A brain-computer interface assisted rehabilitation system based on double feedback closed loop mechanism

Abstract

This invention discloses a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, belonging to the field of rehabilitation technology. It includes a signal acquisition module for acquiring the user's electroencephalogram (EEG) signals and motor state signals during rehabilitation training; a signal processing module, communicatively connected to the signal acquisition module, for preprocessing, feature extraction, and motor intention decoding of the acquired EEG signals; a dual-feedback closed-loop module, communicatively connected to both the signal processing module and the rehabilitation execution module, constructing a bidirectional feedback closed loop; a rehabilitation execution module, communicatively connected to the main control module, for driving the user's limbs to complete targeted rehabilitation movements based on the motor intention decoding results and adjusted training parameters; and a main control module for coordinating the working sequence and data interaction of each module. This invention, employing the aforementioned brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, solves the problems of low decoding accuracy, single feedback, and insufficient personalization in traditional rehabilitation systems.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation technology, and in particular to a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism. Background Technology

[0002] Brain-computer interface (BCI) technology, as a bridge connecting the brain and external devices, offers a new solution for the rehabilitation of motor dysfunction. However, existing BCI-assisted rehabilitation systems often employ a single feedback mechanism (such as visual feedback only or simple mechanical feedback), which has the following drawbacks: It is difficult to balance signal decoding accuracy and real-time performance: Traditional single algorithms (such as simple CSP algorithm or SVM algorithm) do not extract features sufficiently in scenarios with few channels and low sampling rates, and have high computational complexity in scenarios with many channels and high sampling rates, resulting in high latency and unstable accuracy in motion intent decoding. The feedback mechanism is too simplistic: lacking multi-sensory collaborative feedback, users find it difficult to quickly establish a connection between the brain and limb movement, resulting in low efficiency of neural remodeling. Fixed training parameters: It is impossible to dynamically adjust the training difficulty and stimulation parameters according to the user's real-time rehabilitation data, which can easily lead to undertraining or overtraining. Inadequate safety protection mechanisms: Some systems lack real-time monitoring of mechanical driving force, posing a risk of secondary damage; Poor module coordination: High data transmission latency and insufficient synchronization between functional modules affect the overall rehabilitation effect of the system.

[0003] Therefore, there is an urgent need for a brain-computer interface-assisted rehabilitation system with high-precision real-time decoding, multi-dimensional feedback, dynamic parameter adjustment, and safety protection. Summary of the Invention

[0004] The purpose of this invention is to provide a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, comprising: The signal acquisition module is used to acquire the user's electroencephalogram (EEG) signals and motor state signals during rehabilitation training. The EEG signals include motor imagery EEG signals and event-related potential (ERP) signals, while the motor state signals include limb movement angles and movement speed signals. The signal processing module, which communicates with the signal acquisition module, is used to preprocess, extract features, and decode motor intentions of the acquired EEG signals. An improved hybrid algorithm is used to enhance decoding accuracy and real-time performance. The dual feedback closed-loop module is connected to the signal processing module and the rehabilitation execution module respectively. It includes a multi-sensory channel feedback unit and a neural remodeling feedback unit to construct a two-way feedback closed loop. The multi-sensory channel feedback unit provides multi-dimensional visual, auditory and kinesthetic feedback, and the neural remodeling feedback unit adjusts the training parameters according to the rehabilitation data. The rehabilitation execution module, which communicates with the main control module, includes a mechanical rehabilitation device used to drive the user's limbs to complete targeted rehabilitation movements based on the decoding results of the movement intention and the adjusted training parameters. The main control module is connected to the signal acquisition module, signal processing module, dual feedback closed-loop module, and rehabilitation execution module to coordinate the working sequence and data interaction of each module.

[0006] Preferably, the signal acquisition module includes an EEG acquisition unit and a motion state acquisition unit; The EEG acquisition unit uses a non-invasive EEG cap, based on the international 10-20 lead system for electrode distribution. The acquisition channels include at least C3, C4, Cz, AF3, AF4, F7, and F8 channels, with a sampling rate of 250Hz-512Hz, supporting the simultaneous acquisition of EEG signals, electrooculogram (EOG) signals, and electromyogram (EMG) signals. The motion state acquisition unit uses an inertial measurement unit and an angle sensor, which are installed on the rehabilitation execution module and key parts of the user's limbs. The acquisition frequency is greater than or equal to 100Hz, and it is used to acquire limb motion angle, angular velocity and motion trajectory data in real time.

