoskeleton servo joints. It should be understood that in traditional exoskeleton control schemes, the impedance parameter remains constant throughout the gait cycle, the deviation signal contains only a single kinematic or mechanical dimension, and there is a lack of a torque constraint mechanism based on individualized safety judgments. This results in the control output being unable to simultaneously adapt to cortical state fluctuations, multidimensional deviation changes, and safety defense requirements. This step uses time-varying impedance law calculations to apply the impedance parameter, which is dually modulated by EEG compliance and gait phase, to the composite tracking deviation that incorporates three-dimensional deviation information. Then, a safety correction factor is used for multiplicative scaling constraints, so that the final output driving torque can be adjusted in real time according to the patient's physiological state and can automatically execute graded gain constraints under abnormal conditions.

[0085] In one possible implementation of this application, step S6 includes: S61, determining the rate of change of the composite deviation based on the current discrete control cycle sampled value of the composite tracking deviation and the historical value of the previous control cycle in the system cache; S62, performing torque synthesis based on the real-time stiffness coefficient, composite tracking deviation, real-time damping coefficient and rate of change of the composite deviation to obtain the original driving torque; S63, performing safety boundary calibration and servo command dispatch on the original driving torque based on the safety correction factor to obtain the final servo drive command.

[0086] First, execute sub-step S61. Extract the current discrete control cycle sampled value of the composite tracking error and the historical value of the previous control cycle from the system cache, and calculate the composite error rate of change using the first-order backward difference method. The difference between the current value and the historical value, divided by the fixed control cycle time of the system, yields an approximate value of the composite error rate of change. The calculation formula is:

[0087] in, This represents the rate of change of the composite deviation calculated using the first-order backward difference approximation, which is the derivative of the composite tracking deviation with respect to time. This represents the composite tracking deviation value within the current k-th discrete control cycle; This represents the historical value of the composite tracking deviation cached in the system register during the previous (i.e., the (k-1)th) discrete control cycle; This represents the fixed discrete control cycle time of the embedded control system scheduling, for example, 1 to 4 milliseconds. In this embodiment, 2 milliseconds corresponds to a 500Hz scheduling frequency.

[0088] Next, execute sub-step S62. Extract the real-time stiffness coefficient and real-time damping coefficient from the two-dimensional impedance control vector issued in stage S3. Multiply the real-time stiffness coefficient by the composite tracking deviation to obtain the elastic term, which provides a restoring torque proportional to the deviation. Multiply the real-time damping coefficient by the composite deviation change rate to obtain the damping term, which is used to suppress deviation oscillations. Superimpose the elastic and damping terms to obtain the original driving torque.

[0089] In the first embodiment of this application, the time-varying impedance control law treats the elastic recovery term and the damping suppression term as two independent linear channels algebraically superimposed. That is, the original driving torque equals the product of the real-time stiffness coefficient and the composite tracking deviation, plus the product of the real-time damping coefficient and the rate of change of the composite deviation. The calculation formula is as follows:

[0090] in, This represents the raw driving torque of the synthesized output, which has not been determined by safety boundaries. This represents the real-time stiffness coefficient, modulated by the EEG compliance index, derived from the two-dimensional impedance modulation vector. This represents the composite tracking deviation for the current control cycle; This represents the real-time damping coefficient, which is analyzed from the two-dimensional impedance control vector and modulated by the normalized gait phase. This represents the rate of change of the composite deviation calculated in the previous step. Elasticity term. Provides a restoring torque proportional to the deviation, damping term Suppressing deviation oscillations, the two work together to ensure tracking accuracy and smoothness.

[0091] In the closed-loop control architecture of exoskeleton-assisted rehabilitation walking, sub-step S62 undertakes the core computational task of converting abstract impedance parameters into physical driving torque. The time-varying impedance control law it employs treats the elastic recovery term and damping suppression term as two independent linear channels, algebraically superimposed. That is, the original driving torque equals the product of the real-time stiffness coefficient and the composite tracking deviation, plus the product of the real-time damping coefficient and the composite deviation change rate. This linear superposition assumption is reasonable within the normal gait range where the deviation amplitude and deviation change rate are at a high level. However, in several special transitional intervals within the gait cycle, such as the instantaneous push-off transition from the end of the support phase to the beginning of the swing phase, it can lead to severe degradation in control quality.

