Limb rehabilitation training system and method of brain-computer interface

Abstract

The application relates to the technical field of brain-computer interfaces and motor rehabilitation, in particular to a limb rehabilitation training system and method of a brain-computer interface. The system comprises a brain-computer interface device for collecting brain electrical data of a patient, a kinematic sensor for collecting motion data of a healthy limb, an electromyographic sensor for collecting electromyographic data of the healthy limb, a wearable device for assisting the motion of the affected limb and returning execution feedback, and a computer device for processing data; the computer synchronously fuses the brain electrical data, the motion data of the healthy limb and the electromyographic data to generate a control instruction to drive the wearable device, and dynamically adjusts training parameters according to the feedback of the affected limb. Thus, the problems of low motion intention recognition accuracy and poor rehabilitation assistance adaptability in the prior art are solved.

Description

Technical Field

[0001] This application relates to the fields of brain-computer interface and motor rehabilitation technology, and in particular to a brain-computer interface limb rehabilitation training system and method. Background Technology

[0002] In related technologies, brain-computer interface rehabilitation relies on interpreting motor intentions from a single EEG signal. This is susceptible to interference from environmental and physiological noise, resulting in low accuracy in motor intention recognition and affecting the control reliability of rehabilitation equipment. Conventional rehabilitation equipment often uses fixed parameter settings for assistive force, making it difficult to keep up with real-time changes in the actual movement state of the limbs, further reducing the adaptation effect of rehabilitation training for hemiplegic upper limbs. Summary of the Invention

[0003] This application provides a brain-computer interface-based limb rehabilitation training system and method to solve problems such as low accuracy of motor intention recognition and poor adaptability of rehabilitation aids in related technologies.

[0004] The first aspect of this application provides a brain-computer interface-based limb rehabilitation training system, comprising: a brain-computer interface device for collecting electroencephalogram (EEG) data of a rehabilitation target; a kinematic sensor for collecting motion data of the healthy limb of the rehabilitation target; an electromyography (EMG) sensor for collecting EMG data of the healthy limb of the rehabilitation target; a wearable device for assisting the unhealthy limb of the rehabilitation target in performing training movements and collecting feedback data during the execution of training movements by the unhealthy limb of the rehabilitation target; and a computer device for simultaneously acquiring EEG data, motion data, and EMG data, generating motion control commands based on the EEG data, motion data, and EMG data, controlling the wearable device based on the motion control commands, and adjusting training movements based on feedback data.

[0005] Based on the aforementioned technical means, this application embodiment simultaneously integrates multi-source data from EEG, healthy limb movement, and healthy limb electromyography to generate control commands. It leverages measured physiological information from the healthy limb to compensate for the vulnerability of single EEG data to noise interference, effectively improving the accuracy of movement intention recognition and ensuring that the output movement control commands better align with actual human movement patterns. Simultaneously, it dynamically adjusts training parameters based on non-healthy limb movement data fed back from wearable devices, abandoning fixed parameter control modes. This allows the device's assistance intensity and electrical stimulation strategy to adapt to the patient's real-time movement state, avoiding training drawbacks caused by excessive or insufficient assistance intensity, and optimizing the individualized rehabilitation training effect for hemiplegic patients' upper limbs.

[0006] Optionally, the wearable device includes an exoskeleton device and an electrode device, wherein the exoskeleton device responds to motion control commands to perform corresponding actions to drive the movement of the unhealthy limb; and the electrode device responds to motion control commands to output stimulation signals to stimulate the unhealthy limb.

[0007] Based on the above technical means, the embodiments of this application utilize the dual driving methods of mechanical traction of the exoskeleton device and electrical stimulation of the electrode device to assist the unhealthy limb in completing the target movement from two dimensions: mechanical external force and muscle electrical activation. The two forms of assistance work together to improve the reliability of limb movement completion.

[0008] Optionally, the computer device is also used to collect electromyographic data of the healthy limb of the rehabilitation target performing the target action through a kinematic sensor before generating motor control instructions based on the electroencephalogram (EEG) data, motion data, and electromyographic (EMG) data, and to register the stimulation position of the electrode device on the unhealthy limb based on the EMG data.

[0009] Based on the above technical means, the embodiments of this application rely on the electromyographic data corresponding to the movements of the healthy side limbs to complete the electrode device point registration, and determine the stimulation position of the unhealthy side limbs according to the symmetry law of human limbs, so as to ensure that the electrical stimulation can accurately act on the target muscle group and improve the effectiveness of electrical stimulation-assisted movements.

[0010] Optionally, the computer device is also used to control the brain-computer interface device, kinematic sensors, and electromyographic sensors to synchronously collect data during the rehabilitation target's execution of training tasks before generating motor control instructions based on EEG data, motion data, and electromyographic data. A training dataset is constructed based on the synchronously collected data, and a mapping model from the motor EEG characteristics of the healthy limb to the target motor parameters of the unhealthy limb is constructed based on the training dataset. The training task includes: the rehabilitation target's healthy limb performing the target action, while the rehabilitation target's unhealthy limb performs motor imagination and action attempts of the target action.

[0011] Based on the above technical means, this application embodiment uses multiple types of data collected simultaneously from the practical movements of the healthy side limbs and the simultaneous imagination and trial movements of the unhealthy side limbs to build a training dataset, establish a feature mapping model, optimize the accuracy of motion parameter solution based on actual test samples, and further improve the accuracy of subsequent motion control command generation.

[0012] Optionally, the computer device is also used to extract EEG data and motion data from the training dataset, fuse the EEG data and motion data into motion EEG features of the healthy limb, decode and identify the motion intention of the rehabilitation target based on the EEG data, determine the motion type based on the motion intention, extract limb motion features from the motion data, determine the target motion parameters corresponding to the motion type based on the limb motion features, and construct a mapping model based on the motion EEG features and the target motion parameters.

[0013] Based on the above technical means, the embodiments of this application integrate EEG data and motion data to form motion EEG features, combine motion intention and limb motion parameters to build a mapping model, and rely on multi-dimensional features to optimize the model fitting effect and improve the accuracy of motion parameter estimation.

[0014] Optionally, the computer device is also used to fuse EEG data and motion data into motor EEG characteristics of the healthy limb, determine target motion parameters of the unhealthy limb based on the motor EEG characteristics and mapping model of the healthy limb, generate first control parameters of the exoskeleton device based on the target motion parameters, generate second control parameters of the electrode device based on electromyography data, and generate motion control commands based on the first and second control parameters.

[0015] Based on the above technical means, this application embodiment obtains motor EEG characteristics by fusing EEG data and motion data of the healthy side limb, calculates the target motion parameters corresponding to the unhealthy side limb by combining a pre-constructed mapping model, generates mechanical control parameters for the exoskeleton device based on the target motion parameters, and generates electrical stimulation control parameters for the electrode device based on real-time electromyography data. The two types of control parameters are integrated into a unified motion control command, realizing precise synchronization and coordinated regulation of exoskeleton mechanical assistance and electrical stimulation muscle activation, effectively improving the standardization and matching degree of rehabilitation training movements.

[0016] Optionally, the exoskeleton device incorporates at least one of an angle sensor and a torque sensor, wherein the angle sensor acquires the joint angle and joint angular velocity of the unhealthy limb, and the torque sensor acquires the output torque at the unhealthy limb.