[0007] Preferably, the improved hybrid algorithm of the signal processing module includes an improved LBP-SVM algorithm for low sampling rate scenarios with few channels, and an ERP-CSP algorithm for high sampling rate scenarios with many channels. The improved LBP-SVM algorithm divides the frequency band into 8-30Hz segments using a filter bank, employs a sliding window to increase the number of features, and calculates the logarithmic frequency band energy characteristics. ; Where N is the number of sampling points within the window, Let i be the signal value of the i-th sampling point within the window; combining the Pratt scaling probability output formula with the voting strategy to optimize the classification output, the Pratt scaling probability output formula is: ; Where f is the original output of SVM, and A and B are the scaling parameters for cross-validation optimization; The ERP-CSP algorithm extracts the P300 potential and the lateral preparatory potential, locates the starting point of motion imagination, eliminates invalid samples, and extracts spatial domain features through the co-space mode, thereby reducing computational complexity.

[0008] Preferably, the spatial domain feature extraction process of the ERP-CSP algorithm includes: Calculate the covariance matrix of the two types of motion-imagined signals: ; in, Let c be the covariance matrix of the c-th type of signal, where c=1 corresponds to left-hand motor imagery. c=2 corresponds to right-hand movement visualization. Let c be the number of time samples of the c-th type signal. Let c be the channel-time matrix of the signal. for The transpose of the matrix; For the mixed covariance matrix Perform eigenvalue decomposition Construct the whitening matrix: ; in, Mixed covariance matrix The feature vector matrix has dimensions of . Each column is a feature vector, corresponding to the spatial distribution pattern of brain signals; Mixed covariance matrix The eigenvalue diagonal matrix has dimensions of The elements on the diagonal are eigenvalues, arranged in descending order, reflecting the contribution strength of the corresponding eigenvectors; Eigenvector matrix The transpose of the matrix; For eigenvalue diagonal matrices The inverse square root matrix, with dimension . Through the The inverse square root of each eigenvalue on the diagonal is obtained; Through whitening transformation After obtaining the whitened signal, calculate the covariance matrix of the two classes of signals after whitening: ; in, The raw EEG signal, Let be the covariance matrix of the whitened c-th class signal. This is the matrix of the c-th class signal after whitening; right and Joint eigenvalue decomposition, taking the first m and last m eigenvectors to construct a spatial filter. Feature extraction: ; in, This is the spatial filter matrix; Spatial filter matrix The transpose of the matrix, for The transpose of .

[0009] Preferably, the multi-sensory feedback unit includes a virtual reality feedback subunit, an auditory feedback subunit, and a kinesthetic feedback subunit; The virtual reality feedback subunit constructs a rehabilitation training virtual scene based on the Unity engine, including a model training scene and a rehabilitation training scene. The model training scene uses a first-person perspective hand / leg model, while the rehabilitation training scene includes scenarios simulating daily actions such as picking up a water cup and blocking a small ball. The auditory feedback subunit outputs voice prompts based on the results of motion intention recognition and the completion of rehabilitation actions. The kinesthetic feedback subunit, through the mechanical drive of the rehabilitation execution module, synchronously drives the user's limb movements, realizing the synergy between kinesthetic feedback and virtual scene feedback.

[0010] Preferably, the neural remodeling feedback unit includes a rehabilitation data assessment subunit and a parameter adjustment subunit; The rehabilitation data assessment subunit analyzes the accuracy of motor intention decoding, rehabilitation movement completion time, and limb range of motion data in real time to establish a rehabilitation effect assessment model. The parameter adjustment subunit dynamically adjusts the difficulty of training tasks, stimulation parameters, and feedback intensity based on the assessment results. Stimulation parameters include the frequency, pulse width, and amplitude of electrical stimulation, while feedback intensity includes the virtual scene animation frame rate and mechanical drive force.

