Hand rehabilitation equipment and system based on multi-source data analysis

Abstract

The invention relates to the technical field of rehabilitation medical equipment, in particular to hand rehabilitation equipment and system based on multi-source data analys.The hand rehabilitation equipment comprises a wearable glove, the wearable glove is made of a flexible and breathable silica gel-fabric composite material and comprises finger parts and a palm part, and the finger parts correspond to five fingers of a human body one to one; each finger part is divided into three sectional mounting positions including a near section, a middle section and a far section in the length direction, and a polytetrafluoroethylene wear-resistant sleeve is pre-embedded in the wearable glove to serve as a traction rope channel; the flexible array type myoelectricity sensors are fixed to the mounting positions of the proximal section, the middle section and the distal section of each finger part, myoelectricity, postures, tension and interaction signals are comprehensively collected, weighted fusion is conducted through an attention mechanism, the motion intention recognition accuracy is larger than or equal to 92%, functional state evaluation is more accurate, and the motion intention recognition accuracy is larger than or equal to 92%. The training mode and parameters are dynamically adjusted according to the real-time state of the user, excessive / insufficient training is avoided, and the rehabilitation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation medical equipment technology, specifically to a hand rehabilitation device and system based on multi-source data analysis. Background Technology

[0002] Hand motor dysfunction is a common sequela of neurological diseases and trauma, severely impacting patients' daily living abilities. Existing hand rehabilitation equipment suffers from the following core deficiencies:

[0003] The data collection dimensions are limited: passive traction devices only collect motor motion parameters, and active electromyography triggering devices only collect electromyography signals from a single part, which cannot comprehensively reflect the hand's movement status and muscle activity.

[0004] Fixed rehabilitation programs: Training modes and parameters are mostly preset and cannot be dynamically adjusted according to the user's real-time exercise intentions and muscle fatigue levels, resulting in insufficient precision and personalization.

[0005] Lack of closed-loop control: Only open-loop traction drive is implemented, and parameters cannot be adjusted according to joint posture and traction tension feedback, which can easily lead to overtraining or muscle spasm. To address this, a hand rehabilitation device and system based on multi-source data analysis is proposed. Summary of the Invention

[0006] In view of this, the present invention provides a hand rehabilitation device and system based on multi-source data analysis to solve or alleviate the technical problems existing in the prior art, and at least provides a beneficial option.

[0007] The technical solution of this invention is implemented as follows: A hand rehabilitation device based on multi-source data analysis includes: a wearable glove, the wearable glove being made of a flexible and breathable silicone-fabric composite material, including finger parts and palm parts corresponding to the five fingers of the human body, each finger part being divided into three segmented installation positions along its length: proximal phalanx, middle phalanx, and distal phalanx, and a polytetrafluoroethylene wear-resistant sleeve pre-embedded inside the wearable glove as a traction rope channel; and a flexible array-type electromyography (EMG) sensor, with the EMG sensor fixed at each of the proximal, middle, and distal phalanx installation positions of each finger part. The sampling frequency is ≥1000Hz, used to collect raw electromyographic signals from different parts of the fingers, including amplitude, frequency, and time-frequency domain characteristics; the drive component includes five miniature servo motors, five traction wheels, and five sets of traction ropes. Each of the five motors corresponds to one of the five fingers and is fixed to the bottom of the palm of the glove. The output end of each motor is coaxially fixed to the traction wheel, and the motor integrates an encoder for collecting speed and angle signals; each set of traction ropes is made of ultra-high molecular weight polyethylene, with one end wound around and fixed to the corresponding traction wheel, and the other end threaded through the glove. The inner PTFE wear-resistant sleeve is fixedly connected to the end of the corresponding finger; a flexible control screen, which is a flexible OLED screen, is fitted to the top of the back of the hand wearing the glove, integrating a touch interaction unit, a display unit, and a physiological signal acquisition unit; the touch interaction unit supports multi-touch, the display unit is used to display rehabilitation training data, and the physiological signal acquisition unit includes silver electrode pads distributed around the edge of the flexible control screen to collect the user's skin contact impedance signal; auxiliary sensing components include a six-axis posture sensor and a miniature tension sensor; the posture sensor is installed at the joint of each finger to collect the angle, angular velocity, and angular acceleration signals of the finger joint; the tension sensor is connected in series in the middle of each traction rope, with a range of 0-10N and an accuracy of ≤0.1N, to collect the real-time tension signal of the traction rope; the device-side control unit is electrically connected to the electromyography sensor, motor, flexible control screen, posture sensor, and tension sensor, and has a built-in Bluetooth / wireless communication module to receive signals collected by each sensor, drive the motor to rotate, and interact with external systems.

[0008] More preferably, the traction wheel is made of nylon, with a diameter of 8-12mm, and has anti-slip grooves on its surface. The traction rope is wound around the traction wheel and fixed in the grooves by a buckle.

[0009] More preferably, the device-side control unit has a built-in overcurrent and overvoltage protection circuit. When the motor operating current exceeds a preset threshold of ≥500mA, the power supply to the motor is automatically cut off and an alarm is triggered.

[0010] A hand rehabilitation system based on multi-source data analysis, the system includes a local terminal module (tablet computer / industrial computer) and a cloud server module, specifically including the following functional modules: (1) Data acquisition module The data acquisition module is communicatively connected to the device-side control unit, used to receive and summarize multi-source acquired data in real time, and store it according to timestamp classification. The multi-source acquired data includes:

[0011] Electromyography (EMG) signal data: raw EMG signals from the proximal / middle / distal phalanges of the fingers, as well as time-domain characteristics (root mean square, integrated EMG value, number of zero crossings) and frequency-domain characteristics (centroid of the spectrum, average power frequency) after preliminary filtering.