[0017] Based on the above-mentioned technical means, the embodiments of this application can collect motion state information such as joint angle, joint angular velocity and output torque of the non-healthy limb in real time and accurately through the angle sensor and torque sensor built into the exoskeleton device, comprehensively capture the actual motion details of the limb, and provide real and reliable data support for subsequent dynamic adjustment of rehabilitation training parameters and optimization of motion control commands, so as to ensure the accuracy and real-time nature of rehabilitation regulation.

[0018] Optionally, the kinematic sensor includes at least one of an image sensor and an inertial measurement unit. The image sensor acquires motion images of the healthy limb, and the computer device identifies the joint angles, end-effector displacement trajectories, movement speeds, and movement durations of the healthy limb based on the motion images. The inertial measurement unit acquires the acceleration and angular velocity of the unhealthy limb.

[0019] Based on the above-mentioned technical means, the embodiments of this application use image sensors and inertial measurement units to collect motion state data of healthy and unhealthy limbs in multiple dimensions. This enables comprehensive and accurate acquisition of a wealth of motion parameters such as limb joint angles, displacement trajectories, motion speed, acceleration, and angular velocity, enriching the dimensions of data collection and providing a comprehensive and effective data foundation for subsequent multi-source data fusion, motion intention recognition, and accurate generation of rehabilitation control commands.

[0020] Optionally, the motor coverage area of ​​the brain-computer interface device includes at least one of the left and right hemisphere motor cortex, the prefrontal cortex, and the parietal cortex, and the electromyographic data includes at least one of electromyographic amplitude, envelope, integral electromyography, activation sequence, and co-contraction characteristics.

[0021] Based on the above-mentioned technical means, this application embodiment can accurately capture the motor nerve signals of the rehabilitation target by collecting signals from multiple motor-related cortices of the brain. At the same time, it can extract electromyographic features such as electromyographic amplitude, integral electromyography, and activation sequence from multiple dimensions, enriching the signal feature dimensions at the nerve and muscle levels, effectively improving the comprehensiveness and accuracy of motor intention recognition, and providing sufficient and effective data support for the subsequent accurate generation of rehabilitation motor control commands.

[0022] The second aspect of this application provides a control method for a brain-computer interface-based limb rehabilitation training system. The method is applied to the computer device of the brain-computer interface-based limb rehabilitation training system described in the above embodiment. The method includes the following steps: synchronously acquiring EEG data, motion data, and EMG data of the rehabilitation target; generating motion control commands based on the EEG data, motion data, and EMG data; controlling a wearable device to assist the unhealthy limb of the rehabilitation target in performing training movements based on the motion control commands; and adjusting the training movements based on feedback data from the wearable device.

[0023] Based on the aforementioned technical means, this application embodiment achieves comprehensive synchronous acquisition of neural signals, limb movement signals, and muscle activity signals by simultaneously collecting EEG data, motion data, and electromyography data of the rehabilitation target. It generates precise and reliable motion control commands through multi-source data fusion, uniformly regulating wearable devices to assist the unhealthy limb in completing training movements. Simultaneously, it dynamically adjusts training movements based on real-time feedback data from the device, achieving closed-loop rehabilitation training regulation. This effectively improves the problem of insufficient accuracy in single signal recognition, significantly enhancing the accuracy, adaptability, and intelligence of limb rehabilitation training.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is an architectural diagram of a brain-computer interface-based limb rehabilitation training system provided according to an embodiment of this application; Figure 2 This is a flowchart of a multi-source information fusion method based on brain-body symmetry provided in an embodiment of this application; Figure 3This is a flowchart of a control method for a brain-computer interface-based limb rehabilitation training system provided according to an embodiment of this application. Detailed Implementation

[0026] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0027] Currently, BCI (Brain-Computer Interface), a human-computer interaction technology that can directly read thought information through brain electrical activity, is widely used in the field of sports rehabilitation. It can analyze the movement intentions of patients with movement disorders and, combined with muscle stimulation and exoskeleton devices, assist patients in completing limb training movements. Through progressive training, it helps patients gradually recover limb motor abilities and self-care abilities, demonstrating broad application prospects. Related rehabilitation training equipment has gradually integrated BCI technology. Simultaneously, some rehabilitation programs utilize computer vision technology to capture the movements of the patient's healthy limbs, constructing a virtual training environment through mirroring to enhance the immersion and movement guidance effect of rehabilitation training.

[0028] The relevant technologies have been initially applied in the limb rehabilitation training of hemiplegic patients, but there are still some shortcomings: Firstly, most brain-computer interface rehabilitation programs rely on a single EEG signal to analyze motor intentions, which is highly susceptible to environmental noise interference and suffers from low accuracy in motor intention recognition.

[0029] Secondly, the relevant technologies make low use of information from the healthy side of the limbs, remaining only at a superficial level of visual mapping and motion trajectory reference. They fail to effectively apply the kinematic parameters and electromyographic signals of the healthy side of the limbs to the real-time decoding of the movement intentions of the unhealthy side of the limbs and the real-time control of movement parameters, resulting in extremely low information utilization.

[0030] Third, the relevant rehabilitation techniques do not accurately utilize the temporal correlation information between EEG signals and muscle contraction, which cannot guarantee the accuracy of rehabilitation control.

[0031] Fourth, the electrical stimulation intensity and exoskeleton assistance of related rehabilitation systems are mostly fixed parameter levels or rely on manual experience for adjustment. They lack a closed-loop control mechanism that dynamically adjusts the assistance intensity based on the patient's real-time performance, resulting in poor individual and stage adaptability of training.

[0032] In summary, the neural regulation of the left and right sides of the human brain and the motor control of the limbs possess natural structural and functional symmetry, which can provide an effective physiological reference basis for the rehabilitation training of hemiplegic limbs. However, related technologies have not effectively explored and utilized this physiological characteristic, relying on single EEG signals or superficial visual mirror information for rehabilitation regulation. The lack of fusion application of multi-dimensional and multi-modal physiological data results in large errors in decoding motor intentions, insufficient accuracy in estimating motor parameters, and a lack of good adaptive regulation capabilities. This makes it impossible to adapt to the individualized training needs of patients at different stages of rehabilitation, ultimately leading to poor overall rehabilitation training accuracy and inconsistent training effects.

[0033] The following describes a brain-computer interface-based limb rehabilitation training system and method according to embodiments of this application, with reference to the accompanying drawings. Addressing the problems of low accuracy in motor intention recognition and poor adaptability of rehabilitation aids mentioned in the background art, this application provides a brain-computer interface-based limb rehabilitation training system. Before training, the system collects kinematic parameters of the healthy upper limb and electromyographic signals of the target muscle group. Based on the principle of left-right symmetry of the human limbs, it completes electrode stimulation position registration for the unhealthy limb and generates a spatiotemporal control sequence corresponding to the target movement. During training, a training mode is implemented where the healthy upper limb performs a designated action while the unhealthy upper limb simultaneously performs motor imagery and attempt movements. Multi-channel EEG signals, kinematic parameters of the healthy side, and electromyographic signals are simultaneously collected. Electrical stimulation and exoskeleton assistance are simultaneously applied to the unhealthy limb according to a preset spatiotemporal control sequence.

[0034] The system integrates motion information, electromyography (EMG) information, and electroencephalogram (EEG) signals from the healthy side for collaborative decoding, generating refined motion control commands for the unhealthy limb. At the same time, it dynamically adjusts the exoskeleton output force and electrical stimulation intensity based on the real-time motion status of the unhealthy upper limb, assisting the patient in completing standard training movements.