[0092] Specifically, when the exoskeleton joint is in the aforementioned gait transition zone, the affected limb has not yet completed its push-off motion, causing the joint position to deviate significantly from the desired trajectory, and the composite tracking deviation is at its peak level. Simultaneously, the joint is rapidly releasing elastic potential energy to drive forward swing, and the rate of change of the composite deviation also climbs to an extremely high level. Under this condition, the elastic recovery term generates a strong position-correcting torque due to the large deviation, while the damping inhibition term simultaneously applies a large viscous braking torque due to its high rate of change. These two torques highly overlap in the time domain and may have opposing directions. However, the original linear superposition mechanism is completely unaware of this situation. It is neither aware that the stiffness channel is outputting a high-amplitude recovery torque, nor that the damping channel is applying a high-amplitude braking torque, and it cannot assess whether the superposition effect formed by the outputs of the two channels at the physical level exceeds the compliance tolerance range of the human joint.

[0093] This neglected nonlinear coupling cross-effect between stiffness and damping responses leads to three specific problems at the control level. First, torque impact in the gait transition zone: when both deviation and rate of change are high, the combined amplitude of the two independently superimposed forces is much greater than the torque level required for a smooth transition, resulting in spikes in the joint output torque. For spastic patients with abnormal muscle tone, this sudden rigid impact can easily trigger pathological stretch reflexes, not only disrupting the continuity of rehabilitation training but also potentially causing secondary damage to the damaged soft tissues. Second, damping redundancy under low-deviation, high-speed conditions: when the patient's active cooperation is high and the joint position deviation is close to zero but the movement speed is still relatively fast (typically, in the free forward swing phase during the mid-swing), the elastic recovery term naturally decays due to the near-zero deviation, but the damping inhibition term still outputs a high-amplitude viscous braking torque at full coefficient. This artificially applies unnecessary movement resistance at a stage where the patient hardly needs position correction, reducing the smoothness and naturalness of walking. Third, lack of phase margin: linear superposition cannot characterize the relative activity between the deviation amplitude and the rate of change. The controller lacks a smooth gain transition mechanism between the two extreme conditions mentioned above, and torque oscillations are prone to occur at specific phases of the gait cycle, which weakens walking stability.

[0094] To address the structural defect of missing coupling cross-effect between the elastic recovery channel and the damping suppression channel, a second embodiment is proposed. In the second embodiment of this application, step S62 includes: S621, determining the deviation cross-activity index based on the composite tracking deviation and the composite deviation change rate; S622, performing complementary attenuation mapping on the deviation cross-activity index to obtain the stiffness coupling weight and the damping coupling weight; S623, multiplying the stiffness coupling weight by the product of the real-time stiffness coefficient and the composite tracking deviation to obtain the coupling elastic term, multiplying the damping coupling weight by the product of the real-time damping coefficient and the composite deviation change rate to obtain the coupling damping term, and superimposing the coupling elastic term and the coupling damping term to obtain the original driving torque.

[0095] First, sub-step S621 is executed. In the first embodiment, the composite tracking deviation and the composite deviation rate of change are directly fed into their respective multiplication channels, with no information exchange between them. To enable the torque synthesis stage to sense the simultaneous activity of the composite tracking deviation and the composite deviation rate of change within the current control cycle, a dimensionless metric capable of capturing their coordinated activation state needs to be constructed. The absolute values ​​of the composite tracking deviation and the composite deviation rate of change are taken to eliminate the interference of the direction sign on the energy level assessment, resulting in a deviation amplitude scalar and a rate of change amplitude scalar. Then, the product of the two is used as the numerator, and the square root of the sum of their squares, normalized by a scaling factor, is used as the denominator to calculate the deviation cross-activity index. The calculation formula is as follows:

[0096] in, is the deviation cross-activity index, a dimensionless scalar with a value range of [0,1); This represents the composite tracking deviation for the current control cycle. The rate of change of the composite deviation; A normalized scaling factor, consistent with the dimensions of joint angular displacement, is preset in the controller parameter area and is used to map the denominator to a dimensionless space. To prevent division by zero and protect extremely small positive numbers, for example, take... .

