Domestic rehabilitation training method and device based on lower limb exoskeleton and storage medium
By combining non-invasive brain-computer interfaces, surface electromyography patches, and flexible plantar pressure insoles to acquire multimodal signals, and by using sliding window technology and reinforcement learning models to optimize exoskeleton assistance strategies, the problem that existing rehabilitation training equipment cannot meet the needs of home rehabilitation has been solved, and closed-loop rehabilitation training and neural remodeling effects with active patient participation have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XUANWU HOSPITAL OF CAPITAL UNIV OF MEDICAL SCI
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-15
AI Technical Summary
Existing rehabilitation training equipment is large, complex, and expensive, failing to meet patients' needs for convenient and continuous rehabilitation training at home. Furthermore, EEG, sEMG, and foot pressure data lack collaborative optimization strategies, making it impossible to form a closed loop and hindering the promotion of neural remodeling.
By acquiring EEG signals through a non-invasive brain-computer interface, muscle signals through surface electromyography (EMG) patches, and foot pressure signals through flexible plantar pressure insoles, combined with sliding window technology and reinforcement learning models (PPO algorithm), a multimodal data covariance matrix is formed to optimize the exoskeleton assistance strategy in real time. Furthermore, through biofeedback visualization and personalized CNN models to identify user intentions, a closed-loop control is achieved.
It enables patients to actively participate in closed-loop rehabilitation training, enhances the effect of neural remodeling, meets the needs of gait control, and is suitable for home environments.
Smart Images
Figure CN122031231A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rehabilitation training technology, specifically to home rehabilitation training methods, devices, and storage media based on lower limb exoskeletons. Background Technology
[0002] Currently, a large number of patients suffer from lower limb muscle weakness due to various diseases such as spinal cord injury, traumatic brain injury, and stroke. Traditional rehabilitation training methods are mostly passive, with low patient participation and limited rehabilitation effects. Furthermore, existing rehabilitation equipment is often bulky, complex, and expensive, suitable only for specialized medical institutions and unable to meet patients' needs for convenient and continuous rehabilitation training at home. With the continuous development of brain-computer interface (BCI) and exoskeleton technologies, new approaches have been provided to address these issues. However, products and methods that effectively combine non-invasive BCIs with lower limb exoskeletons for home rehabilitation training are still relatively lacking.
[0003] Modern neuroscience research has confirmed that the central nervous system possesses significant plasticity. Repetitive, task-oriented training can promote brain functional reorganization and neural pathway reconstruction. However, this neural remodeling process requires three key conditions: the training intensity must reach a threshold; the training process requires active patient participation; and the training effect needs real-time feedback for reinforcement.
[0004] Brain-computer interface (BCI) is a technology that establishes a direct communication pathway between the brain and external devices. Its core principle is to collect and decode neural activity signals from the brain, converting them into control commands to ultimately control devices such as computers and mechanical devices. Invasive BCI requires surgical implantation of electrodes, posing risks of infection and rejection. Non-invasive BCI has poor home-use compatibility; wet electrodes require conductive gel, making wearing complex, time-consuming, and uncomfortable. Furthermore, it lacks closed-loop feedback, making it impossible to verify in real time whether motor intentions are translated into actual muscle activation (e.g., missing sEMG signals).
[0005] Lower limb exoskeletons are mechanical devices that assist patients in walking, but existing products mostly rely on preset programs or manual control, lacking active patient participation.
[0006] Although existing studies have attempted to combine EEG and sEMG (such as CN202411727581), the signals are analyzed in a fragmented manner, and EEG, sEMG, and foot pressure data are processed independently. No collaborative optimization strategy has been established, which cannot meet the needs of gait control and has not formed a closed loop of "intention recognition → action execution → effect feedback", making it difficult to promote neural remodeling. Summary of the Invention
[0007] In view of this, the purpose of the present invention is to provide a home rehabilitation training method, device and storage medium based on a lower limb exoskeleton, so as to solve the problems of the independent signals of EEG, sEMG and foot pressure data in traditional methods, the lack of a collaborative optimization strategy, the inability to meet the needs of gait control, the failure to form a closed loop of "intention recognition, action execution and effect feedback", and the difficulty in promoting neural remodeling.