[0035] Simultaneously, the system can gradually reduce the exoskeleton's auxiliary output and optimize the electrical stimulation strategy during training, assisting patients in recovering upper limb motor function through a progressive training mode. This system achieves data-driven closed-loop control of non-healthy upper limb movement intention recognition, trajectory estimation, functional electrical stimulation, and exoskeleton-assisted control, effectively improving the stability of movement decoding, optimizing the matching effect at different rehabilitation stages, and ensuring a stable and controllable training process.

[0036] Specifically, Figure 1 This is an architectural diagram of a brain-computer interface-based limb rehabilitation training system provided in an embodiment of this application.

[0037] like Figure 1 As shown, the brain-computer interface-based limb rehabilitation training system 10 includes: a brain-computer interface device 101, a kinematic sensor 102, an electromyography sensor 103, a wearable device 104, and a computer device 105.

[0038] The brain-computer interface device 101 is used to collect electroencephalogram (EEG) data of the rehabilitation target; the kinematic sensor 102 is used to collect motion data of the healthy limbs in the rehabilitation target; the electromyography (EMG) sensor 103 is used to collect EMG data of the healthy limbs in the rehabilitation target; the wearable device 104 is used to assist the unhealthy limbs in the rehabilitation target to perform training movements and collect feedback data during the execution of training movements by the unhealthy limbs in the rehabilitation target; and the computer device 105 is used to simultaneously acquire EEG data, motion data, and EMG data, generate motion control commands based on the EEG data, motion data, and EMG data, control the wearable device 104 based on the motion control commands, and adjust the training movements based on the feedback data.

[0039] It is understood that the embodiments of this application simultaneously integrate multi-source data from EEG, healthy limb movement, and healthy limb electromyography to generate control commands. By relying on measured physiological information from the healthy limb, it compensates for the susceptibility of single EEG data to noise interference, effectively improving the accuracy of movement intention recognition and ensuring that the output movement control commands better match the actual movement patterns of the human body. Simultaneously, it dynamically adjusts training parameters based on non-healthy limb movement data fed back from wearable devices, abandoning fixed parameter control modes. This allows the device's assistance intensity and electrical stimulation strategy to adapt to the patient's real-time movement state, avoiding the training drawbacks caused by excessive or insufficient assistance intensity, and optimizing the individualized rehabilitation training effect of the upper limbs of hemiplegic patients.

[0040] It should be noted that rehabilitation goals can be for hemiplegic patients with cerebrovascular disease who have unilateral upper limb motor dysfunction but intact upper limb motor function; for healthy limbs, rehabilitation goals are upper limbs with intact motor function that can independently complete standard target movements; and for unhealthy limbs, rehabilitation goals are affected upper limbs with motor function impairment, inability to independently complete standard target movements, and motor disorders.

[0041] Specifically, the computer equipment 105 implements unified management and control of three types of acquisition devices to achieve frame-level alignment by synchronously timestamping: the brain-computer interface uses a multi-channel EEG cap covering the motor cortex, prefrontal cortex, and parietal cortex, with a sampling rate ≥256Hz and electrode impedance <5 Ω. The raw EEG data is denoised using a 0.5-70Hz bandpass filter, a 50 / 60Hz power frequency notch filter, and adaptive filtering. Then, ICA (Independent Component Analysis) or a deep neural network is used to remove physiological artifacts from electrooculography (EOG) and electromyography (EMG). A motion sensor is composed of a vision sensor and an IMU (Inertial Measurement Unit). A camera equipped with a deep learning posture estimation model calculates 3D joint coordinates, and the IMU collects triaxial acceleration and angular velocity for data compensation in scenarios with visual occlusion or abnormal lighting. An EMG sensor 103 extracts amplitude, envelope, integrated EMG, activation sequence, and co-contraction features. The system uses two algorithms, DTW (Dynamic Time Warping) and HMM (Hidden Markov Model), to construct a spatiotemporal template library for action classification. Real-time feedback data from the affected side's exoskeleton is introduced to cross-validate the action recognition results. When there are discrepancies in the recognition, the template matching confidence is corrected using weights. Finally, a CNN (Convolutional Neural Network) is used to... Network (Convolutional Neural Network) / Transformer (Transformer Neural Network) / EEG base model extracts spatiotemporal features of EEG, which are then used to generate motor EEG features through a trainable brain region weight fusion layer. (14~30Hz) Energy integral - elbow joint angle (30~70Hz) Peak-to-wrist joint rotation nonlinear mapping was used to initially calculate joint parameters, and Kalman filtering was combined with measured data from the healthy side to dynamically correct the mapping. Intent recognition employed a two-layer decision tree architecture, further subdividing major action categories into columnar gripping, ball gripping, pincer gripping, full arm extension, and forearm independent extension. During the training phase, the dataset was constructed using the brain-body symmetry principle, DTW temporal alignment, and spatial normalization, employing CNN and LSTM (Long Short-Term Memory) algorithms. The system uses a combination of Long Short-Term Memory (LSTM) networks, CNNs, and Transformers to construct a decoding model through transfer learning. Online control employs adaptive Kalman filtering, fusing information from EEG prediction, healthy side motion observation, and healthy side electromyography (EMG) observation. Inverse kinematics is used to generate the first control parameters for the exoskeleton, and EMG mapping generates the second control parameters for electrical stimulation. The system dynamically fine-tunes the exoskeleton's auxiliary torque based on the correlation coefficient of wrist and elbow joint angle changes. Finally, it calculates four indicators—speed, trajectory accuracy, smoothness, and joint matching—based on exoskeleton feedback data, and adaptively lowers the auxiliary power using a five-level gradient. The system incorporates reference thresholds for joint range of motion (elbow flexion 0-145°, wrist flexion 0-80°, and dorsiflexion 0-70°), which, combined with human physiological joint thresholds, enable over-limit emergency stops and locking protection.

[0042] In this embodiment of the application, the wearable device 104 includes an exoskeleton device and an electrode device. The exoskeleton device responds to motion control commands to perform corresponding actions to drive the movement of the unhealthy limb. The electrode device responds to motion control commands to output stimulation signals to stimulate the unhealthy limb.

[0043] It is understood that the embodiments of this application utilize the dual driving methods of mechanical traction from the exoskeleton device and electrical stimulation from the electrode device to assist the unhealthy limb in completing the target movement from two dimensions: mechanical external force and muscle electrical activation. The two forms of assistance work together to improve the reliability of limb movement completion.

[0044] Specifically, the exoskeleton device is a multi-degree-of-freedom upper limb rehabilitation exoskeleton covering the shoulder, elbow, wrist, and fingers. Each joint integrates high-precision angle and torque sensors, providing real-time feedback on joint angles, angular velocities, and output torque. It supports dynamic switching between three modes: position control, speed control, and impedance control. Based on the first control parameter, it outputs corresponding torque, rotational speed, and joint position, mechanically tractioning the affected limb to complete the target motion trajectory. The device has a built-in database of human physiological activity thresholds, using a biomechanical model to limit the extreme rotation angles of each joint. It automatically enters a locking protection state when parameters exceed limits, rotational blockage occurs, or spasticity occurs. The electrode device is a surface functional electrical stimulation device, employing a biphasic pulse stimulation output mode. It can precisely adjust the electrical stimulation trigger time, pulse width, pulse frequency, and current intensity based on the second control parameter. The standard stimulation parameter range is a pulse width of 100-300. With a frequency of 20-60Hz, it can accurately activate the target muscle groups on the affected side, actively induce muscle contraction through muscle electrical stimulation, and work in conjunction with the mechanical traction of the exoskeleton to assist the affected limb in completing standardized and regulated rehabilitation training movements. The stimulation intensity is limited to within the patient's pain tolerance threshold. The electrode attachment points are determined by the previous electromyographic data of the healthy side and the anatomical symmetry of the human body.