[0097] The numerator employs a product structure to ensure that a high exponent value is output only when both the deviation and the rate of change are significant. If either term approaches zero, regardless of the magnitude of the other, the product result will also approach zero, and the deviation cross-activity index will naturally decay, indicating that no coupling correction is needed. Euclidean norm normalization of the denominator eliminates the influence of absolute amplitude and dimensions, making the exponent a relative measure reflecting the proportion of coordinated activity between the two energy channels. In exoskeleton gait control scenarios, when the patient is in the transitional transient from the end of the support phase to the beginning of the swing phase, both joint deviation and the rate of change peak simultaneously, and the deviation cross-activity index rises to its highest level, driving subsequent weights to perform maximum gain decay. In the stable motion range during the middle of the swing phase, where the deviation approaches zero, the exponent automatically falls back, subsequent weights return to full amplitude, and torque synthesis degenerates into a classical linear impedance law, preserving the original tracking accuracy.

[0098] Then, sub-step S622 is executed. After obtaining the deviation cross-activity index, it needs to be converted into effective gain adjustment quantities acting on the elastic recovery channel and the damping suppression channel, respectively. When both deviation and rate of change are active in the gait transition interval, the restoring torque strength of the stiffness term needs to be reduced simultaneously to avoid rigid impact, while the viscous braking of the damping term needs to be suppressed to avoid excessive drag; while in the steady motion interval, both weights should approach their full amplitude values ​​to maintain the original control performance without interference. Based on the deviation cross-activity index, stiffness coupling weights and damping coupling weights are generated through complementary linear decay mapping. The calculation formula is as follows:

[0099]

[0100] in, This represents the stiffness coupling weight, with a value range of (0,1]. This represents the stiffness attenuation sensitivity coefficient, with a value range of (0,1]. It is preset by the therapist based on the patient's degree of spasticity and the passive range of motion of the joint. The larger the value, the more severe the stiffness attenuation in the transition range. For example, it can be 0.7 to 0.9 for patients with severe spasticity and 0.3 to 0.5 for patients with mild spasticity. This represents the damping coupling weight, with a value range of (0,1]. This represents the damping attenuation sensitivity coefficient, with a value range of (0,1]. The larger the value, the more severe the damping attenuation in the transition range. For example, for patients with weak muscle strength who need to retain damping support, a value of 0.2 to 0.4 can be used, while for patients with better muscle strength recovery, a value of 0.5 to 0.8 can be used. This represents the deviation cross-activity index.

[0101] In exoskeleton rehabilitation training scenarios, the physical effect of stiffness coupling weights is equivalent to connecting a variable compliance element dependent on cross-activity in parallel to a virtual joint spring. When both deviation and rate of change are active, the compliance element intervenes to soften the spring and absorb impact; when cross-activity decreases, the compliance element disengages, restoring the spring to its original stiffness. The physical effect of damping coupling weights is equivalent to adding a bypass relief valve with combined velocity-displacement sensing to the virtual damper. When both displacement and velocity are large, the bypass valve opens to reduce equivalent damping and promote natural forward swing of the joint; when both are no longer active simultaneously, the bypass valve closes to restore the original damping level. Sensitivity coefficient. and The independently adjustable design allows therapists to differentiate configurations for different injury types: for patients with more severe spasticity, the size can be increased. To enhance stiffness attenuation and prevent stretch reflex activation, the stiffness can be appropriately reduced for patients with weaker muscles. This is to maintain a certain level of damping support to prevent joint instability.

[0102] Finally, sub-step S623 is executed. In the first embodiment, the elastic recovery term and the damping suppression term each output torque components with a fixed full gain and are directly linearly superimposed, without a gain coordination mechanism between the two channels. After the processing in steps one and two, the stiffness coupling weight and the damping coupling weight have encoded the instantaneous cross-active state between the deviation and the rate of change. At this point, these two weights need to be injected into the corresponding torque generation channels to achieve coupled weighted synthesis.

[0103] The coupling elastic term is obtained by multiplying the stiffness coupling weight by the product of the real-time stiffness coefficient and the composite tracking deviation, and the coupling damping term is obtained by multiplying the damping coupling weight by the product of the real-time damping coefficient and the rate of change of the composite deviation. The original driving torque is obtained by superimposing the two terms.