[0008] According to a first aspect of the present invention, a home-based rehabilitation training method based on a lower limb exoskeleton is provided, the method comprising: The system acquires the user's brainwave signals through a non-invasive brain-computer interface worn by the user, acquires multiple muscle signals through surface electromyography patches on the user's muscles, and acquires the user's foot pressure signals through flexible foot pressure insoles worn by the user. The electroencephalogram (EEG) signal, muscle signal, and foot pressure signal are preprocessed respectively, and the EEG feature vector, muscle feature vector, and foot pressure feature vector are extracted from the preprocessed EEG signal, muscle signal, and foot pressure signal respectively. Based on the aforementioned EEG feature vector, muscle feature vector, and foot pressure feature vector, the sliding window technique is used to obtain the EEG-EM covariance matrix, EEG-foot pressure covariance matrix, and EEM-foot pressure covariance matrix, respectively. The EEG-EM Covariance Matrix, EEG-Foot Pressure Covariance Matrix, and EMG-Foot Pressure Covariance Matrix are fused to obtain the final covariance matrix, and real-time multimodal features are obtained through the final covariance matrix. The real-time multimodal features are input into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angles. The EEG feature vector, muscle feature vector, and foot pressure feature vector are visualized using biofeedback. The user adjusts their force based on the biofeedback visualization, and the EEG signals of the force adjustment are collected. The EEG signals are then input into a pre-trained personalized CNN model, which identifies the user's intention to walk or stop. When the user's intention is to walk, the exoskeleton assists the user in taking one step based on the exoskeleton joint angles and the exoskeleton assistance level.
[0009] Preferably, The preprocessing of the electroencephalogram (EEG) signals, muscle signals, and foot pressure signals includes: The EEG signal was subjected to bandpass filtering to remove low-frequency and high-frequency noise, and the filtered EEG signal was subjected to independent component analysis to remove electrooculogram artifacts. The muscle signal was filtered using a second-order Butterworth bandpass filter to eliminate baseline drift and high-frequency noise; The foot pressure signal is read from the FSR resistance value via the I2C interface. After noise reduction by Kalman filtering, the resistance value is converted into a pressure value according to the calibration parameters.
[0010] Preferably, The extraction of EEG feature vectors, muscle feature vectors, and foot pressure feature vectors from preprocessed EEG signals, muscle signals, and foot pressure signals includes: The μ / β band energy values were extracted from the preprocessed EEG signals and processed by CSP spatial filtering. The filtered EEG signals were then centered and covariance standardized. The first 8 pairs of spatial patterns were extracted by generalized feature decomposition optimization filter to obtain a 16-dimensional EEG feature vector. Obtain the root mean square value of each group of muscle signals after preprocessing, and map the root mean square value of each group of muscle signals to the interval [0,1] to obtain a 5-dimensional muscle feature vector. The preprocessed pressure values are standardized, and the first three principal components are retained by PCA dimensionality reduction, thereby reducing the 16 pressure values to 3 dimensions and obtaining a 3-dimensional foot pressure feature vector.
[0011] Preferably, The training of the PPO model includes: Acquire and preprocess the user's historical EEG signals, historical muscle signals, and historical foot pressure signals; Historical EEG feature vectors, historical muscle feature vectors, and historical foot pressure feature vectors are obtained based on the preprocessed historical EEG signals, historical muscle signals, and historical foot pressure signals. Based on the historical EEG feature vector, historical muscle feature vector, and historical foot pressure feature vector, respectively establish the historical EEG-EMG covariance matrix, the historical EEG-foot pressure covariance matrix, and the historical EEG-foot pressure covariance matrix. The historical EEG-EMG covariance matrix, historical EEG-foot pressure covariance matrix, and historical EMG-foot pressure covariance matrix are fused to obtain the historical final covariance matrix, and historical multimodal features are obtained through the historical final covariance matrix. The pre-built PPO model is iteratively trained using historical multimodal features based on a preset reward function to obtain a trained PPO model.
[0012] Preferably, The training of the personalized CNN model includes: A pre-built lightweight CNN model is trained using a publicly available EEG database, and then the lightweight CNN model is adaptively calibrated based on a preset training task to obtain a personalized CNN model.
[0013] Preferably, The step of adaptively calibrating the lightweight CNN model based on a preset training task to obtain a personalized CNN model includes: Users perform a 5-minute visual leg-raising task in sequence. During the training task, multiple sets of EEG signals are collected as fine-tuning datasets. Freeze the convolutional layers of the lightweight CNN model, set the learning rate and number of iterations of the fully connected layers of the lightweight CNN model, and perform transfer learning on the lightweight CNN model by fine-tuning the dataset to obtain a personalized CNN model.
[0014] Preferably, it further includes: When a sudden muscle spasm or severe shift in the center of gravity is detected, the exoskeleton is braked urgently. The detection of sudden muscle spasm includes: when the root mean square value of any set of muscle signals is greater than a preset threshold and the duration exceeds a preset time threshold, it is determined to be a sudden muscle spasm. The detection of severe center of gravity shift includes: calculating the foot center of gravity shift based on the foot pressure signal; if the foot center of gravity shift exceeds a preset threshold and its duration exceeds a preset time threshold, it is determined to be a severe center of gravity shift.