[0045] In this embodiment of the application, the computer device 105 is further configured to collect electromyographic data of the healthy limb of the rehabilitation target performing the target action through the kinematic sensor 102 before generating the motion control command based on the electroencephalogram data, motion data and electromyographic data, and register the stimulation position of the electrode device on the unhealthy limb based on the electromyographic data.

[0046] It is understood that the embodiments of this application rely on the electromyographic data corresponding to the movement of the healthy side limb to complete the electrode device point registration, and determine the stimulation position of the unhealthy side limb according to the symmetry law of human limbs, so as to ensure that the electrical stimulation can be accurately applied to the target muscle group and improve the effectiveness of electrical stimulation-assisted movement.

[0047] It should be noted that the target movements are standardized upper limb rehabilitation movements preset for rehabilitation training, including daily rehabilitation training movements such as grasping, extension, elbow flexion, arm raising, forearm pronation, forearm supination, and wrist flexion and extension; the stimulation sites are the surface electrode attachment sites of the affected upper limb muscle groups, determined after low-intensity trial stimulation and fine-tuning based on the anatomical symmetry of the left and right limbs of the human body, combined with the activation sites and electromyographic activation characteristics of the target muscle groups on the healthy side, and can stably induce the corresponding target movements within the patient's tolerance threshold.

[0048] Specifically, the computer device 105 pre-collects surface electromyographic (EMG) data of the target muscle groups when the healthy upper limb performs various target movements. It accurately extracts the EMG amplitude range, envelope change patterns, integrated EMG values, muscle activation initiation and termination sequences, and synergistic contraction characteristics of the healthy muscle groups, identifying the core force-generating muscle groups and activation areas on the healthy side for each target movement. Based on the principle of symmetry in the anatomical structure and muscle distribution of the left and right upper limbs, the electrode device on the affected side is initially deployed to the affected muscle group area symmetrically corresponding to the core force-generating muscle groups on the healthy side. Subsequently, test electrical stimulation is output in a low-to-high gradient, recording the muscle contraction effect at each point. The optimal stimulation location is selected that can stably induce the corresponding target movement on the affected limb without causing pain, spasms, or other discomfort in the patient. This completes the precise registration and positioning of the electrode device, ensuring that subsequent electrical stimulation can accurately target and activate the target muscle groups on the affected side. During subsequent training, the system can synchronously modulate various electrical stimulation parameters on the affected side based on real-time EMG characteristics of the healthy side.

[0049] In this embodiment, the computer device 105 is further configured to control the brain-computer interface device 101, kinematic sensor 102, and electromyography sensor 103 to synchronously collect data during the rehabilitation target's execution of a training task before generating motion control instructions based on EEG data, motion data, and electromyography data. Based on the synchronously collected data, a training dataset is constructed, and a mapping model from the EEG characteristics of the healthy limb to the target motion parameters of the unhealthy limb is constructed based on the training dataset. The training task includes: the healthy limb of the rehabilitation target performing the target action, while the unhealthy limb of the rehabilitation target performing the target action through motor imagination and action attempt.

[0050] It is understood that the embodiments of this application use multiple types of data collected simultaneously from the practical movements of the healthy side limbs and the simultaneous imagination and trial movements of the unhealthy side limbs to build a training dataset, establish a feature mapping model, optimize the accuracy of motion parameter solution based on actual test samples, and further improve the accuracy of subsequent motion control command generation.

[0051] It should be noted that the target motion parameters are quantitative parameters that characterize the target motion state of the unhealthy limb, including continuous motion parameters such as joint angle, joint angular velocity, motion acceleration, end displacement trajectory, motion speed, motion duration, motion amplitude, and trajectory direction; the mapping model is an intelligent decoding model based on deep neural network training, used to establish a nonlinear correspondence between the EEG characteristics of the healthy limb and the target motion parameters of the unhealthy limb.

[0052] Specifically, computer device 105 guides the rehabilitation target to perform standardized synchronous training tasks through a human-computer interaction interface. Throughout the process, it strictly controls the brain-computer interface device 101, kinematic sensor 102, and electromyography (EMG) sensor 103 to collect data synchronously in a strict time sequence. That is, while the patient's unaffected upper limb actually performs various target movements, the affected upper limb simultaneously performs motor imagery and active limb attempts at the same target movements. The system synchronously collects multi-channel EEG data, three-dimensional kinematic data of the unaffected upper limb, and EMG data of the target muscle groups on the unaffected side, collecting sufficient training samples through multiple repeated trials. Subsequently, a multimodal data spatiotemporal correlation matrix is ​​established, and the Dynamic Time Warping (DTW) algorithm is used to align the time sequences of EEG, movement, and EMG, eliminating temporal deviations caused by human reaction time differences. Spatial normalization is used to unify the spatial mapping of joint coordinates and muscle group positions on both the unaffected and affected sides. Using synchronously collected unaffected side movement data as supervision labels, unaffected side EMG data as stimulus modulation references, and synchronous affected side EEG data as feature inputs, a paired training dataset containing action category labels and continuous target movement parameter labels is constructed. Based on this dataset, a deep neural network mapping model is trained. The network can adopt a hybrid architecture of (CNN and LSTM) or (CNN and Transformer). Convolutional layers extract spatial features, and temporal networks capture dynamic changes. The training aims to minimize the mean square error between the prediction and the ground truth. During the training process, the electrical stimulation control parameters output by the electromyographic feature modulation model of the healthy side are simultaneously used. At the same time, a transfer learning strategy is introduced. The basic model is pre-trained using a publicly available EEG motor imagery dataset, and then fine-tuned using a small amount of personalized data from patients. Finally, a decoding model is constructed that maps the EEG features of the healthy side limbs to the target motor parameters of the unhealthy side limbs.

[0053] In this embodiment of the application, the computer device 105 is further configured to extract EEG data and motion data from the training dataset, fuse the EEG data and motion data into motion EEG features of the healthy limb, decode and identify the motion intention of the rehabilitation target based on the EEG data, determine the motion type based on the motion intention, extract limb motion features from the motion data, determine the target motion parameters corresponding to the motion type based on the limb motion features, and construct a mapping model based on the motion EEG features and the target motion parameters.

[0054] It is understood that the embodiments of this application integrate EEG data and motion data to form motion EEG features, combine motion intention and limb motion parameters to build a mapping model, and rely on multi-dimensional features to optimize the model fitting effect and improve the accuracy of motion parameter estimation.

[0055] It should be noted that the motor intention is the subjective motor desire generated by the brain of the rehabilitation target, which expects the non-healthy limb to complete the corresponding upper limb movement; the movement type is the standardized upper limb rehabilitation movement category identified and distinguished by the system, including grasping, extension, elbow flexion, arm raising, forearm pronation, forearm supination, and wrist flexion and extension, and can be further subdivided into various movement sub-patterns; the motor EEG features are a highly discriminative joint spatiotemporal feature vector obtained by fusing the motor state information of the healthy limb with the EEG neural activity information.