[0011] Preferably, the rehabilitation execution module includes an upper limb rehabilitation sub-module and a lower limb rehabilitation sub-module, and adopts a multi-degree-of-freedom mechanical structure. The upper limb rehabilitation sub-module supports flexion, extension and rotation movements of the wrist joint and metacarpophalangeal joint, while the lower limb rehabilitation sub-module supports flexion and extension movements of the knee joint and ankle joint. The rehabilitation execution module has a built-in force feedback sensor to detect the mechanical driving force in real time. When the force exceeds the safety threshold, the movement will automatically stop to avoid secondary injury.

[0012] Preferably, the main control module adopts a multi-process architecture and uses the Lab Streaming Layer protocol to realize data transmission and real-time communication and synchronization between modules. The main control module also includes a human-computer interaction unit, which provides a dual interface of control terminal and user terminal. The control terminal can set training parameters and monitor rehabilitation data, while the user terminal displays virtual training scenarios and operation instructions.

[0013] Preferably, the preprocessing process of the signal processing module includes rereference, DC component removal, baseline drift removal, bandpass filtering, and notch filtering. The bandpass filter uses an FIR filter with a cutoff frequency of 8-30Hz, and the notch filter is used to filter out 50Hz power frequency interference.

[0014] Preferably, the workflow of the dual feedback closed-loop module is as follows: S1. The signal acquisition module acquires EEG signals and motion state signals and transmits them to the signal processing module. S2, the signal processing module preprocesses and extracts features from the EEG signal, completes the decoding of motion intention through an improved hybrid algorithm, and outputs the decoding result to the main control module; S3. The main control module sends the decoding results to the rehabilitation execution module, drives the mechanical device to move the user's limbs, and transmits the decoding results and motion status signals to the dual feedback closed loop module. S4. The multi-sensory feedback unit outputs multi-dimensional visual, auditory and kinesthetic feedback based on the decoding results, and the neural remodeling feedback unit adjusts the training parameters based on the rehabilitation data assessment results. S5. The adjusted training parameters are fed back to the signal processing module and the rehabilitation execution module through the main control module, forming a closed-loop adjustment.

[0015] Therefore, the present invention employs the above-mentioned brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, which has the following beneficial effects: (1) An improved hybrid algorithm is adopted to adaptively select the optimal decoding scheme for different scenarios with different number of channels and sampling rates. The improved LBP-SVM algorithm improves the integrity of feature extraction in scenarios with few channels by frequency band segmentation, sliding window feature enhancement and probability output optimization; the ERP-CSP algorithm reduces the computational complexity in scenarios with multiple channels by potential extraction, invalid sample removal and spatial domain feature optimization. Actual tests show that the system decoding accuracy is stable at over 85% and the single frame data processing time is ≤0.1s, which meets the real-time requirements of online rehabilitation training; (2) Design a two-way closed-loop mechanism of multi-sensory channel feedback and neural remodeling feedback. The multi-sensory channel feedback unit integrates virtual reality, auditory cues and mechano-kinesthetic feedback to construct a multi-dimensional collaborative perception of vision, hearing and movement, helping users to quickly establish the connection between brain movement intention and limb movement; the neural remodeling feedback unit dynamically adjusts the training difficulty, electrical stimulation parameters and feedback intensity by analyzing data such as decoding accuracy, movement completion time and limb movement range in real time, so as to achieve adaptive training of "evaluation-adjustment-optimization" and improve the efficiency of neural remodeling. (3) The signal acquisition module supports the simultaneous acquisition of EEG, EOG and EMG signals. The motion state acquisition covers multi-dimensional data such as limb angle, angular velocity and motion trajectory, providing rich basis for personalized assessment. The rehabilitation execution module is divided into upper limb and lower limb sub-modules to adapt to users with different motor function disorders. The main control module provides dual interface interaction. Doctors can customize training plans through the control terminal, and users can obtain intuitive operation guidance through the user terminal to realize personalized rehabilitation of "one person, one policy".