[0104] in, The original driving torque is synthesized after coupling weighting; For stiffness coupling weights; The real-time stiffness coefficient is generated by the bivariate impedance mapping in step three; This is a composite tracking deviation; For damping coupling weights; The real-time damping coefficient is generated by the bivariate impedance mapping in step three. This represents the rate of change of the composite deviation.

[0105] The second embodiment of this application maintains the linear superposition framework of the impedance law in its mathematical structure, when the deviation cross-activity index When approaching zero, and When both approaches 1, the above equation completely degenerates into the original linear impedance control law, and the tracking accuracy and existing control performance are not disturbed in any way; when When the deviation and rate of change are both active during gait transition, the two coupling weights decrease synchronously, the effective gains of the elastic and damping terms are synergistically reduced, and the peak amplitude of the output torque is flexibly constrained. This multiplicative weight injection method ensures that the interface between the improved synthesis result and the downstream safety correction factor scaling stage is fully compatible, without changing the data flow topology of subsequent torque calibration and servo command encoding. In actual exoskeleton operation, the joint torque exhibits an adaptive compliant characteristic of high cross-activity and low gain output throughout the entire gait cycle. The transient torque impact at the end of the support phase is actively softened, the viscous drag during the free forward swing phase in the middle of the swing is timely released, and the control accuracy in the smooth tracking range is fully preserved. The entire improvement only involves the software algorithm upgrade of the embedded controller, without adding additional sensors or hardware computing modules, and without changing the communication protocol and bus topology.

[0106] By introducing a deviation-rate-change cross-modulation factor and injecting it into the elastic recovery and damping inhibition channels in the form of multiplicative coupling weights, the improved torque synthesis mechanism subjectes the peak amplitude of the exoskeleton joint's output torque in the gait transition zone to an adaptive attenuation constraint proportional to the cross-activity level. This effectively suppresses the rigid impact caused by the independent superposition of deviation and rate of change when both are at high levels, reducing the risk of inducing pathological stretch reflexes in spastic patients and the probability of secondary damage to damaged soft tissues. Simultaneously, in the stable movement zone where deviation and rate of change are active, the coupling weights automatically revert to full amplitude, and torque synthesis completely degenerates into a classical linear impedance law, ensuring that tracking accuracy and original control performance are maintained without loss. Independently adjustable stiffness attenuation sensitivity coefficients and damping attenuation sensitivity coefficients provide therapists with the ability to configure differentiated compliance characteristics for different injury types and rehabilitation stages. This allows the exoskeleton's assistive torque to balance the safety and compliance of the transition zone with the tracking accuracy of the stable zone throughout the entire gait cycle, achieving a unity of on-demand compliant assistance and active safety protection.

[0107] Regardless of whether the torque synthesis method of the first embodiment or the second embodiment is used, the subsequent processing procedure is the same.

[0108] Next, execute sub-step S63. Multiply the original driving torque by the safety correction factor output in stage S5 to achieve a hard boundary constraint on the theoretical torque. The calculation formula is as follows:

[0109] in, This indicates the target torque value after safety constraint adjudication; This represents the safety correction factor output by the S5 stage, and its value is a dynamic scaling factor between 0 and 1. This represents the original driving torque. This multiplication, in a physical sense, constitutes a hard boundary constraint on the theoretical torque: in the safe cooperative state... , Full torque output; in a rigid, adversarial state Choose an attenuation constant in the range of 0.4 to 0.7 (e.g., 0.5). The mechanical gain is reduced accordingly to prevent forced stretching damage to soft tissues; in the unstable drop state... Set to 0 or a very small value close to 0 (e.g., 0.05). As the force approaches zero, the active traction is quickly removed, and an emergency locking command is sent to the servo drive, causing the joint to enter a high-damping braking mode, providing rigid support to prevent falls in a short time.

[0110] Ultimately, the target torque value after safety constraint adjudication will be... The torque constant mapping model of the internal current loop of the servo driver is used to convert it into a target current value of the three-phase winding that can be executed by the inverter. This target current sequence is formatted and encoded into a bus communication protocol frame, and the output generates the final servo drive command, which is sent to the exoskeleton servo joint to drive the motor to output torque and guide the affected limb to complete gait movements.