[0056] Specifically, computer device 105 extracts preprocessed EEG data (including time-aligned and noise-reduced data) and healthy limb movement data from the constructed training dataset in batches. It then automatically extracts the spatiotemporal features of the EEG signals and the kinematic features of the limbs through a deep learning network. These features are then weighted and fused using a trainable brain region weight fusion layer to generate the motor EEG features of the healthy limbs. Simultaneously, it decodes and recognizes movement intentions based on the preprocessed, clean EEG signals, utilizing a two-layer decision tree hierarchical model: the first layer determines whether an upper limb movement command is triggered based on shoulder and elbow acceleration and angular velocity thresholds; the second layer distinguishes major movement categories based on joint displacement and opening / closing features. For grasping, it further subdivides into three sub-modes: columnar grasping, ball grasping, and pincer grasping, based on thumb and index finger opening / closing and palm closing angles; for extension, it distinguishes between full arm extension and independent forearm extension based on elbow and wrist activation timing and forearm rotation positive / negative angles. This is combined with… Frequency band (14~30Hz) energy integral correlation with elbow joint rotation angle The physiological mapping law of wrist joint rotation angle associated with frequency band (30~70Hz) was established, and the mapping coefficients were corrected online using measured data from the healthy side in conjunction with Kalman filtering. Furthermore, limb movement features such as joint angles, displacement trajectories, movement speed, and acceleration were extracted from the healthy side's motion data to match and determine the standardized target motion parameters corresponding to the current movement type. Finally, using motor EEG features as input features and target motion parameters as supervised output, the network weights were optimized by minimizing the mean square error between the predicted and actual values. Iterative training yielded a high-precision nonlinear mapping model, achieving accurate correlation and fitting between neural features and motion parameters.

[0057] In this embodiment of the application, the computer device 105 is further configured to fuse EEG data and motion data into motion EEG features of the healthy limb, determine target motion parameters of the unhealthy limb based on the motion EEG features and mapping model of the healthy limb, generate first control parameters of the exoskeleton device based on the target motion parameters, generate second control parameters of the electrode device based on electromyography data, and generate motion control commands based on the first and second control parameters.

[0058] It is understood that the embodiments of this application obtain motor EEG characteristics by fusing EEG data and motion data of the healthy side limb, calculate the target motion parameters corresponding to the unhealthy side limb by combining a pre-built mapping model, generate mechanical control parameters of the exoskeleton device based on the target motion parameters, and generate electrical stimulation control parameters of the electrode device based on real-time electromyography data. The two types of control parameters are integrated into a unified motion control command, realizing precise synchronization and synergistic regulation of exoskeleton mechanical assistance and electrical stimulation of muscle activation, effectively improving the standardization and matching degree of rehabilitation training movements.

[0059] It should be noted that the first control parameter is the mechanical control parameter used to control the movement state of the exoskeleton device, including the target angle, target angular velocity, target output torque, movement trajectory, movement speed and damping coefficient of each joint of the exoskeleton; the second control parameter is the electrical stimulation control parameter used to control the stimulation state of the electrode device, including the electrical stimulation trigger time, pulse width, stimulation frequency, output current intensity and stimulation duration.

[0060] Specifically, during rehabilitation training, computer device 105 integrates synchronously collected EEG data and healthy side motor data in real time, dynamically updates the motor EEG characteristics of the healthy limb, and inputs the real-time motor EEG characteristics into the trained mapping model to accurately decode and output the target motor parameters corresponding to the unhealthy limb. The mapping model is implemented based on a CNN (Convolutional Neural Network) or Transformer (Transformer Neural Network) architecture, and the specific selection is determined based on the scale and characteristics of the training data. The core mapping relationship can be expressed as: ,in, For real-time EEG signal tensors, For the kinematic data tensor of the healthy side, It is a pre-trained single-model mapping network (either CNN or Transformer). The output parameters are the target position and velocity parameters on the affected side. This mapping model is pre-trained offline using mirror motion paired data from healthy subjects, with the training objective being to minimize the mean square error between the predicted trajectory and the measured trajectory.

[0061] The computer device 105 dynamically switches time windows according to the speed of movement based on the target motion parameters. Using an inverse kinematics algorithm combined with the exoskeleton link length and degree-of-freedom structural parameters, it calculates the first control parameters for each joint of the exoskeleton, including target angles, angular velocities, output torque, and damping coefficients. It determines limb coordination by calculating the correlation coefficient of wrist-elbow angle change rate. When coordination is excellent, the output torque is reduced; when coordination is low, the output torque is increased; and when coordination is below a threshold, joint damping is increased for resistance training. Simultaneously, it extracts the amplitude, envelope peak value, integral electromyography, and activation timing features of the healthy side's electromyography data in real time, mapping and matching them to obtain the second control parameters for the affected side's electrical stimulation, including triggering timing, pulse width, frequency, and current intensity. Finally, the first and second control parameters are synchronized in time and coupled together to generate unified, synchronized, and adaptive coordinated motion control commands tailored to the patient's real-time motion state. These commands are then sent to the exoskeleton device and electrode device, achieving complete synchronized and coordinated control of mechanical assistance and electrical stimulation activation. Before outputting the commands, safety constraints are checked against human joint range of motion thresholds from literature.

[0062] In this embodiment, the exoskeleton device incorporates at least one of an angle sensor and a torque sensor. The angle sensor collects the joint angle and joint angular velocity of the unhealthy limb, and the torque sensor collects the output torque at the unhealthy limb.

[0063] It is understood that the embodiments of this application, through the angle and torque sensors built into the exoskeleton device, can collect motion state information such as joint angles, joint angular velocities, and output torque of the unhealthy limb in real time and accurately, comprehensively capture the actual motion details of the limb, and provide real and reliable data support for subsequent dynamic adjustment of rehabilitation training parameters and optimization of motion control commands, ensuring the accuracy and real-time nature of rehabilitation regulation.

[0064] It should be noted that the joint angle refers to the real-time opening and closing angle and rotation angle of the shoulder, elbow, wrist and finger joints of the affected upper limb during movement; the joint angular velocity is the rate of angle change of each joint on the affected side per unit time, used to characterize the speed of limb movement; and the output torque is the mechanical torque output in real time by each drive joint of the exoskeleton, used to characterize the auxiliary force intensity of the exoskeleton.

[0065] Specifically, each drive joint of the exoskeleton device is equipped with high-precision angle and torque sensors, which can collect multi-dimensional mechanical motion feedback data during the movement of the affected limb in real time and at high frequency. The angle sensors sample the absolute angle, relative angle, and angle change per unit time of each joint in real time, and convert them into real-time joint angles and joint angular velocities; the torque sensors collect the actual output torque during the interaction between the exoskeleton drive joints and the patient's limb in real time, and accurately provide feedback on the magnitude of the exoskeleton's assisted force and the patient's active force application. All sensor feedback data is transmitted back to computer device 105 in real time to calculate four quantitative evaluation indicators: movement completion speed, trajectory accuracy, movement smoothness, and joint angle matching degree. The system automatically adjusts parameters using a five-level auxiliary gradient based on the achievement of the indicators: Level 5 is full auxiliary output of rated maximum torque, Level 4 is strong auxiliary output of 70% of rated torque, Level 3 is medium auxiliary output of 50% of rated torque, Level 2 is weak auxiliary output of 30% of rated torque with only gravity compensation, and Level 1 is zero-force mode with no active drive and only gravity compensation. An adaptive gain algorithm is introduced, which downgrades the level in advance if the level is achieved continuously, and extends the training cycle of the current level if the level is not achieved multiple times. All training data is archived in time sequence, and the active force contribution of the affected side is estimated by comprehensively considering torque, movement delay, and trajectory deviation. Visualized rehabilitation reports are generated regularly.