[0015] Preferably, The step of calculating the foot center of gravity offset based on the foot pressure signal includes: The flexible foot pressure insole uses a 16-point FSR array; Obtain the coordinates of each FSR point of the left and right feet in the X direction during the user's movement, and obtain the normalized pressure value of each FSR point of the left and right feet; Calculate the centroid of the left and right feet and the overall plantar pressure center based on the coordinates of each FSR point in the X direction and the normalized pressure value of each FSR point. Calculate the reference center based on the X-axis center of gravity of the left and right feet; The offset of the plantar center of gravity is calculated based on the reference center and the overall plantar pressure center.
[0016] According to a second aspect of the present invention, a home rehabilitation training device based on a lower limb exoskeleton is provided, the device comprising: Data acquisition module: used to acquire the user's brainwave signals through the non-invasive brain-computer interface worn by the user, acquire multiple sets of muscle signals through the surface electromyography patch on the user's muscle surface, and acquire the user's foot pressure signals through the flexible foot pressure insole worn by the user. Signal preprocessing module: used to preprocess the EEG signal, muscle signal and foot pressure signal respectively, and extract EEG feature vector, muscle feature vector and foot pressure feature vector respectively from the preprocessed EEG signal, muscle signal and foot pressure signal; Variance matrix acquisition module: used to acquire the EEG-EM covariance matrix, EEG-foot pressure covariance matrix, and EMG-foot pressure covariance matrix respectively based on the EEG feature vector, muscle feature vector, and foot pressure feature vector using the sliding window technique; Multimodal feature acquisition module: used to fuse the EEG-EMG covariance matrix, EEG-foot pressure covariance matrix and EMG-foot pressure covariance matrix to obtain the final covariance matrix, and to obtain real-time multimodal features through the final covariance matrix; Exoskeleton control signal generation module: used to input the real-time multimodal features into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angle; Intent recognition module: used to perform biofeedback visualization of the EEG feature vector, muscle feature vector and foot pressure feature vector. The user adjusts the force based on the biofeedback visualization, collects the EEG signal of the force adjustment, and inputs the EEG signal into a pre-trained personalized CNN model. The personalized CNN model recognizes the user's walking or stopping intention. Walking Assist Module: When the user's intention is to walk, the exoskeleton is used to assist the user in taking one step based on the joint angles and assistance level of the exoskeleton.
[0017] According to a third aspect of the present invention, a storage medium is provided, the storage medium storing a computer program, which, when executed by a host controller, implements the steps of the above-described method.
[0018] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This application acquires multimodal data and unifies the multimodal data (EEG signals, muscle signals, and foot pressure signals) into a dynamic control input through a covariance matrix fusion algorithm. It then combines this with a reinforcement learning model (PPO algorithm) to optimize the exoskeleton assistance strategy in real time. Users adjust their force exertion through biofeedback visualization of EEG feature vectors, muscle feature vectors, and foot pressure feature vectors. The EEG signals after force adjustment are used to identify the user's walking intention through a personalized CNN model, and the exoskeleton assistance strategy helps the user walk. This forms a closed loop where the patient actively adjusts their force exertion, enhancing the neural remodeling effect.
[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description
[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0021] Figure 1 This is a flowchart illustrating a home rehabilitation training method based on a lower limb exoskeleton, according to an exemplary embodiment. Figure 2 This is a schematic diagram of the dry electrode position according to another exemplary embodiment; Figure 3 This is a schematic diagram of a 16-point FSR (force-sensitive resistor) array distribution according to another exemplary embodiment; Figure 4 This is a schematic diagram of muscle recruitment thermodynamics according to another exemplary embodiment; Figure 5 This is a schematic diagram of a home rehabilitation training device based on a lower limb exoskeleton, according to another exemplary embodiment. In the attached diagram: 1-Data acquisition module, 2-Signal preprocessing module, 3-Variance matrix acquisition module, 4-Multimodal feature acquisition module, 5-Exoskeleton control signal generation module, 6-Intent recognition module, 7-Assisted walking module. Detailed Implementation
[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.