[0111] During continuous operation of the control loop, the joint magnetic encoder and thin-film force sensor are constantly sampling, capturing the actual joint angle and interaction force in real time. These physical quantities are transmitted back to S3 via the bus to update the joint angle position tracking residual and the human-machine interaction force residual; the actual output state of the servo driver is also fed back to S3 to update the servo execution feedback quantity and calculate the neural mismatch residual for the next cycle. This multi-loop feedback of neuro-motor-force enables the exoskeleton's mechanical behavior to continuously track the patient's cortical state and quickly trigger safety protection when the human-machine state changes abruptly, forming a complete neuro-motor coordinated control closed loop.

[0112] The following is a specific implementation example. In an online rehabilitation training program for patients with lower limb motor dysfunction after stroke, the system's control cycle is set to 2 milliseconds (corresponding to a 500Hz scheduling frequency) to ensure that the end-to-end delay of the torque transmission from EEG acquisition does not exceed 8 milliseconds. Within the current control cycle, the system extracts the current value of the composite tracking deviation and the historical value from the previous cycle, and calculates the rate of change of the composite deviation by dividing it by the 2-millisecond control cycle using the first-order backward difference method. Subsequently, the real-time stiffness coefficient and real-time damping coefficient are analyzed from the two-dimensional impedance control vector issued in stage S3.

[0113] In the first embodiment, the system multiplies the real-time stiffness coefficient by the composite tracking deviation to obtain the elastic term, multiplies the real-time damping coefficient by the composite deviation change rate to obtain the damping term, and directly linearly superimposes the two terms to obtain the original driving torque.

[0114] In the second embodiment, the system first takes the absolute values ​​of the composite tracking deviation and the rate of change of the composite deviation, and calculates the deviation cross-activity index. Assuming the current state is a transitional transient from the end of the support phase to the beginning of the swing phase, with both the deviation and the rate of change at relatively high levels, the calculated deviation cross-activity index is 0.6. The system uses the stiffness attenuation sensitivity coefficient... Calculate stiffness coupling weights With damping attenuation sensitivity coefficient Calculate the damping coupling weight Subsequently, the stiffness coupling weight is multiplied by the product of the real-time stiffness coefficient and the composite tracking deviation to obtain the coupled elastic term, and the damping coupling weight is multiplied by the product of the real-time damping coefficient and the composite deviation change rate to obtain the coupled damping term. The two terms are then superimposed to obtain the original driving torque. Compared to the full linear superposition in the first embodiment, the effective gain of the elastic term is reduced to 52%, and the effective gain of the damping term is reduced to 70%, with the torque peak amplitude subject to flexible constraints. When the patient enters the stable motion range in the middle of the swing, the deviation approaches zero, the deviation cross-activity index automatically falls back to near 0, and both the stiffness coupling weight and the damping coupling weight return to near 1.0. The torque synthesis degenerates into the classical linear impedance law, and the tracking accuracy is fully preserved.

[0115] Regardless of the implementation method, the system multiplies the original drive torque by the safety correction factor output in stage S5. The safety correction factor is 1.0 (safety cooperative state) in the current control cycle, and the torque is output at full value. The target torque value after safety constraints is mapped to a three-phase winding current command through the servo driver's internal current loop, encoded into a dual-channel isolated CAN bus communication protocol frame, and then sent to the exoskeleton knee joint servo motor. The CAN bus uses a custom redundancy check frame format; if no response is received within 10 microseconds for a single frame of data, it is automatically retransmitted, suppressing the packet loss rate to within a certain range. Below the specified magnitude, communication reliability is ensured. The aforementioned control strategy can be deployed on an integrated embedded main control board, simultaneously connecting the servo drive nodes of the hip, knee, and ankle joints via a dual-channel isolated CAN bus to form a distributed real-time control network. EEG signal data is transmitted from the portable EEG amplifier to the main control board via a Bluetooth Low Energy link, ensuring both wearability and real-time data transmission requirements. After receiving commands, the servo motor drives the joint to complete the movement of the current control cycle. The joint encoder and force sensor synchronously sample the actual rotation angle and interaction force and transmit them back to S3, providing updated data for the deviation calculation of the next control cycle, forming a continuously operating closed-loop control.