[0066] In this embodiment of the application, the kinematic sensor 102 includes at least one of an image sensor and an inertial measurement unit. The image sensor acquires motion images of the healthy limb, and the computer device 105 identifies the joint angle, end displacement trajectory, movement speed and movement duration of the healthy limb based on the motion images. The inertial measurement unit acquires the acceleration and angular velocity of the unhealthy limb.

[0067] It is understood that the embodiments of this application use image sensors and inertial measurement units to collect motion state data of healthy and unhealthy limbs in multiple dimensions. This enables comprehensive and accurate acquisition of a wealth of motion parameters such as limb joint angles, displacement trajectories, motion speed, acceleration, and angular velocity, enriching the dimensions of data acquisition and providing a comprehensive and effective data foundation for subsequent multi-source data fusion, motion intention recognition, and accurate generation of rehabilitation control commands.

[0068] It should be noted that the end-effector trajectory is the continuous motion displacement path of the unaffected upper limb hand in three-dimensional space; the motion speed is the real-time motion rate and average motion rate of the unaffected limb during the execution of the target motion; and the motion duration is the total time taken for the unaffected limb from the start of the motion to the end of the motion.

[0069] Specifically, the image sensor uses a high-resolution camera equipped with a deep learning pose estimation model to capture continuous motion images of the healthy upper limb in real time. A key point detection algorithm identifies the three-dimensional spatial coordinates of key nodes in the shoulder, elbow, wrist, and hand, continuously resolving the angles of each joint on the healthy side, the three-dimensional displacement trajectory of the hand's end-effector, real-time movement speed, and the duration of a single target movement. An inertial measurement unit (IMU) is worn on the affected upper arm and wrist, independently acquiring triaxial acceleration and angular velocity data during the movement of the affected limb. When the image sensor encounters changes in lighting or limb occlusion causing visual tracking failure, the system performs real-time correction and compensation of the motion data. The system establishes a unified spatial coordinate system, converting visual, IMU, electromyographic, and exoskeleton parameters to the same reference frame. The DTW / HMM algorithm and standard action templates are used for temporal matching to identify the action type. Simultaneously, exoskeleton feedback data is integrated to cross-validate the accuracy of action recognition and correct matching weights. Image sensors focus on acquiring global trajectory and temporal features of the healthy side, while inertial measurement units focus on acquiring high-frequency dynamic features of the affected limb. The two types of sensors complement and fuse to achieve comprehensive and high-precision acquisition of motion data of the healthy and affected limbs, providing complete kinematic data support for subsequent template matching, intent decoding, parameter fusion, and command generation.

[0070] In this embodiment of the application, the motor coverage area of ​​the brain-computer interface device 101 includes at least one of the left and right hemisphere motor cortex, the prefrontal cortex, and the parietal cortex, and the electromyographic data includes at least one of electromyographic amplitude, envelope, integral electromyography, activation sequence, and co-contraction characteristics.

[0071] It is understood that the embodiments of this application can accurately capture the motor nerve signals of the rehabilitation target by collecting signals from multiple motor-related cortical areas of the brain. At the same time, it extracts electromyographic features such as electromyographic amplitude, integral electromyography, and activation sequence from multiple dimensions, enriching the signal feature dimensions at the nerve and muscle levels, effectively improving the comprehensiveness and accuracy of motor intention recognition, and providing sufficient and effective data support for the subsequent accurate generation of rehabilitation motor control commands.

[0072] Specifically, the brain-computer interface device 101 uses a multi-channel EEG acquisition cap, with an electrode array covering brain regions highly related to limb movement regulation, such as the left and right hemisphere motor cortex, prefrontal cortex, and parietal cortex. This allows for comprehensive capture of brain neural electrical activity signals during motor imagery and active trial movements, with a signal sampling rate ≥256Hz and electrode impedance controlled within 5 Hz. The following preprocessing steps, including 0.5~70Hz bandpass filtering, 50 / 60Hz power frequency notch filtering, adaptive filtering, and ICA / deep neural network artifact removal, avoid the problems of missing neural signals and one-sided intention recognition caused by single brain region acquisition, ensuring the integrity of neural signal acquisition. The electromyography (EMG) sensor 103 can acquire and analyze multi-dimensional EMG features of the healthy target muscle group, specifically including: EMG amplitude representing muscle exertion intensity, EMG envelope representing continuous muscle activation state, integral EMG representing total muscle work, activation sequence representing muscle movement timing pattern, and coordinated contraction features representing multi-muscle group coordinated exertion state. The system integrates multi-brain region neural signals and multi-dimensional EMG activity features, feeding them into CNN, Transformer, or large EEG-based models for feature extraction and weighted fusion. This achieves comprehensive perception of neuromuscular dual-layer information, greatly enriching the model input feature dimensions, effectively suppressing noise interference, and significantly improving the accuracy of motion intention decoding and the stability of motion parameter estimation. This provides a sufficient and reliable multimodal data foundation for accurately generating coordinated rehabilitation control commands.

[0073] The brain-computer interface-based limb rehabilitation training system proposed in this application integrates multi-source data from electroencephalography (EEG), healthy limb movement, and electromyography (EMG) of the healthy limb to generate control commands. It leverages measured physiological information from the healthy limb to compensate for the susceptibility of single EEG data to noise interference, effectively improving the accuracy of movement intention recognition and ensuring that the output movement control commands better match the actual movement patterns of the human body. Simultaneously, it dynamically adjusts training parameters based on non-healthy limb movement data fed back from wearable devices, abandoning fixed parameter control modes. This allows the device's assistance intensity and electrical stimulation strategy to adapt to the patient's real-time movement state, avoiding the training drawbacks caused by excessive or insufficient assistance intensity, and optimizing the individualized rehabilitation training effect for the upper limbs of hemiplegic patients.

[0074] The following will combine Figure 2 The process of multi-source information fusion based on brain-body symmetry is described in detail. For example... Figure 2 As shown, the specific steps are as follows: In step one, EEG acquisition is performed. A multi-channel EEG acquisition cap covering the left and right hemispheres' motor cortex, prefrontal cortex, and parietal cortex is deployed to acquire multi-channel EEG signals from the entire brain. The sampling rate is set to be no less than 256 Hz, and the electrode impedance is controlled at 5 ohms. The raw EEG signal is then sequentially filtered through a 0.5 to 70 Hz bandpass filter, a 50 or 60 Hz power frequency notch filter, and an adaptive filter to remove environmental and motion noise. Then, ICA or a deep neural network is used to remove physiological artifacts from electrooculography (EOG) and electromyography (EMG) to obtain clean EEG data. Subsequently, spatiotemporal features of EEG are extracted using CNN, Transformer, or a large EEG-based model. Finally, a trainable brain region weighted fusion layer is used to weight and fuse multi-channel features to generate standardized motion EEG features. The entire process is synchronized with the motion and EMG acquisition terminals to achieve frame-level temporal alignment.