[0012] Motion posture data: finger joint angle (0-90°), angular velocity (0-10° / ms), angular acceleration (0-5° / ms²), and speed (0-60rpm) and rotation angle (0-360°) signals collected by the motor encoder;

[0013] Tension feedback data: Real-time tension value of the traction rope, sampling frequency ≥200Hz;

[0014] Interactive sensing data: touch position, touch pressure (0-5N), touch duration, and skin contact impedance (100-1000Ω) signal of the flexible control screen;

[0015] (2) Data preprocessing module

[0016] The data preprocessing module is communicatively connected to the data acquisition module and is used to perform noise reduction, calibration, feature extraction, and normalization processing on multi-source data, specifically including:

[0017] Electromyography signal preprocessing: 50Hz notch filtering is used to eliminate power frequency interference, wavelet denoising (db4 wavelet, 5-level decomposition) is used to remove baseline drift, and linear interpolation is used to complete missing data. Finally, standardized electromyography feature vectors are output.

[0018] Motion posture data preprocessing: Kalman filtering is used to eliminate measurement noise from the posture sensor, outlier removal (3σ criterion) is performed on the motor encoder data, and smoothed joint motion parameters are output.

[0019] Tension feedback data preprocessing: A 50ms sliding window is used for smoothing to remove abnormal peaks >8N and output a continuous tension curve;

[0020] Interactive perception data preprocessing: Physically calibrate the touch coordinates, normalize the skin contact impedance signal to the 0-1 range, and extract touch intent features (such as single-point long press corresponding to "pause training" and multi-point swipe corresponding to "adjust mode").

[0021] (3) Multi-source data fusion analysis module

[0022] The multi-source data fusion analysis module is communicatively connected to the data preprocessing module, and adopts a two-level fusion algorithm of "feature layer + decision layer", specifically including:

[0023] Feature layer fusion: The preprocessed electromyographic features (12-dimensional), posture features (9-dimensional), tension features (1-dimensional), and interaction features (5-dimensional) are mapped to the same high-dimensional feature space, and an attention weight matrix is ​​constructed (the weight values ​​are dynamically adjusted according to the user's historical rehabilitation data, with electromyographic feature weights of 0.4, posture features of 0.3, tension features of 0.15, and interaction features of 0.15), and the weighted fusion of the 27-dimensional global feature vector is output.

[0024] Decision layer fusion: The global feature vector is input into a pre-trained CNN-LSTM hybrid model, which includes 2 convolutional layers (extracting spatial features), 2 LSTM layers (extracting temporal features), and 1 fully connected layer. The output is:

[0025] Motor intention recognition results: active flexion, active extension, passive rehabilitation, cessation of training, mode switching (accuracy ≥ 92%).

[0026] Hand function status assessment results: muscle strength grade (1-5), joint range of motion grade (1-3), degree of muscle spasm (0-2);

[0027] (4) Adaptive Rehabilitation Program Generation Module

[0028] The adaptive rehabilitation program generation module is communicatively connected to the multi-source data fusion and analysis module, and is used to generate a real-time adjustable rehabilitation training program based on movement intention and functional status, specifically including:

[0029] Training mode selection: Active training mode (assisted movement driven by traction components based on user electromyographic signals), passive training mode (driving finger movement according to a preset trajectory), resistance training mode (dynamically adjusting traction tension), relaxation mode (low speed, low tension traction);

[0030] Training parameter dynamic adjustment rules:

[0031] Muscle strength grade ≤ 2: motor speed ≤ 5° / s, rotation range is 50% of joint range of motion, tension threshold ≤ 1N;

[0032] Muscle strength grade 3-4: motor speed 8-12° / s, rotation range of 80%-100% of joint range of motion, tension threshold 1-3N;

[0033] Muscle strength level ≥ 5: Motor speed 10-15° / s, start resistance training, tension threshold 3-5N;

[0034] Muscle spasm severity ≥ Grade 2: Immediately switch to relaxation mode, reduce motor speed to 2° / s, and tension threshold ≤ 0.5N;

[0035] Fatigue characteristics of electromyography signals (integrated electromyography value decrease rate ≥20%): Automatically pause training and trigger rest prompt;

[0036] (5) Closed-loop drive control module

[0037] The closed-loop drive control module is communicatively connected to the adaptive rehabilitation plan generation module and the equipment-side control unit, respectively, and uses a PID control algorithm to achieve precise traction control, specifically including:

[0038] Command issuance: Convert the motor control parameters (target speed, target angle, tension threshold) in the rehabilitation plan into PWM drive commands and send them to the device control unit;

[0039] Real-time feedback adjustment: The actual values ​​of joint angle and traction tension transmitted back from the receiving device are compared with the target values. If the angle deviation is ≥3° or the tension deviation is ≥0.5N, the motor PWM duty cycle is adjusted through the PID algorithm (Kp=0.8, Ki=0.2, Kd=0.1).