[0023] Example 1 Figure 1 This is a schematic flowchart illustrating a home-based rehabilitation training method based on a lower limb exoskeleton, according to an exemplary embodiment. Figure 1 As shown, the method includes: S1 acquires the user's brainwave signals through a non-invasive brain-computer interface worn by the user, acquires multiple muscle signals through surface electromyography patches on the user's muscle surface, and acquires the user's foot pressure signals through flexible foot pressure insoles worn by the user. S2, preprocess the EEG signal, muscle signal and foot pressure signal respectively, and extract the EEG feature vector, muscle feature vector and foot pressure feature vector respectively from the preprocessed EEG signal, muscle signal and foot pressure signal. S3, based on the EEG feature vector, muscle feature vector and foot pressure feature vector, the sliding window technique is used to obtain the EEG-EM covariance matrix, EEG-foot pressure covariance matrix and EEM-foot pressure covariance matrix respectively; S4, the EEG-EMG covariance matrix, EEG-foot pressure covariance matrix, and EMG-foot pressure covariance matrix are fused to obtain the final covariance matrix, and real-time multimodal features are obtained through the final covariance matrix; S5, input the real-time multimodal features into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angle; S6, the EEG feature vector, muscle feature vector and foot pressure feature vector are visualized using biofeedback. The user adjusts the force based on the biofeedback visualization, and the EEG signal of the force adjustment is collected. The EEG signal is input into a pre-trained personalized CNN model, and the personalized CNN model recognizes the user's intention to walk or stop. S7, when the user's intention is to walk, the exoskeleton assists the user to take one step based on the exoskeleton joint angle and the exoskeleton assistance level; It is understood that, in order to better understand this solution, this embodiment first describes the hardware devices involved in this method, specifically including: Wearable data collection terminal: This is a non-invasive 8-channel dry electrode EEG headgear. The electrode deployment conforms to the international 10-20 system standard. Electrodes include Cz, FCz, CPz, C1, C2, FC1, FC2, CP1, A1 / A2 (reference electrodes), plus Fpz (ground). The electrode distribution is shown in the attached figure. Figure 2 As shown; Signal processing flow: Signal amplification: Preamplifier (gain 1000x, bandwidth 0.5-100Hz); Analog-to-digital conversion: 24-bit ADC (512Hz sampling rate, quantization noise <1μV); Preprocessing: Removal of electrooculography artifacts by bandpass filtering (5-30Hz) and independent component analysis (ICA); The pre-processed EEG features are transmitted to the exoskeleton main control unit via Bluetooth; The exoskeleton master control unit performs feature extraction: extracts μ / β band energy values and uses the CSP algorithm to enhance the signal differences between target categories (such as leg raising / resting).
[0024] Control strategy: Equipped with a PID controller (1kHz frequency) to adjust joint angle error in real time; Device side (exoskeleton main control unit / EEG headband): The exoskeleton's main control unit is the primary computing unit at the Level 1 device level, responsible for integrating and processing all data from the EEG headband, sEMG patch, and foot pressure insole.
[0025] Receive pre-processed EEG signals from the EEG headband; Real-time processing of sEMG and foot pressure signals (RMS calculation, center of gravity trajectory analysis); Run basic intent recognition algorithms (such as walking / resting classification): The exoskeleton's emergency braking control is triggered when a sudden spasm or a severe shift in the center of gravity is detected. Dynamic offloading mechanism: When CPU load > 70%, the reinforcement learning inference task will be offloaded to the home gateway; EEG headband: integrates an EEG signal acquisition module (dry electrodes) and a pre-processing circuit (amplification, filtering, analog-to-digital conversion); Responsible for the initial acquisition and preprocessing of EEG signals (such as ICA artifact removal); The pre-processed EEG features are transmitted to the exoskeleton main control unit via Bluetooth; Level 2: Home Gateway (Smart Router): Multimodal data alignment (Bluetooth 5.0 timestamp synchronization); Run lightweight PPO model inference to generate exoskeleton control commands; Level 3: Mobile devices (smartphones / tablets) User interface (thought bar, heat map, etc.) Cloud platform: Model training: PyTorch distributed training, with personalized reinforcement learning models (PPO algorithm) updated weekly based on patient training data. Rehabilitation database: MySQL stores patients' historical data (EEG / sEMG / foot pressure / training records) and provides a WebSocket interface for hospitals to access the data.