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of an embodiment of a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to the present invention. Figure 2 This is a schematic diagram illustrating the accuracy of an embodiment of a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example Please see Figure 1 This invention provides a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, comprising: The signal acquisition module is used to acquire the user's electroencephalogram (EEG) signals and motor state signals during rehabilitation training. The EEG signals include motor imagery EEG signals and event-related potential (ERP) signals, while the motor state signals include limb movement angles and speed signals. The signal acquisition module includes an EEG acquisition unit and a motor state acquisition unit. The EEG acquisition unit uses a non-invasive EEG cap, with electrodes distributed based on the international 10-20 lead system. The acquisition channels include at least C3, C4, C2, AF3, AF4, F7, and F8 channels, with a sampling rate of 250Hz-512Hz. It supports the simultaneous acquisition of EEG signals, EEG signals, and EMG signals. The electrodes are silver / silver chloride wet electrodes, which are used with conductive gel to ensure good contact with the scalp, and the impedance is controlled below 5kΩ.

[0021] The motion acquisition unit employs an inertial measurement unit (IMU) and angle sensors, installed on the rehabilitation execution module and key parts of the user's limbs. The acquisition frequency is greater than or equal to 100Hz, used to acquire real-time data on limb movement angles, angular velocities, and motion trajectories. In this embodiment, a MEMS inertial measurement unit (IMU) and a high-precision angle sensor are selected and installed on the user's left forearm, wrist joint, and rehabilitation robotic hand, respectively. The acquisition frequency is set to 100Hz, acquiring real-time data on limb movement angles (measurement range ±180°, accuracy ±0.5°), angular velocities (measurement range ±2000° / s, accuracy ±0.1° / s), and motion trajectories.

[0022] The signal processing module, communicating with the signal acquisition module, is used for preprocessing, feature extraction, and motion intent decoding of the acquired EEG signals. An improved hybrid algorithm is employed to enhance decoding accuracy and real-time performance. This embodiment uses an Intel Core i7-12700H processor (14 cores, 20 threads), 16GB of DDR5 memory, and runs Windows 11. The signal processing program is developed using Python 3.9, integrating improved LBP-SVM and ERP-CSP algorithms. Multi-process parallel processing improves data processing efficiency, with a single-frame data processing time ≤0.1s, meeting the real-time requirements of online rehabilitation training.

[0023] The improved hybrid algorithms include the improved LBP-SVM algorithm for scenarios with few channels and low sampling rates, and the ERP-CSP algorithm for scenarios with many channels and high sampling rates. The improved LBP-SVM algorithm divides the frequency band into 8-30Hz segments using a filter bank, employs a sliding window to increase the number of features, and calculates the logarithmic frequency band energy characteristics. ; Where N is the number of sampling points within the window, Let i be the signal value of the i-th sampling point within the window; combining the Pratt scaling probability output formula with the voting strategy to optimize the classification output, the Pratt scaling probability output formula is: ; Where f is the original output of SVM, and A and B are the scaling parameters for cross-validation optimization; The ERP-CSP algorithm extracts the P300 potential and the lateral preparatory potential, locates the starting point of motion imagination, eliminates invalid samples, and extracts spatial domain features through the co-space mode, thereby reducing computational complexity.

[0024] The spatial domain feature extraction process of the ERP-CSP algorithm includes: Calculate the covariance matrix of the two types of motion-imagined signals: ; in, Let c be the covariance matrix of the c-th type of signal, where c=1 corresponds to left-hand motor imagery. c=2 corresponds to right-hand movement visualization. Let c be the number of time samples of the c-th type signal. Let c be the channel-time matrix of the signal. for The transpose of the matrix; For the mixed covariance matrix Perform eigenvalue decomposition Construct the whitening matrix: ; in, Mixed covariance matrix The feature vector matrix has dimensions of . Each column is a feature vector, corresponding to the spatial distribution pattern of brain signals; Mixed covariance matrix The eigenvalue diagonal matrix has dimensions of The elements on the diagonal are eigenvalues, arranged in descending order, reflecting the contribution strength of the corresponding eigenvectors; Eigenvector matrix The transpose of the matrix; For eigenvalue diagonal matrices The inverse square root matrix, with dimension . Through the The inverse square root of each eigenvalue on the diagonal is obtained; Through whitening transformation After obtaining the whitened signal, calculate the covariance matrix of the two classes of signals after whitening: ; in, The raw EEG signal, Let be the covariance matrix of the whitened c-th class signal. This is the matrix of the c-th class signal after whitening; right and Joint eigenvalue decomposition, taking the first m and last m eigenvectors to construct a spatial filter. Feature extraction: ; in, This is the spatial filter matrix; Spatial filter matrix The transpose of the matrix, for The transpose of .