[0040] Safety protection control: When the tension value is ≥5N, the joint angle exceeds the safe range (>90°) or muscle spasm is detected, the motor drive signal is immediately cut off, triggering the audible and visual alarm and voice prompt of the flexible control screen;

[0041] (6) Human-computer interaction module

[0042] The human-computer interaction module is communicatively connected to the flexible control screen and the multi-source data fusion analysis module to achieve multi-dimensional interaction, specifically including:

[0043] Visualization: The flexible control screen displays the training mode, current joint angle, muscle strength level, training duration, and fatigue level in real time, and simultaneously displays the animation of the finger movement target trajectory;

[0044] Touch interaction: Supports users to manually switch training modes, adjust motor speed / tension threshold, and trigger emergency pause by pressing and holding the touch area for ≥3 seconds;

[0045] Voice interaction: It integrates a speech recognition unit (recognition rate ≥90%) and a speech synthesis unit, supports voice commands (such as "switch to passive training" and "speed up"), and provides voice feedback on training progress and evaluation results;

[0046] (7) Data storage and feedback module

[0047] The data storage and feedback module is communicatively connected to the multi-source data fusion and analysis module and the adaptive rehabilitation plan generation module, respectively, for full-cycle rehabilitation management, specifically including:

[0048] Data storage: Store user data collected from multiple sources, fusion analysis results, rehabilitation plans and execution data by timestamp to form personalized rehabilitation records, with a storage period of ≥12 months;

[0049] Rehabilitation effect assessment: Based on historical data, the improvement in muscle strength and joint range of motion is calculated through a trend analysis model, and weekly / monthly rehabilitation reports are generated;

[0050] Remote monitoring: The cloud server module communicates with the rehabilitation physician's terminal, allowing the physician to remotely view rehabilitation data, adjust training program parameters, and simultaneously push reports to the user's mobile APP.

[0051] More preferably, the CNN-LSTM hybrid model is trained using transfer learning, first pre-trained on a publicly available hand rehabilitation dataset, and then fine-tuned online using user-personalized data, with a model update cycle of ≤7 days.

[0052] Preferably, the data storage and feedback module supports encrypted data transmission and uses the AES-256 encryption algorithm to protect user privacy data.

[0053] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions:

[0054] I. This invention comprehensively collects electromyography, posture, tension, and interaction signals, and uses attention mechanisms to weightedly fuse them, achieving a motor intention recognition accuracy of ≥92%, more accurate functional status assessment, and dynamically adjusting training modes and parameters according to the user's real-time status to avoid overtraining / undertraining and improve rehabilitation efficiency.

[0055] Second, this invention adjusts the motor drive based on real-time feedback of joint angle and traction tension, with an angle deviation of ≤±2° and a tension deviation of ≤±0.5N, reducing the risk of muscle spasms. The flexible screen that fits the back of the hand integrates multiple interactive functions and, combined with voice control, is convenient to operate and comfortable to wear.

[0056] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a structural diagram of the present invention;

[0059] Figure 2 This is a structural diagram from another perspective of the present invention;

[0060] Figure 3 This is a system flowchart of the present invention;

[0061] Figure 4 This is a flowchart of the fusion algorithm of the present invention.

[0062] Reference numerals: 1. Wearing gloves; 2. Fingers; 3. Electromyography sensor; 4. Miniature servo motor; 5. Traction wheel; 6. Traction rope; 7. Flexible control panel. Detailed Implementation

[0063] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0064] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0065] like Figure 1-4As shown, this embodiment of the invention provides a hand rehabilitation device based on multi-source data analysis, comprising: a wearable glove 1, which is made of a flexible and breathable silicone-fabric composite material, including finger parts 2 corresponding to the five fingers of the human body and a palm part, each finger part 2 being divided into three segmented installation positions along its length: proximal phalanx, middle phalanx, and distal phalanx, and a polytetrafluoroethylene wear-resistant sleeve pre-embedded inside the wearable glove 1 as a traction rope channel; and a flexible array-type electromyography (EMG) sensor 3, with an EMG sensor 3 fixed at each of the proximal, middle, and distal phalanx installation positions of each finger part 2, the EMG sensor 3 sampling frequency... The frequency is ≥1000Hz, used to collect raw electromyographic signals from different parts of the fingers, including amplitude, frequency, and time-frequency domain characteristics; the drive component includes five miniature servo motors 4, five traction wheels 5, and five sets of traction ropes 6. The five motors 4 correspond one-to-one with the five fingers of the human body and are fixed to the bottom of the palm of the glove 1. The output end of the motor 4 is coaxially fixed to the traction wheel 5, and the motor 4 integrates an encoder for collecting speed and angle signals; each set of traction ropes 6 is made of ultra-high molecular weight polyethylene, one end of which is wound around and fixed to the corresponding traction wheel 5, and the other end is threaded through the glove. The PTFE wear-resistant sleeve inside the glove 1 is fixedly connected to the end of the corresponding finger part 2; the flexible control screen 7, which is a flexible OLED screen, is attached to the top of the back of the hand wearing glove 1, integrating a touch interaction unit, a display unit, and a physiological signal acquisition unit; the touch interaction unit supports multi-touch, the display unit is used to display rehabilitation training data, and the physiological signal acquisition unit includes silver electrode pads distributed on the edge of the flexible control screen 7 for collecting the user's skin contact impedance signal; the auxiliary sensing components include a six-axis posture sensor and a miniature tension sensor; the posture sensor is installed at the joint of each finger part 2 for collecting the angle, angular velocity, and angular acceleration signals of the finger joint; the tension sensor is connected in series in the middle of each traction rope 6, with a range of 0-10N and an accuracy of ≤0.1N, for collecting the real-time tension signal of the traction rope 6; the device-side control unit is electrically connected to the electromyography sensor 3, the motor 4, the flexible control screen 7, the posture sensor, and the tension sensor, and has a built-in Bluetooth / wireless communication module for receiving signals collected by each sensor, driving the motor 4 to rotate, and interacting with the external system.