[0026] Based on the aforementioned hardware facilities, the specific steps for implementing this method are as follows: The system acquires the user's electroencephalogram (EEG) signals through a head-mounted device, acquires multiple sets of muscle signals (sEMG) through sEMG patches on the surface of the user's muscles, and acquires the user's foot pressure signals through flexible plantar pressure insoles. Synchronize EEG, sEMG, and foot pressure signals via Bluetooth 5.0 broadcast timestamps; Signal preprocessing: EEG: Low-frequency noise (such as electromyography interference) and high-frequency noise (such as equipment noise) are filtered out by bandpass filtering (5-30Hz), and electrooculography artifacts are removed by independent component analysis (ICA). sEMG: A second-order Butterworth bandpass filter (20-500Hz) is used to eliminate baseline drift and high-frequency noise; Foot pressure: The FSR resistance value is read through the I2C interface, and after noise reduction by Kalman filtering, the resistance value is converted into a pressure value according to the calibration parameters; Signal feature extraction: EEG: Extract the energy values of the μ / β band, process the EEG signal through CSP spatial filtering, center the filtered EEG data (mean to zero) and standardize the covariance, and extract the first 8 pairs of spatial patterns (a total of 16 feature vectors) through generalized feature decomposition optimization filter to obtain the above-mentioned EEG feature vectors. sEMG: Calculate the root mean square (RMS) value of each muscle signal group, and map the RMS value to the [0,1] interval. The formula for calculating the root mean square value is as follows:
[0027] In the formula, X i This represents the amplitude of the sEMG (surface electromyography) signal at the nth sampling point (unit: usually μV). N This indicates the total number of sampling points within the current calculation window (e.g., the signal length over a period of time). Foot pressure signal: The 16 pressure points are standardized (mean zeroed, variance 1), and then dimensionality is reduced using PCA, retaining the first 3 principal components. The cumulative contribution rate must be >85%, i.e., the 16 pressure points are reduced to 3 dimensions using PCA. A 16-point FSR (force-sensitive resistor) array is evenly distributed on the insole, as shown in the attached image. Figure 3 As shown, the sampling rate is 100Hz; Covariance matrix fusion algorithm: Input feature alignment: Align the modal features according to the time window (EEG: 16-dimensional, sEMG: 5-dimensional, foot pressure signal: 3-dimensional). Sliding window calculation: Calculate the covariance matrices of EEG-EMG, EEG-foot pressure, and EMG-foot pressure using a 500ms window.
[0028] In the formula, x i and y i The mode at time n. X and Y Feature vector and For modality X and Y The mean vector within the sliding window, N The number of samples within the window. T This refers to the duration of the sliding window. Principal component extraction: for the covariance matrix Perform eigenvalue decomposition and take the top 3 largest eigenvalues as multimodal features; Pre-training the PPO model using historical data: The input for model training is: multimodal features obtained by fusing the covariance matrix from historical data (1D-CNN extracts temporal features). The output of model training is: exoskeleton assistance level, also known as exoskeleton assistance percentage (negative value in impedance mode, positive value in assistance mode) and exoskeleton joint angle; The model building function is designed as follows:
[0029] In the formula, α, β, and γ are dynamically adjusted according to the patient's recovery status (e.g., in the early stage, α=0.6, β=0.3, γ=0.1, and in the later stage, when muscle strength recovers, α=0.3, β=0.6, γ=0.1). MVC The maximum amplitude of the electromyographic signal recorded when the subject contracts the target muscle with maximum effort; Generate exoskeleton control strategies based on model output: PID control: Calculate the motor drive current based on the angle error (θ_target - θ_actual); 𝜃target: Desired joint angle (generated from preset gait); 𝜃actual: Current joint angle (measured by encoder); Early rehabilitation support mode, later resistance control mode; Joint torque calculation: ; in J For Jacobian matrices, Δθ Δω is the angular error, and Δω is the angular velocity error (the error is the difference between the angular velocity in the preset gait and the current angular velocity (obtained by differentiating the angle measured by the encoder)). and The gain coefficients for angle and angular velocity; Impedance control mode: In the formula, : Desired speed (generated from preset gait); The current speed can be calculated using the Jacobian matrix based on the angle, angular velocity, and link length. : End target position (generated by preset gait); The current position can be calculated using the forward kinematics formula by adding the angular velocity and the length of the link. B =15Ns / m, K =200N / m (can be adjusted according to the percentage of assist); Link length: The structural length between two adjacent joints; After the PPO model training is completed, the user's real-time multimodal features are input into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level; based on the exoskeleton assistance level, an exoskeleton control strategy is generated, such as the exoskeleton joint control angle.