[0025] This embodiment improves the LBP-SVM algorithm: Frequency band segmentation: The 8-30Hz frequency band is divided into 11 sub-bands (8-10Hz, 10-12Hz...28-30Hz) by a filter bank, using a 110th-order Kaiser window FIR bandpass filter (Beta parameter 0.01).

[0026] Sliding window: window length 1s, step size 0.5s, generating multiple sets of feature vectors for each sub-band.

[0027] SVM parameters: The optimal parameters are determined using a Gaussian kernel function and 5-fold cross-validation (penalty coefficient k=10, kernel width). Pratt scaling parameters A=0.8, B=0.2, and confidence threshold set to 0.7.

[0028] ERP-CSP Algorithm: Covariance matrix calculation: number of time samples (Corresponding to a 3-second motion visualization task, 250Hz sampling rate), number of channels L=7 (core acquisition channels).

[0029] Spatial filter: Take the first two and last two feature vectors to construct a spatial filter W (7×4 dimension), and extract feature vectors with a dimension of 4×1.

[0030] The dual-feedback closed-loop module communicates with both the signal processing module and the rehabilitation execution module. It includes a multi-sensory feedback unit and a neural remodeling feedback unit, constructing a two-way feedback loop. The multi-sensory feedback unit provides multi-dimensional visual, auditory, and kinesthetic feedback, while the neural remodeling feedback unit adjusts training parameters based on rehabilitation data. The multi-sensory feedback unit includes virtual reality feedback subunits, auditory feedback subunits, and kinesthetic feedback subunits.

[0031] The virtual reality feedback subunit constructs a rehabilitation training virtual scene based on the Unity engine, including a model training scene and a rehabilitation training scene. The model training scene uses a first-person perspective hand / leg model, while the rehabilitation training scene includes scenarios simulating daily actions such as picking up a water cup and blocking a small ball. The auditory feedback subunit outputs voice prompts based on the results of motion intention recognition and the completion of rehabilitation actions. The kinesthetic feedback subunit, through the mechanical drive of the rehabilitation execution module, synchronously drives the user's limb movements, realizing the synergy between kinesthetic feedback and virtual scene feedback.

[0032] In this embodiment, the virtual reality feedback subunit is built based on the Unity 2021.3 engine to construct rehabilitation training scenarios, including "model training scenario" and two rehabilitation training scenarios: "picking up a water cup" and "blocking a small ball". It runs on a VR all-in-one machine (resolution 2160×2160, refresh rate 90Hz) and presents the hand model and training scenario from a first-person perspective.

[0033] Auditory feedback subunit: Uses a portable Bluetooth speaker to output voice prompts (such as "Please imagine grasping with your right hand" "Training successful"), with a voice frequency range of 300Hz-3kHz, and the volume can be adjusted by the main control module (30-80dB).

[0034] Kinesthetic feedback subunit: Linked with the rehabilitation execution module, it synchronously drives the user's limb movement through mechanical driving force, with a kinesthetic feedback delay of ≤0.3s.

[0035] The neural remodeling feedback unit includes a rehabilitation data assessment subunit and a parameter adjustment subunit; The rehabilitation data assessment subunit analyzes in real time the accuracy of motor intention decoding, rehabilitation movement completion time, and limb range of motion data to establish a rehabilitation effect assessment model. The parameter adjustment subunit dynamically adjusts the training task difficulty, stimulation parameters, and feedback intensity based on the assessment results. Stimulation parameters include the frequency, pulse width, and amplitude of electrical stimulation, while feedback intensity includes the virtual scene animation frame rate and mechanical drive force. The main control module analyzes rehabilitation data in real time, establishes an assessment model, and dynamically adjusts training parameters. The parameter adjustment step size is as follows: training task difficulty (levels 1-10, step size 1), electrical stimulation frequency (20-50Hz, step size 5Hz), pulse width (0.2-0.6ms, step size 0.1ms), and amplitude (20-80V, step size 10V).