[0066] In a further preferred embodiment, the traction wheel 5 is made of nylon, with a diameter of 8-12mm, and has anti-slip grooves on its surface. The traction rope 6 is wound around the traction wheel 5 and fixed in the grooves of the traction wheel 5 by a buckle.

[0067] In a further preferred embodiment, the device-side control unit has a built-in overcurrent and overvoltage protection circuit. When the operating current of motor 4 exceeds a preset threshold of ≥500mA, the power supply to the motor is automatically cut off and an alarm is triggered.

[0068] A hand rehabilitation system based on multi-source data analysis, the system includes a local terminal module (tablet computer / industrial computer) and a cloud server module, specifically including the following functional modules: (1) Data acquisition module The data acquisition module is connected to the device control unit for real-time reception and aggregation of multi-source acquired data, which is stored according to timestamp classification. The multi-source acquired data includes:

[0069] Electromyography (EMG) signal data: raw EMG signals from the proximal / middle / distal phalanges of the fingers, as well as time-domain characteristics (root mean square, integrated EMG value, number of zero crossings) and frequency-domain characteristics (centroid of the spectrum, average power frequency) after preliminary filtering.

[0070] Motion posture data: finger joint angle (0-90°), angular velocity (0-10° / ms), angular acceleration (0-5° / ms²), and speed (0-60rpm) and rotation angle (0-360°) signals collected by the motor encoder;

[0071] Tension feedback data: Real-time tension value of the traction rope, sampling frequency ≥200Hz;

[0072] Interactive sensing data: touch position, touch pressure (0-5N), touch duration, and skin contact impedance (100-1000Ω) signal of the flexible control screen;

[0073] (2) Data preprocessing module

[0074] The data preprocessing module communicates with the data acquisition module and is used to perform noise reduction, calibration, feature extraction, and normalization on multi-source data to ensure data quality. Specifically, this includes:

[0075] Electromyography signal preprocessing: 50Hz notch filtering is used to eliminate power frequency interference, wavelet denoising is achieved through 5-level decomposition of db4 wavelets (high frequency noise removal threshold is set to 0.02mV), baseline drift is removed, and missing data is filled by linear interpolation (interpolation error ≤5%), and finally a standardized 12-dimensional electromyography feature vector is output.

[0076] Motion posture data preprocessing: Kalman filtering (process noise covariance Q=diag([1e-4,1e-4,1e-4]), observation noise covariance R=diag([1e-3,1e-3,1e-3])) is used to eliminate measurement noise from the posture sensor. Outliers are removed from the motor encoder data using the 3σ criterion (σ is the standard deviation of the data, and data that deviates from the mean by more than 3σ are removed). Smoothed 9-dimensional joint motion parameters are output.

[0077] Tension feedback data preprocessing: A 50ms sliding window is used for smoothing to remove abnormal peaks >8N (this threshold is determined based on the human hand muscle safety force test, the maximum safe tolerance tension of human hand muscles is 8N), and a continuous 1D tension curve is output.

[0078] Interactive perception data preprocessing: Physically calibrate the touch coordinates (calibration error ≤ 0.5mm), normalize the skin contact impedance signal to the 0-1 range, and extract 5-dimensional touch intent features (such as single-point long press ≥ 3s corresponding to "pause training", multi-point swipe corresponding to "adjust mode");

[0079] This module improves the signal-to-noise ratio by an average of 35% through multi-step data processing, providing high-quality data support for subsequent fusion analysis, which is superior to the existing equipment that only uses single filtering processing.

[0080] (3) Multi-source data fusion analysis module

[0081] The multi-source data fusion analysis module fuses four types of signals—electromyography, posture, tension, and interaction—through a two-level process of "feature layer + decision layer," and introduces a dynamic weight adjustment mechanism. This significantly improves recognition accuracy and personalization. Its non-obviousness lies in breaking through the limitations of existing technologies that rely on "single signal dominance" or "fixed weight fusion." It dynamically adjusts the weights of each signal according to the user's rehabilitation stage, making the fusion results more closely match the user's real-time state. Specifically, this includes:

[0082] Feature layer fusion: Preprocessed electromyographic features (12-dimensional), posture features (9-dimensional), tension features (1-dimensional), and interaction features (5-dimensional) are mapped to the same high-dimensional feature space (32-dimensional) to construct an attention weight matrix. The weight values ​​are dynamically adjusted based on the user's historical rehabilitation data. The initial weight allocation is based on: electromyographic feature weight 0.4 (EMG signals directly reflect muscle activity and are the core representation of motor intention; tests show that the correlation coefficient between EMG signals and motor intention is ≥0.85), posture features 0.3 (joint movement state is a direct manifestation of intention execution, with a correlation coefficient ≥0.7), and tension features 0.4 (tension). Features are weighted at 0.15 (tension feedback reflects training load adaptability, correlation coefficient ≥ 0.5) and 0.15 (user-initiated interaction commands have the highest priority, correlation coefficient ≥ 0.9). For example, when the user's muscle strength is weak (≤ level 2), the electromyography signal-to-noise ratio decreases (≤ 0.3), and the system automatically increases the posture feature weight to 0.35 and adjusts the electromyography feature weight to 0.35 to ensure the reliability of the fusion result. When the user triggers touch interaction, the interaction feature weight is temporarily increased to 0.5 to prioritize responding to user-initiated commands. Finally, a weighted fusion of 27-dimensional global feature vectors is output.