[0030] The next step is to visualize the biofeedback of the EEG feature vector, muscle feature vector, and foot pressure feature vector, specifically including: EEG intensity "mind bar": ERD / ERS calculation: Using the resting state as a baseline, the μ / β energy change is calculated in real time, as shown in the following expression: ; In the formula, Ecurrent This indicates the μ / β band energy of the current EEG (based on EEG signals). Erest Indicates the resting baseline energy (the energy of the corresponding band when stopped); CSP Feature Mapping: Log-ratio of Walk / Rest Variance Linearly Scaled to 0-100 Range (The energy of the band corresponding to walking is less than the energy of the band corresponding to stopping). The CSP feature vector and ERD% value are input together into the pre-trained personalized CNN model, and the CNN model outputs a continuous value from 0 to 100. Dynamic display: The intensity of intention is displayed in real time on the mobile terminal in the form of a progress bar, with color gradient (blue → red corresponds to 0 → 100). Muscle recruitment heatmap: RMS calculation: in, N This represents the number of sampling points within the window. x i The denoised electromyographic signal value; Muscle activation intensity = RMS / MVC * 100%; Color is used to indicate the degree of muscle recruitment in human muscles, as shown in the attached image. Figure 4 As shown, red → blue corresponds to 0 → 100%; Plantar pressure distribution diagram: Convert the pressure value Pi at each FSR point into a heat map, with blue → red corresponding to 0 → 1; The "Mind Bar" displays EEG intensity, helping users better grasp walking / resting motion imagery; the muscle recruitment heat map can show that a patient's muscles are not working properly, and users can try to contract that muscle more forcefully; the plantar pressure distribution map shows gait symmetry, allowing users to identify gait problems during walking, and thus better control their gait, forming a self-regulating and optimized training process.
[0031] Training a personalized CNN includes: Users perform a 5-minute visual leg-raising task in sequence. During the training task, multiple sets of EEG signals are collected as fine-tuning datasets. The fine-tuning dataset includes the CSP feature vectors and ERD% values mentioned above. The input of the lightweight CNN model is the CSP feature vector + ERD%, and the label is the intent intensity, resting = 0, walking imagination = 100, and the output is a continuous value from 0 to 100. Freeze the convolutional layers of the lightweight CNN model, set the learning rate and number of iterations (learning rate 1e-4, 10 iterations) of the fully connected layers of the lightweight CNN model, and perform transfer learning on the lightweight CNN model through fine-tuning the dataset to obtain a personalized CNN model; By setting a threshold for the "thought bar," the exoskeleton is able to recognize the user's intention to walk at a threshold of 80%. The exoskeleton then sends a "walk" command to the user and assists the user in walking based on the aforementioned exoskeleton assistance level and control strategy. Each time the exoskeleton receives a "walk" command, it assists the user in taking one step.
[0032] Visualize recovery progress: Generate daily / weekly reports.
[0033] It is understood that this embodiment also discloses a protection mechanism, including: if a sudden muscle spasm is detected (sEMG RMS > 300 μV and lasts for 200 ms) or a severe shift in the center of gravity is detected (Δ... X When the force is greater than 25% and lasts for 300ms, the exoskeleton emergency braking control is triggered, and it automatically rolls back to the previous stable state (assistance percentage, joint angle, etc.) to avoid secondary injury. Wherein, offset Δ X The calculation method is as follows: Calculate the center of gravity: The center of gravity along the X-axis (left-right direction) is calculated for the left and right feet respectively. and : In the formula, x i Indicates the first i The coordinate position of a pressure sensor in the X direction; Indicates the first i Normalized pressure values of each sensor ( ); Dynamic trajectory: The center of gravity movement path is generated through time-series data at a frequency of 100Hz; Symmetry assessment: The midpoint of the X-coordinate of the center of gravity of the left and right feet is taken as the "ideal reference center". ; Overall plantar pressure center (i.e., the overall CoP of both feet) ; Offset is ; Mild asymmetry: ΔX > 15% of the width of the flexible foot pressure insole; Severe asymmetry: ΔX(t) > 25% of the width of the flexible foot pressure insole; If the duration of severe asymmetry is ≥ 300ms, it is considered a severe offset.
[0034] Example 2 Figure 5 This is a schematic diagram of a home rehabilitation training device based on a lower limb exoskeleton, according to another exemplary embodiment, the device comprising: Data acquisition module 1: used to acquire the user's brainwave signals through the non-invasive brain-computer interface worn by the user, acquire multiple sets of muscle signals through the surface electromyography patch on the surface of the user's muscles, and acquire the user's foot pressure signals through the flexible foot pressure insole worn by the user. Signal preprocessing module 2: is used to preprocess the EEG signal, muscle signal and foot pressure signal respectively, and extract EEG feature vector, muscle feature vector and foot pressure feature vector respectively from the preprocessed EEG signal, muscle signal and foot pressure signal. Variance matrix acquisition module 3: used to acquire the EEG-EMG covariance matrix, EEG-foot pressure covariance matrix, and EMG-foot pressure covariance matrix respectively based on the EEG feature vector, muscle feature vector, and foot pressure feature vector using the sliding window technique; Multimodal feature acquisition module 4: used to fuse the EEG-EMG covariance matrix, EEG-foot pressure covariance matrix and EMG-foot pressure covariance matrix to obtain the final covariance matrix, and to obtain real-time multimodal features through the final covariance matrix; Exoskeleton control signal generation module 5: used to input the real-time multimodal features into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angle; Intent recognition module 6: used to perform biofeedback visualization of the EEG feature vector, muscle feature vector and foot pressure feature vector. The user adjusts the force based on the biofeedback visualization, collects the EEG signal of the force adjustment, and inputs the EEG signal into the pre-trained personalized CNN model. The personalized CNN model recognizes the user's walking or stopping intention. Walking Assist Module 7: When the user's intention is to walk, the exoskeleton is used to assist the user in taking one step based on the joint angles and assist level of the exoskeleton.