[0075] In step two, motion monitoring is performed. A high-resolution camera equipped with a deep learning pose estimation model enables visual tracking of the unaffected upper limb. Simultaneously, an IMU (Inertial Measurement Unit) and an unaffected side surface electromyography (EMG) sensor 103 collect various data, including 3D motion trajectory, joint angles, acceleration, angular velocity, and EMG amplitude envelope integral EMG. A unified spatial coordinate system is established to normalize the coordinates of the multi-source data. A motion template library is built using DTW (Dynamic Time Warping) and HMM (Hidden Markov Model) to achieve temporal matching and motion type recognition. Real-time feedback data from the affected side's exoskeleton is introduced to cross-validate the motion recognition results. Template matching weights are dynamically adjusted based on recognition deviations. A two-layer decision tree hierarchical structure is used to further subdivide various sub-motions under actions such as grasping and extending. Then, through… Frequency band energy integral related to elbow joint angle, The nonlinear mapping relationship between frequency band energy peak and wrist joint rotation angle was initially established to establish the corresponding rules of EEG motion parameters, and the mapping coefficients were dynamically calibrated using Kalman filtering combined with measured data.

[0076] In step three, a training set is constructed. Based on the left-right symmetry of human limb and brain nerve encoding, the initial symmetrical deployment of stimulation electrodes on the affected side is completed using the multi-dimensional electromyographic features of the healthy side. The optimal stimulation points are then determined through fine-tuning with low-intensity graded trial stimulation. The patient is then guided to complete a synchronous training task. While the healthy side performs the target movement, the affected side simultaneously performs corresponding movement imagination and movement attempts. Full-link EEG and EMG data are collected synchronously. The DTW algorithm is used to align multimodal temporal sequences and spatially normalize and unify the spatial coordinates of the healthy and affected sides. A pairwise joint training dataset with dual labels of movement category and continuous movement parameters is constructed. Based on the dataset, a decoding mapping model with a CNN plus LSTM or CNN plus Transformer hybrid architecture is built. The network weights are iteratively optimized with the goal of minimizing the mean square error between the predicted and true parameters. Transfer learning is also introduced. After pre-training the basic model based on a public EEG dataset, the model is fine-tuned using a small sample of the patient's personalization. The electrical stimulation parameters of the affected side are synchronously modulated according to the electromyographic features of the healthy side during the training phase.

[0077] In step four, fusion decoding is performed. During the rehabilitation training phase, three input information channels are summarized in real time: the predicted movement intention information of the affected side obtained by decoding through the mapping model, the actual movement observation data output by the healthy side movement monitoring module, and the muscle group activation observation information collected by the healthy side electromyography sensor 103. The optimal fusion of the three types of information is completed by using adaptive Kalman filtering, and the expected movement parameters of the affected side are calculated. The first control parameters corresponding to the exoskeleton drive are obtained by kinematic inverse kinematics conversion combined with the exoskeleton mechanical structure parameters. Simultaneously, the real-time healthy side electromyography feature mapping is extracted to generate the second control parameters corresponding to electrical stimulation. The two types of control parameters are temporally coupled to generate a coordinated movement control command.

[0078] In step five, rehabilitation is performed and adjusted. The generated control commands are sent to the multi-degree-of-freedom upper limb exoskeleton worn on the affected side and the surface functional electrical stimulation device. The exoskeleton outputs corresponding torque to complete mechanical traction based on the received first control parameter, and the electrical stimulation device outputs 100 to 300 according to the second control parameter. The system uses biphasic pulse stimulation with pulse width and a frequency of 20 to 60 Hz to synergistically activate the affected side's muscle groups. The system collects feedback data from the exoskeleton's built-in angle and torque sensors in real time, quantifying the quality of the movement based on four indicators: speed of movement completion, trajectory accuracy, smoothness of movement, and joint angle matching. The system adaptively adjusts the assistance intensity according to a five-level gradient assistance rule, progressively decreasing from level 5 (full assistance) to level 4 (70% torque strong assistance), level 3 (50% torque medium assistance), level 2 (30% torque weak assistance), and finally level 1 (zero force with only gravity compensation). During the process, the system flexibly adjusts the pace of advancement and demotion based on training performance using an adaptive gain algorithm. The system stores all rehabilitation data throughout the process and calculates the contribution of the affected side's active force exertion, generating visualized rehabilitation reports periodically to complete phased closed-loop management. At the same time, it sets safety constraints based on the human physiological joint activity threshold database, automatically triggering deceleration, locking, and emergency stop protection when conditions such as over-limit spasticity, stunting, or communication abnormalities occur.

[0079] In summary, the embodiments of this application have at least the following beneficial effects: (1) Whole brain region EEG acquisition with noise reduction preprocessing is adopted, combined with deep learning to decode motor intention, and visual, inertial sensing and electromyography multimodal monitoring are integrated. Based on the symmetry characteristics of the brain and body, the parameters and electrode positions of the affected side are calibrated with the physiological data of the healthy side, which effectively improves the accuracy of intention recognition and the reliability of electrode registration.

[0080] (2) Using healthy side movement and electromyography data to construct a symmetrical supervised training set to optimize the decoding model, during the training and rehabilitation stages, adaptive Kalman multi-source information fusion is used to generate collaborative instructions, linking the exoskeleton and electrical stimulation for synchronous treatment, thereby improving the robustness of motion parameter estimation on the affected side.

[0081] (3) Relying on multi-dimensional quantitative indicators to assess the rehabilitation level, the exoskeleton and electrical stimulation intensity are adjusted adaptively and progressively to achieve individualized closed-loop rehabilitation training that matches the patient's recovery ability.

[0082] The control method of the brain-computer interface-based limb rehabilitation training system proposed in this application will be described next through a specific embodiment, such as... Figure 3 As shown, this method is applied to the computer device 105 of the brain-computer interface-based limb rehabilitation training system in the above embodiment. The specific steps are as follows: In step S301, the electroencephalogram (EEG), motor data, and electromyogram (EMG) data of the rehabilitation target are acquired simultaneously.

[0083] It is understood that the embodiments of this application, through multimodal synchronous acquisition combined with preprocessing, rely on a brain-body symmetrical data acquisition architecture to avoid the defects of single signals being susceptible to noise interference, and provide a complete data source for subsequent accurate decoding and instruction generation.

[0084] Specifically, the brain-computer interface uses a multi-channel EEG cap covering the left and right motor cortex, prefrontal cortex, and parietal lobe to collect EEG data, with a sampling rate ≥256Hz and electrode impedance < The raw EEG data was denoised by 0.5-70Hz bandpass filtering, power frequency notch filtering, adaptive filtering, and ICA artifact removal. Fine motor parameters of the healthy side were collected using a high-definition visual camera (deep learning posture calculation) and an IMU. The electromyography sensor 103 extracted five types of features: amplitude, envelope, integral electromyography, activation sequence, and co-contraction. The three types of data were unified with a timestamp to achieve frame-level temporal alignment, and a common spatial coordinate system was established to complete data normalization.

[0085] In step S302, motion control commands are generated based on EEG data, motion data, and electromyography data.

[0086] It is understood that the embodiments of this application construct a decoding model based on brain-body symmetry, combine an adaptive Kalman multi-source fusion strategy, dynamically optimize parameter estimation accuracy, and split dual-type control parameters to achieve coordinated control of exoskeleton and electrical stimulation.