[0083] Decision layer fusion: The global feature vector is input into a pre-trained CNN-LSTM hybrid model. The model is trained using transfer learning, first pre-trained on a public hand rehabilitation dataset (such as the NinaProDB5 dataset, containing 10 classes of hand movement intentions, with a sample size ≥100,000), and then fine-tuned online using personalized user data (model update cycle ≤7 days) to ensure the model's adaptability to different users. The model structure includes two convolutional layers (3×3 kernel size, stride 1, padding method SAME, outputting 32-dimensional and 64-dimensional feature maps respectively for extracting spatial features), two LSTM layers (128 and 64 hidden units respectively, used for extracting temporal features), and one fully connected layer (output layer activation function is Softmax). The final output is:

[0084] Motor intention recognition results: active flexion, active extension, passive rehabilitation, cessation of training, and mode switching. Laboratory tests on 20 subjects (including 10 patients with post-stroke hand dysfunction and 10 healthy subjects) showed an average recognition accuracy of 93.2% (≥95% accuracy for healthy subjects and ≥90% accuracy for the patient group). Compared to existing single electromyography signal recognition devices (accuracy ≤80%), the recognition accuracy is improved by 16.5%.

[0085] Hand function status assessment results: muscle strength grade (1-5, referring to the MMT muscle strength assessment standard), joint range of motion grade (1-3, grade 1 ≤30°, grade 2 30-60°, grade 3 ≥60°), and muscle spasticity (0-2, referring to the Ashworth spasticity assessment standard).

[0086] (4) Adaptive Rehabilitation Program Generation Module

[0087] The adaptive rehabilitation program generation module communicates with the multi-source data fusion and analysis module to generate real-time adjustable rehabilitation training programs based on movement intentions and functional states, specifically including:

[0088] Training mode selection:

[0089] Active training mode: The traction component is driven by the user's electromyography (EMG) signal to assist movement. When the amplitude of the EMG signal reaches the preset threshold (0.2-0.5mV, dynamically adjusted according to the user's muscle strength), the motor provides auxiliary traction force.

[0090] Passive training mode: Drives finger movement along a preset trajectory (based on the normal finger movement angle curve), suitable for patients with muscle strength ≤2;

[0091] Resistance training mode: dynamically adjusts traction tension (gradually increases resistance according to muscle strength level), suitable for patients with muscle strength ≥5, to enhance muscle strength;

[0092] Relaxation mode: low speed, low tension traction (motor speed ≤2° / s, tension threshold ≤0.5N), used to relieve muscle spasms or relax after training;

[0093] Dynamic adjustment rules for training parameters (parameter range determined based on human hand biomechanical testing to avoid joint injury; test subjects are healthy individuals aged 18-60 and rehabilitation patients; maximum safe joint movement speed ≤15° / s):

[0094] Muscle strength grade ≤ 2: motor speed ≤ 5° / s, rotation range is 50% of joint range of motion, tension threshold ≤ 1N;

[0095] Muscle strength grade 3-4: motor speed 8-12° / s, rotation range of 80%-100% of joint range of motion, tension threshold 1-3N;

[0096] Muscle strength level ≥ 5: Motor speed 10-15° / s, start resistance training, tension threshold 3-5N;

[0097] Muscle spasm severity ≥ Grade 2: Immediately switch to relaxation mode, reduce motor speed to 2° / s, and tension threshold ≤ 0.5N;

[0098] Fatigue characteristics of electromyography signals (integrated electromyography value decrease rate ≥20%, continuous monitoring for 30s): Automatically pause training and trigger rest prompt (rest duration 5-10min, adjusted according to fatigue level);

[0099] In a 4-week clinical trial involving 10 stroke patients with hand dysfunction, this adaptive regimen improved patient training compliance by 40% and reduced the incidence of muscle spasticity by 65% ​​(from 32% to 11.2%), which was significantly better than the traditional fixed regimen.

[0100] (5) Closed-loop drive control module

[0101] The closed-loop drive control module is communicatively connected to the adaptive rehabilitation program generation module and the equipment-side control unit, respectively. It uses a PID control algorithm to achieve precise traction control, specifically including:

[0102] Command issuance: Convert the motor control parameters (target speed, target angle, tension threshold) in the rehabilitation plan into PWM drive commands (frequency 1kHz) and send them to the device control unit;

[0103] Real-time feedback adjustment: The actual values ​​of joint angle and traction tension transmitted back from the receiving device are compared with the target values. If the angle deviation is ≥3° or the tension deviation is ≥0.5N, the motor PWM duty cycle is dynamically adjusted through a PID algorithm (proportional coefficient Kp=0.8, integral coefficient Ki=0.2, derivative coefficient Kd=0.1). The adjustment response time is ≤100ms, ultimately achieving an angle deviation ≤±2° and a tension deviation ≤±0.5N (this accuracy index has been verified by laboratory calibration tests, and the repeatability error is ≤±1° within the 0-90° angle range).

[0104] Safety protection control: When the tension value is ≥5N (the maximum safe tolerance value of human hand muscles, determined based on human biomechanical testing), the joint angle exceeds the safe range (>90°), or muscle spasm is detected, the motor drive signal is immediately cut off, triggering the flexible control screen's audible and visual alarm (alarm intensity ≥60dB, screen flashing frequency 2Hz) and voice prompts to prevent secondary injury;

[0105] Compared to existing open-loop control equipment (angle deviation ≥ ±5°, tension deviation ≥ ±1N), this module improves control accuracy by more than 60% and safety protection response speed by 50%.

