Intention-driven active-passive conversion pneumatic rehabilitation glove system and control method

Abstract

The application discloses an intention-driven active-passive conversion pneumatic rehabilitation glove system and a control method, relates to the fields of bioelectric signal processing and pneumatic flexible driving control technology, and the system comprises a signal acquisition layer, an intention decoding and conversion control layer and an execution feedback layer. The signal acquisition layer collects electroencephalogram, surface electromyogram and motion sensing signals; the intention decoding and conversion control layer quantifies motor imagery electroencephalogram signals into continuous quality scores through a brain intention quality evaluator, drives automatic conversion of three stages of passivity, assistance and activity by a gradual conversion strategy engine with the quality scores, dynamically and adaptively adjusts thresholds, and maps the quality scores into assistance instructions by an adaptive assistance decision maker; and the execution feedback layer drives the pneumatic flexible glove to execute actions and provide feedback. The application solves the problems of intention detection binary, mode switching relying on manual operation, lack of neural plasticity-oriented optimization and multi-loop coordination in the prior art.

Description

Technical Field

[0001] This application relates to the fields of bioelectric signal processing and pneumatic flexible drive control technology, and in particular to an intention-driven active-passive switching pneumatic rehabilitation glove system and control method. Background Technology

[0002] Approximately 60% to 80% of stroke patients experience residual hand motor dysfunction, requiring a step-by-step training path of "passive activation—assisted transition—active control." Existing hand function rehabilitation equipment mainly includes passive / multi-modal pneumatic rehabilitation gloves, brain-computer interface triggered rehabilitation systems, EEG-EMG fusion rehabilitation equipment, and EMG intention-driven rehabilitation gloves.

[0003] However, the aforementioned existing technologies generally share the following common technical problems: Binarization of brain intent detection: Existing brain-computer interfaces only make trigger-based judgments of "having / not having intent", and cannot continuously quantify the quality of motor imagery EEG signals, resulting in the device being unable to distinguish the strength of intent and the output lacking specificity.

[0004] Mode switching relies on manual intervention: switching between passive, assisted, and active training modes depends on the therapist's manual selection, lacking an automatic and gradual conversion mechanism based on the patient's real-time status, and the switching process is discrete and abrupt, resulting in an abrupt experience for the patient.

[0005] Lack of neuroplasticity guidance: It does not consider the temporal locking relationship between intention and action feedback, and lacks switching timing optimization and delay control based on Heb relational plasticity theory.

[0006] Lack of coordination among multi-source signals: EEG, EMG, motion sensing and other multi-loop signals are mostly controlled by single or double loops, lacking a three-way coordinated fusion closed loop, making it difficult to balance the accuracy and safety of assistance. Summary of the Invention

[0007] To overcome the technical problems of the prior art, this application provides an intention-driven active-passive switching pneumatic rehabilitation glove system and control method. The system quantifies motor imagery EEG signals into continuous quality scores using a brain intention quality evaluator. Driven by these scores, a progressive switching strategy engine automatically manages the judgment and switching between passive, assisted, and active stages. The switching threshold is dynamically and adaptively adjusted based on sliding window statistics and rehabilitation assessment. A Hebbe pre-inflation strategy optimizes the delay of intentional actions. An adaptive assist decision-maker achieves additive fusion closed-loop control of EEG, EMG, and motion sensing signals. This system achieves continuous quantification and differentiated assistance of brain intention quality, eliminating the dependence on manual switching of training modes. The switching threshold is personalized as rehabilitation progresses. Pre-inflation shortens the delay to improve the efficiency of neural plasticity induction. Multi-source signal synergistic fusion balances accuracy and safety in assisting the glove.

[0008] This application provides an intention-driven active-passive switching pneumatic rehabilitation glove system, comprising a three-layer closed-loop architecture: a signal acquisition layer, an intention decoding and conversion control layer, and an execution feedback layer. The signal acquisition layer includes an EEG signal acquisition module, a surface electromyography (EMG) acquisition module, and a motion sensing module, which are used to acquire EEG signals, surface EMG signals, and joint flexion angle data in real time. The intention decoding and conversion control layer includes a brain intention quality evaluator, an adaptive assist decision-maker, and a progressive conversion strategy engine. It is used to receive multimodal signals from the signal acquisition layer, calculate a comprehensive brain intention quality score including event-related desynchronization strength, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency, map the quality score to an assist level, and combine electromyographic inhibition correction and motion error compensation correction to output an assist command. The execution feedback layer includes a pneumatic flexible glove execution module, a multimodal feedback module, and a rehabilitation assessment module. It is used to receive the assistance commands and target air pressure values ​​output by the intention decoding and conversion control layer, drive the execution pneumatic glove, and transmit the motion feedback and assessment results back to the intention decoding and conversion control layer to achieve real-time closed loop.