[0036] The rehabilitation execution module, communicating with the main control module, includes a mechanical rehabilitation device used to drive the user's limbs to complete targeted rehabilitation movements based on the decoding results of the movement intention and the adjusted training parameters. The rehabilitation execution module comprises an upper limb rehabilitation submodule and a lower limb rehabilitation submodule, employing a multi-degree-of-freedom mechanical structure. The upper limb rehabilitation submodule supports flexion, extension, and rotation movements of the wrist and metacarpophalangeal joints, while the lower limb rehabilitation submodule supports flexion and extension movements of the knee and ankle joints. The rehabilitation execution module incorporates a force feedback sensor to detect the mechanical driving force in real time; when the force exceeds a safe threshold, the movement automatically stops to prevent secondary injury.

[0037] The main control module communicates with the signal acquisition module, signal processing module, dual feedback closed-loop module, and rehabilitation execution module to coordinate the working sequence and data interaction of each module. The main control module adopts a multi-process architecture, using the LabStreaming Layer protocol to achieve data transmission and real-time communication and synchronization between modules. The main control module also includes a human-computer interaction unit, providing a dual interface of control and user. The control end allows setting training parameters and monitoring rehabilitation data, while the user end displays a virtual training scenario and operation instructions.

[0038] The preprocessing process of the signal processing module includes rereference, DC component removal, baseline drift removal, bandpass filtering, and notch filtering. The bandpass filtering uses an FIR filter with a cutoff frequency of 8-30Hz, and the notch filtering is used to filter out 50Hz power frequency interference.

[0039] The above dual-feedback closed-loop workflow is as follows: S1. The signal acquisition module acquires EEG signals and motion state signals and transmits them to the signal processing module. S2, the signal processing module preprocesses and extracts features from the EEG signal, completes the decoding of motion intention through an improved hybrid algorithm, and outputs the decoding result to the main control module; S3. The main control module sends the decoding results to the rehabilitation execution module, drives the mechanical device to move the user's limbs, and transmits the decoding results and motion status signals to the dual feedback closed loop module. S4. The multi-sensory feedback unit outputs multi-dimensional visual, auditory and kinesthetic feedback based on the decoding results, and the neural remodeling feedback unit adjusts the training parameters based on the rehabilitation data assessment results. S5. The adjusted training parameters are fed back to the signal processing module and the rehabilitation execution module through the main control module, forming a closed-loop adjustment.

[0040] The process in this embodiment is as follows: Signal acquisition: The signal acquisition module simultaneously acquires the user's motor imagery EEG signals (such as imagining the right hand grasping) and the left upper limb motor state signals, and transmits them wirelessly to the signal processing module.

[0041] Signal processing and decoding: After preprocessing the EEG signal, the signal processing module automatically selects an algorithm based on the number of acquisition channels and the sampling rate (in this embodiment, it is a multi-channel scenario and the ERP-CSP algorithm is used), extracts spatial domain features, completes motion intent decoding, and outputs decoding results such as "grasp" and "extend" to the main control module.

[0042] Rehabilitation execution: The main control module converts the decoding results into control commands and sends them to the rehabilitation execution module. The mechanical device drives the user's left upper limb to complete the corresponding movements (such as flexion of the metacarpophalangeal joints and adduction of the wrist joint when grasping). The force feedback sensor monitors the driving force in real time.

[0043] Dual feedback regulation: Multi-sensory feedback: The VR scene synchronously presents the virtual hand grasping animation, the Bluetooth speaker outputs the "training successful" voice prompt, and the kinematic feedback of the mechanical device allows the user to perceive the body movement, forming a multi-dimensional feedback of visual, auditory and kinematic senses.

[0044] Neural remodeling feedback: The neural remodeling feedback unit analyzes the decoding accuracy (target ≥80%), action completion time (target ≤3s), and limb movement angle (target wrist flexion ≥60°) in real time. If the decoding accuracy is below 70% for three consecutive sets, the training task difficulty is automatically reduced by 1 level, and the electrical stimulation amplitude is increased by 10V. If the decoding accuracy is above 85% for five consecutive sets, the training task difficulty is increased by 1 level to enhance the training challenge. Two subjects (p1, p2) performed the operation according to the experimental protocol. The decoding accuracy is as follows: Figure 2 .