[0106] (6) Human-computer interaction module

[0107] Visualization: The flexible control screen displays the training mode, current joint angle, muscle strength level, training duration, and fatigue level in real time, and simultaneously displays the finger movement target trajectory animation (animation frame rate ≥15fps), making it easy for users to intuitively understand the training status.

[0108] Touch interaction: Supports users to manually switch training modes, adjust motor speed / tension threshold (adjustment step size 0.5° / s, 0.1N), and trigger emergency pause by long-pressing the touch area for ≥3 seconds;

[0109] Voice interaction: It integrates a speech recognition unit (recognition rate ≥90%, supports 10 core commands) and a speech synthesis unit, supports voice commands (such as "switch to passive training", "speed up", "pause training"), and provides voice feedback on training progress (such as "training has been going on for 10 minutes, 5 minutes remaining") and evaluation results (such as "muscle strength improved by 0.3 levels in this training"), which is suitable for patients with limited hand operation.

[0110] According to user experience testing, the ease of use of this module scored 4.2 out of 5, which is significantly higher than that of traditional devices (below 3.0).

[0111] (7) Data storage and feedback module

[0112] The data storage and feedback module is integrated with the multi-source data fusion analysis module and the adaptive rehabilitation plan generation module. Data storage: User multi-source collected data, fusion analysis results, rehabilitation plans and execution data are stored by timestamp to form personalized rehabilitation records, with a storage period of ≥12 months; it supports encrypted data transmission and storage, using the AES-256 encryption algorithm (key length 256 bits, encryption mode CBC) to protect user privacy data, in accordance with the medical data security standard (GB / T31491-2015).

[0113] Rehabilitation effect assessment: Based on historical data, the improvement in muscle strength and joint range of motion are calculated using a trend analysis model (using linear regression + moving average algorithm), and weekly / monthly rehabilitation reports are generated.

[0114] Remote monitoring: The cloud server module (using Alibaba Cloud ECS server with a data storage capacity of ≥10TB) communicates with the rehabilitation physician's terminal, allowing physicians to remotely view rehabilitation data, adjust training program parameters, and simultaneously push reports to the user's mobile APP, thus combining home rehabilitation with professional guidance.

[0115] Further optimized, the CNN-LSTM hybrid model is trained using transfer learning. It is first pre-trained on a publicly available hand rehabilitation dataset, and then fine-tuned online using personalized user data, with a model update cycle of ≤7 days.

[0116] Furthermore, the data storage and feedback module supports encrypted data transmission and uses the AES-256 encryption algorithm to protect user privacy data.

[0117] When the present invention is in operation: device startup and wearing: after the user puts on gloves 1, the system starts up and completes sensor self-test (including electromyography sensor signal integrity, posture sensor calibration, motor no-load operation test, and flexible screen display normality). If the detection is abnormal, the flexible screen displays the fault type (such as "electromyography sensor not connected" or "motor stuck") and provides voice prompts; after the self-test is passed, the flexible control screen 7 displays "initialization completed".

[0118] Data Acquisition and Preprocessing: The system acquires multi-source data in real time, including user finger electromyography (EMG) signals (such as proximal phalanx root mean square value 0.2-0.5mV), joint angle (initial value 0°), traction tension (initial value 0N), and touch interaction signals; the EMG signals are subjected to 50Hz notch filtering and 5-layer db4 wavelet denoising to extract integrated EMG values ​​(iEMG); the posture signals are subjected to Kalman filtering to eliminate measurement noise; and the tension signals are subjected to 50ms sliding window smoothing to remove abnormal peaks.

[0119] Multi-source data fusion analysis: Preprocessed electromyography, posture, tension, and interaction features are fused through an attention mechanism to construct a dynamic weight matrix (e.g., for patients with muscle strength grade 2, the weights of electromyography features are 0.35, posture features are 0.35, tension features are 0.15, and interaction features are 0.15), outputting a 27-dimensional global feature vector; the global features are input into a CNN-LSTM hybrid model to identify the user's movement intention (e.g., "active flexion") and assess functional status (e.g., muscle strength grade 2, joint range of motion grade 1, and muscle spasticity grade 0).

[0120] Rehabilitation program generation and execution: The adaptive rehabilitation program generation module generates an active training mode based on the recognition results, sets the motor speed to 4° / s, the rotation angle range to 45° (50% of the joint range of motion), and the tension threshold to 1N; the closed-loop drive control module converts the parameters into PWM drive commands and sends them to the device control unit to drive the motor to rotate.

[0121] Closed-loop adjustment and safety control: During training, the system collects the actual joint angle and traction tension signals in real time. If the actual joint angle is detected to be 42° (deviation 3°), the PID algorithm (Kp=0.8, Ki=0.2, Kd=0.1) is used to adjust the duty cycle of the 4PWM motor, and the angle is quickly corrected to 45°. The final angle deviation is ≤±2°. The flexible control screen 7 displays "Active flexion training in progress, current angle 45°, muscle strength level 2" in real time, and plays the finger flexion animation simultaneously.