[0035] Example 3: This embodiment provides a storage medium storing a computer program, which, when executed by a host controller, implements the various steps in the above method. It is understood that the storage medium mentioned above can be a read-only memory, a hard disk, or an optical disk, etc.
[0036] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.
[0037] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.
[0038] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention 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 will be understood by those skilled in the art to which embodiments of the invention pertain.
[0039] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in 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 one or a combination of the following techniques known in the art: 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 (PGAs), field-programmable gate arrays (FPGAs), etc.
[0040] Those skilled in the art will understand that all or part of the steps of the methods in 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.
[0041] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0042] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.
[0043] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. 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.
[0044] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A home-based rehabilitation training method based on a lower limb exoskeleton, characterized in that, The method includes: The system acquires the user's brainwave signals through a non-invasive brain-computer interface worn by the user, acquires multiple muscle signals through surface electromyography patches on the user's muscles, and acquires the user's foot pressure signals through flexible foot pressure insoles worn by the user. The electroencephalogram (EEG) signal, muscle signal, and foot pressure signal are preprocessed respectively, and the EEG feature vector, muscle feature vector, and foot pressure feature vector are extracted from the preprocessed EEG signal, muscle signal, and foot pressure signal respectively. Based on the aforementioned EEG feature vector, muscle feature vector, and foot pressure feature vector, the sliding window technique is used to obtain the EEG-EM covariance matrix, EEG-foot pressure covariance matrix, and EEM-foot pressure covariance matrix, respectively. The EEG-EM Covariance Matrix, EEG-Foot Pressure Covariance Matrix, and EMG-Foot Pressure Covariance Matrix are fused to obtain the final covariance matrix, and real-time multimodal features are obtained through the final covariance matrix. The real-time multimodal features are input into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angles. The EEG feature vector, muscle feature vector, and foot pressure feature vector are visualized using biofeedback. The user adjusts their force based on the biofeedback visualization, and the EEG signals of the force adjustment are collected. The EEG signals are then input into a pre-trained personalized CNN model, which identifies the user's intention to walk or stop. When the user's intention is to walk, the exoskeleton assists the user in taking one step based on the exoskeleton joint angles and the exoskeleton assistance level.
2. The method according to claim 1, characterized in that, The preprocessing of the electroencephalogram (EEG) signals, muscle signals, and foot pressure signals includes: The EEG signal was subjected to bandpass filtering to remove low-frequency and high-frequency noise, and the filtered EEG signal was subjected to independent component analysis to remove electrooculogram artifacts. The muscle signal was filtered using a second-order Butterworth bandpass filter to eliminate baseline drift and high-frequency noise; The foot pressure signal is read from the FSR resistance value via the I2C interface. After noise reduction by Kalman filtering, the resistance value is converted into a pressure value according to the calibration parameters.
3. The method according to claim 2, characterized in that, The extraction of EEG feature vectors, muscle feature vectors, and foot pressure feature vectors from preprocessed EEG signals, muscle signals, and foot pressure signals includes: The μ / β band energy values were extracted from the preprocessed EEG signals and processed by CSP spatial filtering. The filtered EEG signals were then centered and covariance standardized. The first 8 pairs of spatial patterns were extracted by generalized feature decomposition optimization filter to obtain a 16-dimensional EEG feature vector. Obtain the root mean square value of each group of muscle signals after preprocessing, and map the root mean square value of each group of muscle signals to the interval [0,1] to obtain a 5-dimensional muscle feature vector. The preprocessed pressure values are standardized, and the first three principal components are retained by PCA dimensionality reduction, thereby reducing the 16 pressure values to 3 dimensions and obtaining a 3-dimensional foot pressure feature vector.