[0045] Closed-loop optimization: The adjusted training parameters are fed back to the signal processing module and the rehabilitation execution module through the main control module. The signal processing module updates the algorithm parameters to adapt to the user's current state, and the rehabilitation execution module adjusts the driving force and range of motion to form a continuously optimized closed-loop training.

[0046] Therefore, the present invention employs a brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, which effectively solves the problems of low decoding accuracy, single feedback, and lack of personalization in traditional rehabilitation systems, and provides an efficient, safe, and personalized rehabilitation solution for patients with limb motor dysfunction.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism, characterized in that, include: The signal acquisition module is used to acquire the user's electroencephalogram (EEG) signals and motor state signals during rehabilitation training. The EEG signals include motor imagery EEG signals and event-related potential (ERP) signals, while the motor state signals include limb movement angles and movement speed signals. The signal processing module, which communicates with the signal acquisition module, is used to preprocess, extract features, and decode motor intentions of the acquired EEG signals. An improved hybrid algorithm is used to enhance decoding accuracy and real-time performance. The dual feedback closed-loop module is connected to the signal processing module and the rehabilitation execution module respectively. It includes a multi-sensory channel feedback unit and a neural remodeling feedback unit to construct a two-way feedback closed loop. The multi-sensory channel feedback unit provides multi-dimensional visual, auditory and kinesthetic feedback, and the neural remodeling feedback unit adjusts the training parameters according to the rehabilitation data. The rehabilitation execution module, which communicates with the main control module, includes a mechanical rehabilitation device used to drive the user's limbs to complete targeted rehabilitation movements based on the decoding results of the movement intention and the adjusted training parameters. The main control module is connected to the signal acquisition module, signal processing module, dual feedback closed-loop module, and rehabilitation execution module to coordinate the working sequence and data interaction of each module.

2. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The signal acquisition module includes an EEG acquisition unit and a motion state acquisition unit; The motion state acquisition unit uses an inertial measurement unit and an angle sensor, which are installed on the rehabilitation execution module and key parts of the user's limbs. The acquisition frequency is greater than or equal to 100Hz, and it is used to acquire limb motion angle, angular velocity and motion trajectory data in real time.

3. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 2, characterized in that: The improved hybrid algorithms of the signal processing module include the improved LBP-SVM algorithm for low-channel, low-sampling-rate scenarios, and the ERP-CSP algorithm for high-channel, high-sampling-rate scenarios. The improved LBP-SVM algorithm divides the frequency band into 8-30Hz segments using a filter bank, employs a sliding window to increase the number of features, and calculates the logarithmic frequency band energy characteristics. ； Where N is the number of sampling points within the window, Let i be the signal value of the i-th sampling point within the window; combining the Pratt scaling probability output formula with the voting strategy to optimize the classification output, the Pratt scaling probability output formula is: ； Where f is the original output of SVM, and A and B are the scaling parameters for cross-validation optimization; The ERP-CSP algorithm extracts the P300 potential and the lateral preparatory potential, locates the starting point of motion imagination, eliminates invalid samples, and extracts spatial domain features through the co-space mode, thereby reducing computational complexity.

4. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 3, characterized in that, The spatial domain feature extraction process of the ERP-CSP algorithm includes: Calculate the covariance matrix of the two types of motion-imagined signals: ； in, Let c be the covariance matrix of the c-th type of signal, where c=1 corresponds to left-hand motor imagery. c=2 corresponds to right-hand movement visualization. Let c be the number of time samples of the c-th type signal. Let c be the channel-time matrix of the signal. for The transpose of the matrix; For the mixed covariance matrix Perform eigenvalue decomposition Construct the whitening matrix: ； in, Mixed covariance matrix The feature vector matrix has dimensions of . Each column is a feature vector, corresponding to the spatial distribution pattern of brain signals; Mixed covariance matrix The eigenvalue diagonal matrix has dimensions of The elements on the diagonal are eigenvalues, arranged in descending order, reflecting the contribution strength of the corresponding eigenvectors; Eigenvector matrix The transpose of the matrix; For eigenvalue diagonal matrices The inverse square root matrix, with dimension . Through the The inverse square root of each eigenvalue on the diagonal is obtained; Through whitening transformation After obtaining the whitened signal, calculate the covariance matrix of the two classes of signals after whitening: ； in, The raw EEG signal, Let be the covariance matrix of the whitened c-th class signal. This is the matrix of the c-th class signal after whitening; right and Joint eigenvalue decomposition, taking the first m and last m eigenvectors to construct a spatial filter. Feature extraction: ； in, This is the spatial filter matrix; Spatial filter matrix The transpose of the matrix, for The transpose of .

5. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The multisensory feedback unit includes a virtual reality feedback subunit, an auditory feedback subunit, and a kinesthetic feedback subunit; The virtual reality feedback subunit is built on the Unity engine to construct a virtual scene for rehabilitation training, including a model training scene and a rehabilitation training scene. The model training scene adopts a first-person perspective model. The auditory feedback subunit outputs voice prompts based on the results of motor intention recognition and the completion of rehabilitation actions; the kinesthetic feedback subunit, through the mechanical drive of the rehabilitation execution module, synchronously drives the user's limb movements, realizing the synergy between kinesthetic feedback and virtual scene feedback.

6. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The neural remodeling feedback unit includes a rehabilitation data assessment subunit and a parameter adjustment subunit; The rehabilitation data assessment subunit analyzes the accuracy of motor intention decoding, rehabilitation movement completion time, and limb range of motion data in real time to establish a rehabilitation effect assessment model. The parameter adjustment subunit dynamically adjusts the difficulty of training tasks, stimulation parameters, and feedback intensity based on the assessment results. Stimulation parameters include the frequency, pulse width, and amplitude of electrical stimulation, while feedback intensity includes the virtual scene animation frame rate and mechanical drive force.

7. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The rehabilitation execution module includes an upper limb rehabilitation sub-module and a lower limb rehabilitation sub-module. It adopts a multi-degree-of-freedom mechanical structure. The upper limb rehabilitation sub-module supports flexion, extension and rotation movements of the wrist and metacarpophalangeal joints, while the lower limb rehabilitation sub-module supports flexion and extension movements of the knee and ankle joints. The rehabilitation execution module has a built-in force feedback sensor that detects the mechanical drive force in real time. When the force exceeds the safety threshold, the exercise is automatically stopped to avoid secondary injury.

8. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The main control module adopts a multi-process architecture and uses the Lab Streaming Layer protocol to realize data transmission and real-time communication and synchronization between modules. The main control module also includes a human-computer interaction unit, which provides a dual interface of control terminal and user terminal. The control terminal can set training parameters and monitor rehabilitation data, while the user terminal displays virtual training scenarios and operation instructions.

9. The brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that: The preprocessing process of the signal processing module includes rereference, DC component removal, baseline drift removal, bandpass filtering, and notch filtering. The bandpass filtering uses an FIR filter with a cutoff frequency of 8-30Hz, and the notch filtering is used to filter out 50Hz power frequency interference.

10. A brain-computer interface-assisted rehabilitation system based on a dual-feedback closed-loop mechanism according to claim 1, characterized in that, The workflow of the dual feedback closed-loop module is as follows: S1. The signal acquisition module acquires EEG signals and motion state signals and transmits them to the signal processing module. S2, the signal processing module preprocesses and extracts features from the EEG signal, completes the decoding of motion intention through an improved hybrid algorithm, and outputs the decoding result to the main control module; S3. The main control module sends the decoding results to the rehabilitation execution module, drives the mechanical device to move the user's limbs, and transmits the decoding results and motion status signals to the dual feedback closed loop module. S4. The multi-sensory feedback unit outputs multi-dimensional visual, auditory and kinesthetic feedback based on the decoding results, and the neural remodeling feedback unit adjusts the training parameters based on the rehabilitation data assessment results. S5. The adjusted training parameters are fed back to the signal processing module and the rehabilitation execution module through the main control module, forming a closed-loop adjustment.

Images