[0116] To verify the effectiveness of the above method, a 4-week clinical controlled trial was conducted: Ten patients with lower limb motor dysfunction after stroke were recruited and randomly divided into a traditional fixed impedance control group and the neuro-coordination experimental group proposed in this application. The control group used constant gain PID position control, while the experimental group used the aforementioned multi-channel EEG rhythm analysis-feature optimization-compliance assessment closed-loop strategy. Assessment indicators included the root mean square value of surface electromyography (EMG) of the affected lower limb (reflecting active participation), gait symmetry index, peak human-computer interaction force, and patient subjective comfort visual analog scale (VAS) score. The results showed that the experimental group exhibited a 28.6% increase in the root mean square value of EMG, a 19.3% improvement in the gait symmetry index, a 35.2% reduction in the peak abnormal interaction force, and a 32% increase in the subjective comfort score. These data indicate that the neuro-motor coordination control method proposed in this application can significantly improve walking safety and human-computer interaction compliance while stimulating patient active participation.

[0117] In an extended implementation paradigm for spinal cord injury trainees, researchers expanded the frequency band for EEG rhythm feature extraction from 8-30Hz to 4-35Hz for users with incomplete T10 segment injury resulting in bilateral lower limb muscle strength grade 2. This was to incorporate low-frequency components in the theta rhythm that might be related to changes in cortical state. Simultaneously, they increased the inter-channel coherence analysis dimension to capture more subtle spatial coordination differences. Furthermore, the compliance assessment model's input layer incorporated temporal features of the plantar pressure center trajectory. Pressure distribution in different plantar regions was obtained using a force-sensitive resistor array. In offline preprocessing, the top 8 most discriminative pressure features were selected after sorting by inter-class separability index and then stored in an index table. Experimental results showed that after introducing plantar pressure features, the compliance assessment model's sensitivity in recognizing foot drop gait improved by 22.4%, and the frequency of false triggering of emergency locking decreased from 1.8 times per thousand steps to 0.4 times, significantly improving the wearer's walking continuity experience.

[0118] In another adaptation program targeting Parkinson's disease trainees, considering their resting tremor and difficulty initiating movement, the system expanded the frequency band for extracting EEG rhythm features from 8-30Hz to 4-35Hz, incorporating low-frequency components in theta rhythms that may be related to changes in cortical state. A tremor detection bypass was added to the compliance assessment module: when multi-channel sensor FFT analysis showed a significant energy peak in the 4-6Hz frequency band that persisted for more than three gait cycles, the system temporarily relaxed the upper boundary threshold of the human-machine compliance index by 30%, preventing frequent safety downgrades caused by physiological tremors. Preliminary trials showed that this adaptation strategy increased the average continuous walking distance for Parkinson's trainees from 45 meters to 82 meters, without any falls due to tremor misjudgment.

[0119] In summary, the exoskeleton neuromotor coordination control method based on EEG signals according to the embodiments of this application is elucidated. It establishes a closed-loop regulation pathway from cortical neural activity to exoskeleton servo joint driving torque by incorporating continuous compliance decoding of multi-channel EEG rhythm features, offline optimization of multi-source sensor features based on inter-class separability index, construction of multi-dimensional composite bias fused with neural mismatch residuals, and calibration of graded correction factors based on individualized safety boundaries into a unified time-varying impedance control architecture. This enables the exoskeleton to adaptively adjust joint compliance characteristics according to the real-time changes in the patient's cortical activation level and gait phase to achieve on-demand assistance throughout the gait cycle. It can also automatically trigger graded gain constraints to achieve active safety protection when the human-machine compliance exceeds the safety envelope. This solves the technical problems in the prior art, such as insufficient depth of EEG signal utilization leading to unsmooth transition of control gain, single dimension of physical interaction closed loop leading to the inability to perceive and compensate for loss of agility in neural intent and electromechanical response, and coarse compliance assessment without individualized graded safety mechanisms leading to the risk of secondary injury to patients.