[0087] Specifically, it relies on a pre-trained deep network mapping model to analyze motor EEG features, combined with , The nonlinear mapping relationship between frequency band EEG energy and joint angle is used to initially calculate the target motion parameters on the affected side. By integrating three types of information—EEG intention prediction, healthy side motion observation, and electromyography observation—through adaptive Kalman filtering, the first parameter of exoskeleton mechanical control is calculated through inverse kinematics. The second parameter—electric stimulation trigger time, pulse width, frequency, and current—is obtained by mapping real-time electromyography features. The two types of parameters are coupled in time to generate an integrated coordinated motion control command. In step S303, the wearable device 104 is controlled based on motion control commands to assist the unhealthy limb in the rehabilitation target to perform training movements, and the training movements are adjusted based on the feedback data from the wearable device 104.

[0088] It is understood that the embodiments of this application rely on real-time feedback from the exoskeleton to quantify rehabilitation effectiveness, and adjust the auxiliary intensity in a graded and adaptive manner to form individualized rehabilitation regulation, taking into account both training safety and adaptability.

[0089] Specifically, control commands are sent to the multi-degree-of-freedom exoskeleton and the biphasic pulse electrical stimulation module, with the electrical stimulation parameters limiting the pulse width to 100~300. The frequency is 20~60Hz; the device has built-in angle and torque sensors that transmit feedback data in real time. The system evaluates the rehabilitation status based on four indicators: movement completion speed, trajectory accuracy, smoothness, and joint matching. It adaptively adjusts the assistance intensity according to a five-level gradient (level 5 full assistance to level 1 zero force compensation). At the same time, it sets a safety locking mechanism based on the human body's standard joint activity threshold, and automatically stops the machine to protect against abnormal over-limit, spasm, and jamming.

[0090] It should be noted that the foregoing explanation of the embodiment of the brain-computer interface-based limb rehabilitation training system also applies to the control method of the brain-computer interface-based limb rehabilitation training system of this embodiment, and will not be repeated here.

[0091] The control method for a brain-computer interface-based limb rehabilitation training system proposed in this application achieves comprehensive synchronous acquisition of neural signals, limb movement signals, and muscle activity signals by simultaneously collecting EEG, motion, and electromyography data of the rehabilitation target. It generates precise and reliable motion control commands based on multi-source data fusion, and uniformly regulates wearable devices to assist the unhealthy limb in completing training movements. Simultaneously, it dynamically adjusts training movements based on real-time feedback data from the device, achieving closed-loop rehabilitation training control. This effectively improves the insufficient accuracy of single signal recognition, significantly enhancing the accuracy, adaptability, and intelligence of limb rehabilitation training.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0094] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0095] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0096] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0097] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A limb rehabilitation training system based on a brain-computer interface, characterized in that, include: Brain-computer interface devices are used to collect electroencephalogram (EEG) data for rehabilitation goals; Kinematic sensors are used to collect motion data of the healthy limb in the rehabilitation target. An electromyography (EMG) sensor is used to collect EMG data of the healthy limb in the rehabilitation target. Wearable devices are used to assist the unhealthy limbs in the rehabilitation target to perform training movements and to collect feedback data during the process of the unhealthy limbs in the rehabilitation target performing training movements. A computer device is used to synchronously acquire the electroencephalogram (EEG) data, the motion data, and the electromyogram (EMG) data, generate motion control commands based on the EEG data, the motion data, and the EMG data, control the wearable device based on the motion control commands, and adjust the training movements based on the feedback data.

2. The limb rehabilitation training system based on a brain-computer interface according to claim 1, characterized in that, The wearable device includes an exoskeleton device and an electrode device. The exoskeleton device responds to the motion control command to perform corresponding actions to drive the movement of the unhealthy limb. The electrode device responds to the motion control command to output stimulation signals to stimulate the unhealthy limb.

3. The limb rehabilitation training system based on a brain-computer interface according to claim 2, characterized in that, The computer device is further configured to, before generating motion control commands based on the electroencephalogram (EEG) data, the motion data, and the electromyogram (EMG) data, acquire EMG data of the healthy limb of the rehabilitation target performing the target action using the kinematic sensor, and register the electrode device at the stimulation position of the unhealthy limb based on the EMG data.

4. The limb rehabilitation training system based on a brain-computer interface according to claim 2, characterized in that, The computer device is also used to control the brain-computer interface device, the kinematic sensor, and the electromyographic sensor to synchronously collect data during the rehabilitation target's execution of a training task before generating motion control instructions based on the EEG data, the motion data, and the electromyographic data. A training dataset is constructed based on the synchronously collected data, and a mapping model is built from the EEG characteristics of the healthy limb to the target motion parameters of the unhealthy limb based on the training dataset. The training task includes: the healthy limb of the rehabilitation target performing a target action, while the unhealthy limb of the rehabilitation target simultaneously performs motor imagery and action attempts related to the target action.

5. The limb rehabilitation training system based on a brain-computer interface according to claim 4, characterized in that, The computer device is also used to extract the EEG data and the motion data from the training dataset, fuse the EEG data and the motion data into motion EEG features of the healthy limb, decode and identify the motion intention of the rehabilitation target based on the EEG data, determine the motion type based on the motion intention, extract limb motion features from the motion data, determine the target motion parameters corresponding to the motion type based on the limb motion features, and construct the mapping model based on the motion EEG features and the target motion parameters.

6. The limb rehabilitation training system based on a brain-computer interface according to claim 4, characterized in that, The computer device is further configured to fuse the EEG data and the motion data into the motion EEG characteristics of the healthy limb, determine the target motion parameters of the unhealthy limb based on the motion EEG characteristics of the healthy limb and the mapping model, generate the first control parameters of the exoskeleton device based on the target motion parameters, generate the second control parameters of the electrode device based on the electromyography data, and generate the motion control command based on the first control parameters and the second control parameters.

7. The limb rehabilitation training system based on a brain-computer interface according to claim 2, characterized in that, The exoskeleton device has at least one built-in angle sensor and torque sensor. The angle sensor collects the joint angle and joint angular velocity of the unhealthy limb, and the torque sensor collects the output torque at the unhealthy limb.

8. The limb rehabilitation training system based on a brain-computer interface according to claim 1, characterized in that, The kinematic sensor includes at least one of an image sensor and an inertial measurement unit. The image sensor acquires motion images of the healthy limb, and the computer device identifies the joint angles, end-effector displacement trajectories, movement speeds, and movement durations of the healthy limb based on the motion images. The inertial measurement unit acquires the acceleration and angular velocity of the unhealthy limb.

9. The limb rehabilitation training system based on a brain-computer interface according to claim 1, characterized in that, The motor coverage area of ​​the brain-computer interface device includes at least one of the left and right hemisphere motor cortex, the prefrontal cortex, and the parietal cortex, and the electromyographic data includes at least one of electromyographic amplitude, envelope, integral electromyography, activation sequence, and co-contraction characteristics.

10. A control method for a limb rehabilitation training system based on a brain-computer interface, characterized in that, The method is applied to the computer device of the brain-computer interface-based limb rehabilitation training system according to any one of claims 1-9, wherein the method includes the following steps: Simultaneously acquire EEG, motor, and electromyography data for rehabilitation goals; Motor control commands are generated based on the electroencephalogram (EEG) data, the motion data, and the electromyogram (EMG) data. The wearable device is controlled based on the motion control commands to assist the non-healthy limb in the rehabilitation target to perform training movements, and the training movements are adjusted based on the feedback data from the wearable device.

Images