[0122] Data storage and feedback: After 20 minutes of training, the system automatically stores the multi-source data, fusion analysis results and execution parameters of this training; it calculates the rehabilitation effect through a trend analysis model, generates a weekly report showing "muscle strength improved by 0.5 grade, joint range of motion improved by 10°", and pushes it to the user's mobile APP and the doctor's terminal at the same time. The doctor can remotely adjust the subsequent training parameters based on the report.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hand rehabilitation device based on multi-source data analysis, characterized in that: include: Wearing gloves (1), the wearing gloves (1) are made of flexible and breathable silicone-fabric composite material, including finger parts (2) corresponding to the five fingers of the human body and palm parts, each finger part (2) is divided into three segmented installation positions along the length direction: proximal phalanx, middle phalanx and distal phalanx, and the wearing gloves (1) are pre-embedded with polytetrafluoroethylene wear-resistant sleeves as traction rope channels; flexible array electromyography sensors (3), each finger part (2) is fixed with the electromyography sensor (3) at the proximal phalanx, middle phalanx and distal phalanx installation positions, the electromyography sensor (3) has a sampling frequency ≥1000Hz, and is used to collect the original electromyography signals of different parts of the fingers, including amplitude, frequency and time frequency domain characteristics; The drive assembly includes five micro servo motors (4), five traction wheels (5), and five sets of traction ropes (6). The five motors (4) correspond one-to-one with the five fingers of the human body and are fixed to the bottom of the palm of the glove (1). The output end of the motor (4) is coaxially fixed to the traction wheel (5), and the motor (4) integrates an encoder for collecting speed and angle signals. Each set of traction ropes (6) is made of ultra-high molecular weight polyethylene. One end of the rope is wound around and fixed to the corresponding traction wheel (5), and the other end passes through the polytetrafluoroethylene wear-resistant sleeve inside the glove (1) and is fixed to the end of the corresponding finger (2). The flexible control screen (7) is a flexible OLED screen, which is fitted and installed on the top of the back of the hand of the glove (1). It integrates a touch interaction unit, a display unit, and a physiological signal acquisition unit. The touch interaction unit supports multi-touch, and the display unit uses... To display rehabilitation training data, the physiological signal acquisition unit includes silver electrode pads distributed on the edge of the flexible control screen (7) for acquiring the user's skin contact impedance signal; auxiliary sensing components, including a six-axis posture sensor and a miniature tension sensor; the posture sensor is installed at the joint of each finger (2) for acquiring the angle, angular velocity, and angular acceleration signals of the finger joint; the tension sensor is connected in series in the middle of each traction rope (6), with a range of 0-10N and an accuracy of ≤0.1N, for acquiring the real-time tension signal of the traction rope (6); the device end control unit is electrically connected to the electromyography sensor (3), motor (4), flexible control screen (7), posture sensor, and tension sensor respectively, and has a built-in Bluetooth / wireless communication module for receiving the signals acquired by each sensor, driving the motor (4) to rotate, and interacting with the external system.

2. The hand rehabilitation device based on multi-source data analysis according to claim 1, characterized in that: The traction wheel (5) is made of nylon and has a diameter of 8-12mm. Anti-slip grooves are opened on the surface. The traction rope (6) is wrapped around the traction wheel (5) and fixed in the grooves of the traction wheel (5) by a buckle.

3. The hand rehabilitation device based on multi-source data analysis according to claim 1, characterized in that: The device control unit has a built-in overcurrent and overvoltage protection circuit. When the operating current of the motor (4) exceeds the preset threshold (≥500mA), the power supply to the motor is automatically cut off and an alarm is triggered.