[0009] Furthermore, the brain intent quality evaluator is used to convert the binary intent detection of traditional brain-computer interfaces into a continuous comprehensive quality score. The comprehensive quality score It is obtained by weighted fusion of the following four sub-components normalized to [0,1]: Event-related desynchronization strength Calculate the power decline rate of the contralateral motor cortex leads during the task period relative to the baseline period in the μ and β bands, and normalize the power decline rate; 1 is obtained when the healthy side or rehabilitation target level is reached, and 0 is obtained when no effective desynchronization occurs; Frequency band energy ratio shift : Reflects the degree of shift of the activation mode relative to the baseline, and is obtained by mapping the ratio of μ-band power to β-band power during the mission period and the baseline period through the sigmoid function; Classifier confidence After spatial filtering using the Common Spatial Pattern (CSP), a Support Vector Machine (SVM) or Linear Discriminant Analysis (LDA) is used to perform binary classification of the motor imagery state or resting state on the affected side, and the posterior probability of the output motor imagery state is obtained; the classifier is trained offline and updated online using the patient's historical labeled data; Spatial side-ization consistency The value was obtained by normalizing the ratio of the desynchronization amplitude of the contralateral and ipsilateral motor cortexes. In healthy individuals, desynchronization of the contralateral hemisphere was stronger than that of the ipsilateral hemisphere during motor imagery. In stroke patients, desynchronization of the contralateral hemisphere was impaired, and the level of spatial lateralization consistency was reduced.

[0010] The overall quality score The calculation process uses the following formula:

[0011] in, , , and These represent the event-related desynchronization strength, band energy ratio shift, classifier confidence, and spatial desynchronization consistency, respectively. , , and This represents the weighting coefficient.

[0012] The weighting coefficients are based on the patient's actual athletic performance. Perform online adaptive updates, including: After each training unit is completed, the actual motion performance score given by the motion sensing module will be used as the benchmark. To achieve the goal of improving the overall quality score track The actual motion performance is divided into The normalized fusion of action completion rate, trajectory smoothness, and active force ratio ; Using the mean square error loss function The goal of online adaptive updating is to quantify the deviation between the output of the current brain intention quality assessor and the patient's actual motor performance, and to adjust the weighting coefficients accordingly. Minimize; the loss function Expressed as a formula:

[0013] in This represents the total number of consecutive action samples contained within a training unit, i.e., the number of samples collected. The total number of data pairs; and They represent the first The overall quality score and actual athletic performance score of each sample As the target value for supervised learning, it represents the first... The actual quality of the movement completed by the patient's affected hand during the second action; Based on calculations Mean squared error loss function value of the next action The fusion weights of the four sub-components in the brain intent quality evaluator were adjusted using the gradient descent method. , , and After updating, the projection back is non-negative and the sum is 1, thus improving the overall quality score. The correlation with the patient's actual motor ability is gradually maximized.

[0014] Furthermore, the adaptive assist decision-maker continuously maps the comprehensive quality score to an assist level within the allowable assist range of the current training phase, and corrects it using two signals: electromyography and motion error. The progressive transition strategy engine provides a current stage support level range. Internally, the overall quality score is calculated using an S-shaped function. Mapped to baseline boost level :

[0015] in, and These represent the lower and upper limits of the current training stage's support level range, respectively. The steepness of the transition of the sigmoid function. The midpoint of the auxiliary area and This determines the center position of the S-curve, that is, at this point When the value is reached, the benchmark assist level is the midpoint of the interval. ,in and The stage division threshold set for the progressive transition strategy engine; Residual muscle strength was obtained by root mean square normalization of each channel of surface electromyography. Calculate the electromyographic inhibition correction amount ,in The suppression coefficient is calculated by comparing the actual angle obtained from the bending sensor with the target trajectory, and then calculating the normalized error. The motion error compensation correction amount is expressed as ,in This is the compensation coefficient; The final assist level is obtained by additively fusing the above three data sources. :

[0016] in, This represents a limiting function, used to restrict the value within the parentheses to a certain value. Within the range; and converted to target air pressure. ,in This is the maximum air pressure of the system.