[0120] Figure 6 This is a block diagram of an exoskeleton neuromotor coordination control system based on electroencephalogram (EEG) signals according to an embodiment of this application, such as... Figure 6 As shown, the exoskeleton neuromotor collaborative control system 600 based on EEG signals includes: a feature preprocessing module 610, used to perform feature screening and weight solidification on the offline multi-channel sensing time series channel by channel based on the inter-class separability index to obtain an optimal feature index table; an EEG decoding module 620, used to perform compliance decoding on the acquired real-time scalp EEG signals to obtain an EEG compliance index; and an impedance mapping module 630, used to perform bivariate impedance interpolation mapping on the EEG compliance index and normalized gait phase to obtain a two-dimensional impedance control vector, and quantify the neural mismatch residual between the EEG compliance index and the servo execution feedback quantity through a sliding window mechanism, and combine the joint angle position tracking residual with the human-machine interface. The interaction force residual is used to obtain the composite tracking deviation; the compliance evaluation module 640 is used to extract simplified features from real-time multi-channel sensor data based on the preferred feature index table, and to perform weighted scoring by combining the joint angle change rate and the actuator output torque change rate to obtain the human-machine compliance index; the safety discrimination module 650 is used to compare the human-machine compliance index with the individualized safety boundary interval to obtain the safety correction factor; the torque output module 660 is used to perform time-varying impedance torque conversion on the composite tracking deviation based on the two-dimensional impedance control vector to obtain the original driving torque, and to scale and constrain the original driving torque with the safety correction factor to form the final servo drive command and send it to the exoskeleton servo joint.

[0121] Here, those skilled in the art will understand that the specific operations of each step in the above-described exoskeleton neuromotor coordination control system based on electroencephalogram (EEG) signals have been referenced above. Figures 1 to 5 The description of the exoskeleton neuromotor coordination control method based on EEG signals is detailed here, and therefore, its repeated description will be omitted.

[0122] The terms "first," "second," etc., are used only to distinguish different technical features and do not imply their importance or number. The designation of a feature as "first" or "second" does not preclude the possibility of additional similar features. The scope of protection of this application shall be determined by the claims; any modifications, substitutions, or combinations that do not depart from the core idea of ​​this application shall fall within the scope of protection of this application.

[0123] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has been disclosed above with reference to preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A method for coordinated exoskeleton neuromotor control based on electroencephalogram (EEG) signals, characterized in that, include: S1. Based on the inter-class separability index, feature filtering and weight solidification are performed on each channel of the offline multi-channel sensing time series to obtain the preferred feature index table. S2, perform compliance decoding on the acquired real-time scalp EEG signals to obtain the EEG compliance index; S3, a two-dimensional impedance control vector is obtained by bivariate impedance interpolation mapping of EEG compliance index and normalized gait phase, and the neural mismatch residual between EEG compliance index and servo execution feedback is quantified by sliding window mechanism, and the joint angle position tracking residual and human-computer interaction force residual are combined to obtain composite tracking bias. S4. Based on the preferred feature index table, the real-time multi-channel sensing data is simplified feature extraction, and the joint angle change rate and the actuator output torque change rate are combined to perform a weighted score to obtain the human-machine compliance index. S5, compares the human-machine compliance index with the individualized safety boundary interval to obtain the safety correction factor; S6 performs time-varying impedance torque conversion on the composite tracking deviation based on the two-dimensional impedance control vector to obtain the original driving torque, and applies scaling constraints on the original driving torque with a safety correction factor to form the final servo drive command and sends it to the exoskeleton servo joint.

2. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, The offline multi-channel sensing time series includes timing data from joint encoders, torque sensors, surface electromyography (EMG) devices, and inertial measurement units.

3. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 2, characterized in that, Step S1 includes: S11, by using a sliding window of preset length and step, window framing and multi-dimensional feature extraction are performed on the offline multi-channel sensing time series to obtain an initial high-dimensional feature matrix; S12, after performing feature standardization on the initial high-dimensional feature matrix to obtain a standardized candidate feature set, the inter-class separability of each feature dimension in the standardized candidate feature set is quantitatively evaluated based on the pre-labeled compliance level labels to obtain the inter-class separability index. S13. After sorting the features in descending order of their inter-class separability indices, the first few feature indices are extracted independently within each channel. The inter-class separability indices of the extracted features are normalized and mapped to weights and paired with the indices to obtain the preferred feature index table.

4. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S2 includes: S21, through a digital bandpass filter, baseline drift elimination and electromyography artifact suppression are performed on the real-time scalp EEG signal to obtain the rhythmic frequency band timing signal; S22, Multidimensional rhythmic feature extraction and splicing of rhythmic frequency band time signals to obtain high-dimensional EEG feature vectors; S23, The high-dimensional EEG feature vector is fed into a pre-configured compliance assessment model to obtain a preliminary compliance score; S24. Based on the upper and lower benchmark values ​​of the score, extreme value normalization constraints are applied to the preliminary compliance score to obtain the EEG compliance index.

5. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S3 includes: S31, using the EEG compliance index as the stiffness interpolation weight and the normalized gait phase as the damping interpolation variable, performs bivariate impedance parameter interpolation mapping to obtain a two-dimensional impedance control vector. S32, within the first-in-first-out buffer, maintain a time window of a preset length, calculate the absolute difference between the EEG compliance index and the servo execution feedback value for each sampling point, and take the arithmetic mean of all the absolute differences within the time window to obtain the neural mismatch residual. S33, based on the position weight, force weight and neural mismatch weight read from the parameter library at the current rehabilitation stage, performs multi-source multi-dimensional residual weighted fusion of joint angle position tracking residual, human-computer interaction force residual and neural mismatch residual to obtain composite tracking bias.

6. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S4 includes: S41, based on the channel feature numbers and dimension indexes fixed in the preferred feature index table, extract the corresponding dimension feature values ​​from the real-time multi-channel sensor data one by one through the direct memory access addressing method to obtain a simplified sensor feature subset. S42, the simplified sensor feature subset is concatenated with the joint angle change rate and the actuator output torque change rate to obtain a joint description vector, and the normalized weights in the preferred feature index table are used to perform a weighted summation of each component of the joint description vector to obtain the human-machine compliance index.

7. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S5 includes: comparing the human-machine compliance index with the lower boundary threshold and upper boundary threshold of the individualized safety boundary interval, wherein, based on the three discrimination results of falling within the interval, below the lower boundary, or above the upper boundary, corresponding gain scaling values ​​are assigned to obtain the safety correction factor.

8. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S6 includes: S61, Based on the current discrete control cycle sampled value of the composite tracking deviation and the historical value of the previous control cycle in the system cache, determine the composite deviation change rate; S62, based on the real-time stiffness coefficient, composite tracking deviation, real-time damping coefficient and composite deviation change rate, torque synthesis is performed to obtain the original driving torque; S63, based on the safety correction factor, performs safety boundary calibration and servo command dispatch on the original driving torque to obtain the final servo drive command.

9. The exoskeleton neuromotor coordination control method based on electroencephalogram (EEG) signals according to claim 8, characterized in that, Step S62 includes: S621, Based on the composite tracking deviation and the composite deviation change rate, determine the deviation cross-activity index; S622, perform complementary attenuation mapping on the deviation cross-activity index to obtain stiffness coupling weight and damping coupling weight; S623, the coupling elastic term is obtained by multiplying the stiffness coupling weight by the product of the real-time stiffness coefficient and the composite tracking deviation, and the coupling damping term is obtained by multiplying the damping coupling weight by the product of the real-time damping coefficient and the composite deviation change rate. The coupling elastic term and the coupling damping term are superimposed to obtain the original driving torque.

10. An exoskeleton neuromotor coordination control system based on electroencephalogram (EEG) signals, used to execute the exoskeleton neuromotor coordination control method based on EEG signals according to any one of claims 1-9, characterized in that, include: The feature preprocessing module is used to perform feature filtering and weight solidification on the offline multi-channel sensing time series based on the inter-class separability index to obtain the preferred feature index table. The EEG decoding module is used to decode the acquired real-time scalp EEG signals to obtain the EEG compliance index. The impedance mapping module is used to perform bivariate impedance interpolation mapping between the EEG compliance index and the normalized gait phase to obtain a two-dimensional impedance control vector. It also quantifies the neural mismatch residual between the EEG compliance index and the servo execution feedback through a sliding window mechanism, and combines the joint angle position tracking residual and the human-computer interaction force residual to obtain the composite tracking bias. The compliance assessment module is used to extract simplified features from real-time multi-channel sensor data based on the preferred feature index table, and to perform a weighted score by combining the joint angle change rate and the actuator output torque change rate to obtain the human-machine compliance index. The safety discrimination module is used to compare the human-machine compliance index with the individualized safety boundary interval to obtain the safety correction factor; The torque output module is used to perform time-varying impedance torque conversion on the composite tracking deviation based on the two-dimensional impedance control vector to obtain the original driving torque, and to scale and constrain the original driving torque with a safety correction factor to form the final servo drive command and send it to the exoskeleton servo joint.

Images