4. A hand rehabilitation system based on multi-source data analysis, in conjunction with a hand rehabilitation device based on multi-source data analysis as described in any one of claims 1-3, characterized in that: The system includes a local terminal module (tablet PC / industrial control computer) and a cloud server module, specifically including the following functional modules: (1) Data acquisition module The data acquisition module is connected to the device control unit for real-time reception and aggregation of multi-source acquired data, which is stored according to timestamp classification. The multi-source acquired data includes: Electromyography (EMG) signal data: raw EMG signals from the proximal / middle / distal phalanges of the fingers, as well as time-domain characteristics (root mean square, integrated EMG value, number of zero crossings) and frequency-domain characteristics (centroid of the spectrum, average power frequency) after preliminary filtering. Motion posture data: finger joint angle (0-90°), angular velocity (0-10° / ms), angular acceleration (0-5° / ms²), and speed (0-60rpm) and rotation angle (0-360°) signals collected by the motor encoder; Tension feedback data: Real-time tension value of the traction rope, sampling frequency ≥200Hz; Interactive sensing data: touch position, touch pressure (0-5N), touch duration, and skin contact impedance (100-1000Ω) signal of the flexible control screen; (2) Data preprocessing module The data preprocessing module is communicatively connected to the data acquisition module and is used to perform noise reduction, calibration, feature extraction, and normalization processing on multi-source data, specifically including: Electromyography signal preprocessing: 50Hz notch filtering is used to eliminate power frequency interference, wavelet denoising (db4 wavelet, 5-level decomposition) is used to remove baseline drift, and linear interpolation is used to complete missing data. Finally, standardized electromyography feature vectors are output. Motion posture data preprocessing: Kalman filtering is used to eliminate measurement noise from the posture sensor, outlier removal (3σ criterion) is performed on the motor encoder data, and smoothed joint motion parameters are output. Tension feedback data preprocessing: A 50ms sliding window is used for smoothing to remove abnormal peaks >8N and output a continuous tension curve; Interactive perception data preprocessing: Physically calibrate the touch coordinates, normalize the skin contact impedance signal to the 0-1 range, and extract touch intent features (such as single-point long press corresponding to "pause training" and multi-point swipe corresponding to "adjust mode"). (3) Multi-source data fusion analysis module The multi-source data fusion analysis module is communicatively connected to the data preprocessing module, and adopts a two-level fusion algorithm of "feature layer + decision layer", specifically including: Feature layer fusion: The preprocessed electromyographic features (12-dimensional), posture features (9-dimensional), tension features (1-dimensional), and interaction features (5-dimensional) are mapped to the same high-dimensional feature space, and an attention weight matrix is ​​constructed (the weight values ​​are dynamically adjusted according to the user's historical rehabilitation data, with electromyographic feature weights of 0.4, posture features of 0.3, tension features of 0.15, and interaction features of 0.15), and the weighted fusion of the 27-dimensional global feature vector is output. Decision layer fusion: The global feature vector is input into a pre-trained CNN-LSTM hybrid model, which includes 2 convolutional layers (extracting spatial features), 2 LSTM layers (extracting temporal features), and 1 fully connected layer. The output is: Motor intention recognition results: active flexion, active extension, passive rehabilitation, cessation of training, mode switching (accuracy ≥ 92%). Hand function status assessment results: muscle strength grade (1-5), joint range of motion grade (1-3), degree of muscle spasm (0-2); (4) Adaptive Rehabilitation Program Generation Module The adaptive rehabilitation program generation module is communicatively connected to the multi-source data fusion and analysis module, and is used to generate a real-time adjustable rehabilitation training program based on movement intention and functional status, specifically including: Training mode selection: Active training mode (assisted movement driven by traction components based on user electromyographic signals), passive training mode (driving finger movement according to a preset trajectory), resistance training mode (dynamically adjusting traction tension), relaxation mode (low speed, low tension traction); Training parameter dynamic adjustment rules: Muscle strength grade ≤ 2: motor speed ≤ 5° / s, rotation range is 50% of joint range of motion, tension threshold ≤ 1N; Muscle strength grade 3-4: motor speed 8-12° / s, rotation range of 80%-100% of joint range of motion, tension threshold 1-3N; Muscle strength level ≥ 5: Motor speed 10-15° / s, start resistance training, tension threshold 3-5N; Muscle spasm severity ≥ Grade 2: Immediately switch to relaxation mode, reduce motor speed to 2° / s, and tension threshold ≤ 0.5N; Fatigue characteristics of electromyography signals (integrated electromyography value decrease rate ≥20%): Automatically pause training and trigger rest prompt; (5) Closed-loop drive control module The closed-loop drive control module is communicatively connected to the adaptive rehabilitation plan generation module and the equipment-side control unit, respectively, and uses a PID control algorithm to achieve precise traction control, specifically including: Command issuance: Convert the motor control parameters (target speed, target angle, tension threshold) in the rehabilitation plan into PWM drive commands and issue them to the device control unit; Real-time feedback adjustment: The actual values ​​of joint angle and traction tension transmitted back from the receiving device are compared with the target values. If the angle deviation is ≥3° or the tension deviation is ≥0.5N, the motor PWM duty cycle is adjusted through the PID algorithm (Kp=0.8, Ki=0.2, Kd=0.1). Safety protection control: When the tension value is ≥5N, the joint angle exceeds the safe range (>90°) or muscle spasm is detected, the motor drive signal is immediately cut off, triggering the audible and visual alarm and voice prompt of the flexible control screen; (6) Human-computer interaction module The human-computer interaction module is communicatively connected to the flexible control screen and the multi-source data fusion analysis module to achieve multi-dimensional interaction, specifically including: Visualization: The flexible control screen displays the training mode, current joint angle, muscle strength level, training duration, and fatigue level in real time, and simultaneously displays the animation of the finger movement target trajectory; Touch interaction: Supports users to manually switch training modes, adjust motor speed / tension thresholds, and trigger an emergency pause by pressing and holding the touch area for ≥3 seconds; Voice interaction: It integrates a speech recognition unit (recognition rate ≥90%) and a speech synthesis unit, supports voice commands (such as "switch passive training" "speed up"), and provides voice feedback on training progress and evaluation results; (7) Data storage and feedback module The data storage and feedback module is communicatively connected to the multi-source data fusion and analysis module and the adaptive rehabilitation plan generation module, respectively, for full-cycle rehabilitation management, specifically including: Data storage: Store user data collected from multiple sources, fusion analysis results, rehabilitation plans and execution data by timestamp to form personalized rehabilitation records, with a storage period of ≥12 months; Rehabilitation effect assessment: Based on historical data, the improvement in muscle strength and joint range of motion is calculated through a trend analysis model, and weekly / monthly rehabilitation reports are generated; Remote monitoring: The cloud server module communicates with the rehabilitation physician's terminal, allowing the physician to remotely view rehabilitation data, adjust training program parameters, and simultaneously push reports to the user's mobile APP.

5. A hand rehabilitation system based on multi-source data analysis according to claim 4, characterized in that: The CNN-LSTM hybrid model is trained using transfer learning. It is first pre-trained on a publicly available hand rehabilitation dataset, and then fine-tuned online using personalized user data. The model update cycle is ≤7 days.

6. A hand rehabilitation system based on multi-source data analysis according to claim 4, characterized in that: The data storage and feedback module supports encrypted data transmission and uses the AES-256 encryption algorithm to protect user privacy data.

Images