[0017] Furthermore, the progressive transition strategy engine manages the determination of passive, auxiliary, and active stages, divides the assistance level range for each stage, and performs threshold adaptation and smooth transition. The determination of passive, auxiliary, and active stages and the corresponding assistance level range division strategy include: When the comprehensive quality score Less than the preset first threshold When the time is right, it is determined to be the passive stage, and the assistance level range is set to [0.7, 1.0]. At this time, the rehabilitation glove is the main driving force, and the goal is to maintain joint range of motion, establish an initial connection between intention-driven pneumatic movement and receiving proprioceptive feedback. When the comprehensive quality score Greater than or equal to the preset first threshold And less than the preset second threshold When the patient actively exerts force, the rehabilitation gloves are adjusted to the final assistance level, with the goal of strengthening the muscle pathway between the cortex and muscles. When the comprehensive quality score Greater than or equal to the preset second threshold When the time is right, it is determined to be the active phase, and the assistance level range is set to [0, 0.3], which mainly focuses on monitoring or mild resistance, with the goal of consolidating autonomous movement.

[0018] Furthermore, the threshold adaptation includes: Maintenance length is A sliding window is used to calculate the overall quality score within the window. mean with standard deviation ;when Continuously above the preset average threshold and When it is less than the preset standard deviation threshold, according to the formula Raise the threshold The updated threshold is obtained. ,in Stage division thresholds set for the progressive transition strategy engine and , The default value is 0.02. The recovery progress factor is calculated using the following formula:

[0019] and These represent the current and previous scores on the Fugl-Meyer Upper Limb Motor Assessment Scale, respectively. The full score of the Fugl-Meyer Upper Limb Motor Assessment Scale is given. This function limits the values ​​within the brackets to the range [0,1]; it includes a backlash mechanism to prevent values ​​from rolling back only when the values ​​within the window are consecutive. Second-rate Value is lower than the current threshold Only then was it lowered; a safety lower limit was set. ≥0.15 ≥0.5.

[0020] Furthermore, the intent decoding and conversion control layer of the intent-driven active-passive switching pneumatic rehabilitation glove system also includes a Hebbian coupling window optimization unit, used to monitor in real time the delay from the moment of brain intent confirmation to the moment when the pneumatic telescopic tube establishes effective torque. Set target delay windows at different training stages, passive stage ≤300ms, auxiliary phase 150≤ ≤200ms, active phase ≤100ms; The Herb coupling window optimization unit adopts a pre-inflation strategy. Before the comprehensive quality score is confirmed, the probability of intention occurrence is estimated based on the trend of EEG characteristics. When the estimated probability exceeds the pre-inflation threshold, the pneumatic telescopic tube is pre-inflated to a low-pressure standby state. Once the comprehensive quality score is confirmed, the pressure is increased from the standby pressure to the target pressure, shortening the effective torque establishment time.

[0021] Furthermore, the system also includes a security protection mechanism: Abnormal signal protection: If the overall quality score changes by more than 0.5 within 1 second, it is judged as an abnormal signal, the assist is paused and a recalibration prompt is given; Pressure limits: Pressure not exceeding 200 kPa, rate of change not exceeding 50 kPa / s; Spasm protection: The bending sensor immediately releases pressure and pauses when it detects involuntary rapid flexion; Fatigue monitoring: When the overall quality score of multiple consecutive movements shows a significant downward trend, it is considered fatigue, and the intensity is reduced or a rest is suggested.

[0022] This application also provides an intention-driven active-passive switching pneumatic rehabilitation glove control method, including: Step S1: Initialization, loading patient historical data and current parameters, including assistance level range thresholds. and , Steepness of transition of sigmoid function Weighting coefficients in the overall quality score calculation formula , , and Each stage has a range of support levels; Step S2: Simultaneously collect EEG, surface electromyography, and motion sensing signals; Step S3: Preprocess the EEG and extract event-related desynchronization intensity, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency; calculate the overall quality score. ; Step S4, by Determine the current training phase and identify the assistance level range. ; Step S5: Calculate the baseline assist level within the interval using S-shaped mapping. ; Step S6, calculate residual muscle strength With electromyographic inhibition ; Step S7, calculate the normalized motion error With compensation amount ; Step S8: Additively integrate the baseline assistance level, electromyographic inhibition, and compensation level, and limit the amplitude to obtain the final assistance level. Further conversion to target air pressure ; Step S9: Implement the pre-inflation strategy, monitor and optimize the intention action delay. ; In step S10, the proportional valve drives the pneumatic telescopic tube to perform bending and stretching, and smoothly transitions across the threshold according to the transition window. Step S11: Collect motion and evaluation feedback, and update the sliding window. statistics; Step S12, dynamically adjust according to rehabilitation assessment and and weight , , and ; Step S13: Determine if the process has ended. If not, return to step S2 and repeat.