4. The method according to claim 3, characterized in that, The training of the PPO model includes: Acquire and preprocess the user's historical EEG signals, historical muscle signals, and historical foot pressure signals; Historical EEG feature vectors, historical muscle feature vectors, and historical foot pressure feature vectors are obtained based on the preprocessed historical EEG signals, historical muscle signals, and historical foot pressure signals. Based on the historical EEG feature vector, historical muscle feature vector, and historical foot pressure feature vector, respectively establish the historical EEG-EMG covariance matrix, the historical EEG-foot pressure covariance matrix, and the historical EEG-foot pressure covariance matrix. The historical EEG-EMG covariance matrix, historical EEG-foot pressure covariance matrix, and historical EMG-foot pressure covariance matrix are fused to obtain the historical final covariance matrix, and historical multimodal features are obtained through the historical final covariance matrix. The pre-built PPO model is iteratively trained using historical multimodal features based on a preset reward function to obtain a trained PPO model.
5. The method according to claim 4, characterized in that, The training of the personalized CNN model includes: A pre-built lightweight CNN model is trained using a publicly available EEG database, and then the lightweight CNN model is adaptively calibrated based on a preset training task to obtain a personalized CNN model.
6. The method according to claim 5, characterized in that, The step of adaptively calibrating the lightweight CNN model based on a preset training task to obtain a personalized CNN model includes: Users perform a 5-minute visual leg-raising task in sequence. During the training task, multiple sets of EEG signals are collected as fine-tuning datasets. Freeze the convolutional layers of the lightweight CNN model, set the learning rate and number of iterations of the fully connected layers of the lightweight CNN model, and perform transfer learning on the lightweight CNN model by fine-tuning the dataset to obtain a personalized CNN model.
7. The method according to claim 6, characterized in that, Also includes: When a sudden muscle spasm or severe shift in the center of gravity is detected, the exoskeleton is braked urgently. The detection of sudden muscle spasm includes: when the root mean square value of any set of muscle signals is greater than a preset threshold and the duration exceeds a preset time threshold, it is determined to be a sudden muscle spasm. The detection of severe center of gravity shift includes: calculating the foot center of gravity shift based on the foot pressure signal; if the foot center of gravity shift exceeds a preset threshold and its duration exceeds a preset time threshold, it is determined to be a severe center of gravity shift.
8. The method according to claim 7, characterized in that, The step of calculating the foot center of gravity offset based on the foot pressure signal includes: The flexible foot pressure insole uses a 16-point FSR array; Obtain the coordinates of each FSR point of the left and right feet in the X direction during the user's movement, and obtain the normalized pressure value of each FSR point of the left and right feet; Calculate the centroid of the left and right feet and the overall plantar pressure center based on the coordinates of each FSR point in the X direction and the normalized pressure value of each FSR point. Calculate the reference center based on the X-axis center of gravity of the left and right feet; The offset of the plantar center of gravity is calculated based on the reference center and the overall plantar pressure center.
9. A home-use rehabilitation training device based on a lower limb exoskeleton, characterized in that, The device includes: Data acquisition module: used to acquire the user's brainwave signals through the non-invasive brain-computer interface worn by the user, acquire multiple sets of muscle signals through the surface electromyography patch on the user's muscle surface, and acquire the user's foot pressure signals through the flexible foot pressure insole worn by the user. Signal preprocessing module: used to preprocess the EEG signal, muscle signal and foot pressure signal respectively, and extract EEG feature vector, muscle feature vector and foot pressure feature vector respectively from the preprocessed EEG signal, muscle signal and foot pressure signal; Variance matrix acquisition module: used to acquire the EEG-EM covariance matrix, EEG-foot pressure covariance matrix, and EMG-foot pressure covariance matrix respectively based on the EEG feature vector, muscle feature vector, and foot pressure feature vector using the sliding window technique; Multimodal feature acquisition module: used to fuse the EEG-EMG covariance matrix, EEG-foot pressure covariance matrix and EMG-foot pressure covariance matrix to obtain the final covariance matrix, and to obtain real-time multimodal features through the final covariance matrix; Exoskeleton control signal generation module: used to input the real-time multimodal features into the pre-trained PPO model, and the trained PPO model outputs the exoskeleton assistance level and exoskeleton joint angle; Intent recognition module: used to perform biofeedback visualization of the EEG feature vector, muscle feature vector and foot pressure feature vector. The user adjusts the force based on the biofeedback visualization, collects the EEG signal of the force adjustment, and inputs the EEG signal into a pre-trained personalized CNN model. The personalized CNN model recognizes the user's walking or stopping intention. Walking Assist Module: When the user's intention is to walk, the exoskeleton is used to assist the user in taking one step based on the joint angles and assistance level of the exoskeleton.
10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by the main controller, implements the various steps of the home rehabilitation training method based on a lower limb exoskeleton as described in any one of claims 1-8.