[0023] This application discloses the following technical effects: This application provides an intention-driven active / passive switching pneumatic rehabilitation glove system and control method, addressing the limitations of existing brain-computer interfaces that only support intentional / passive switching. Figure 2 To overcome the limitations of quantitative judgment and the inability to distinguish between strong and weak intentions, a continuous brain intention quality score is obtained by fusing four dimensions of features: event-related desynchronization intensity, frequency band energy ratio, classifier confidence, and spatial lateralization consistency. This continuous quantitative index can distinguish between strong and weak intentions, allowing the system to dynamically adjust the level of assistance based on the strength of the intention. When the intention is strong, assistance is automatically reduced to encourage active patient participation; when the intention is weak, assistance is increased to ensure the completion of the action, thus ensuring that the assistance output is accurately matched with the patient's real-time central nervous system state.

[0024] To address the shortcomings of existing equipment training modes, which rely on manual switching and have fixed thresholds that are difficult to adapt to individual differences, this invention uses brain intention quality scores as a driver to automatically switch between passive, assisted, and active stages. The assisted intervals between stages partially overlap and a transition window with linear interpolation is used to achieve smooth, seamless mode switching, eliminating reliance on therapist experience. Simultaneously, the switching threshold is dynamically and adaptively adjusted based on sliding window quality score statistics and rehabilitation assessment scores: the threshold increases rapidly for those with faster rehabilitation progress to accelerate advancement, while the threshold remains stable for those with slower progress to consolidate training. Furthermore, anti-regression lag and a safety lower limit are implemented to ensure the training pace truly matches the individual's rehabilitation progress.

[0025] Furthermore, based on the Hebbian correlation plasticity principle, this invention employs a pre-inflation strategy to inflate the pneumatic telescopic tube to low pressure before intent confirmation, and then rapidly increase pressure after confirmation, thus delaying and compressing the intended action within the target time window. This tightly locks the central intent and peripheral action feedback in time, effectively improving the efficiency of neural plasticity induction. On the other hand, it integrates the three signals—EEG intent, surface electromyography residual muscle strength, and motor sensor feedback—in an additive structure and is protected by stage interval amplitude limiting. This ensures that electromyographic inhibition and error compensation remain effective at each stage, guaranteeing the accuracy of assistance under complex conditions while avoiding the safety risks caused by excessive assistance or insufficient drive. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of an intent-driven active-passive switching pneumatic rehabilitation glove system provided in an embodiment of this application.

[0027] Figure labels: 1. EEG signal acquisition module; 2. Surface electromyography acquisition module; 3. Motion sensing module; 4. Brain intention quality evaluator; 5. Adaptive assist decision-making system; 6. Progressive transition strategy engine; 7. Heb pre-inflated window optimization unit; 8. Pneumatic flexible glove execution module; 9. Multimodal feedback module; 10. Rehabilitation assessment module.

[0028] Figure 2 A flowchart of the brain intent quality evaluator algorithm provided in the embodiments of this application.

[0029] Figure 3 A flowchart of the three-way additive fusion of the adaptive assist decision-maker provided in the embodiments of this application.

[0030] Figure 4 This is a schematic diagram of a three-stage progressive transition strategy provided in an embodiment of this application.

[0031] Figure 5 The timing diagram for the optimized Hebrew coupling window (pre-inflation strategy) provided for the embodiments of this application.

[0032] Figure 6This is a flowchart illustrating an intent-driven active-passive switching pneumatic rehabilitation glove control method provided in an embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Example 1: This application provides an intent-driven active-passive switching pneumatic rehabilitation glove system, such as... Figure 1 As shown, the system comprises a three-layer closed-loop architecture: a signal acquisition layer, an intent decoding and conversion control layer, and an execution feedback layer. The signal acquisition layer includes an EEG signal acquisition module 1, a surface electromyography (EMG) acquisition module 2, and a motion sensing module 3, which are used to acquire EEG signals, surface EMG signals, and joint flexion angle data in real time. In this embodiment, the EEG signal acquisition module uses a non-invasive dry electrode or semi-dry electrode EEG cap with no less than 16 channels, a sampling rate of no less than 500Hz, covering the motor cortex-related leads of C3, C4, Cz, FC3, FC4, CP3, and CP4. The EEG signal is output after pre-amplification, bandpass filtering, and power frequency notch filtering. The surface electromyography (EMG) acquisition module uses no fewer than 4 channels of electrodes, which are attached to the flexor digitorum superficialis, flexor digitorum profundus, and extensor digitorum commonis muscles of the affected forearm. The sampling rate is no less than 1000Hz, and the surface EMG signal is output after being bandpass filtered by 20–500Hz. The motion sensing module includes a flexible bending sensor or inertial measurement unit (IMU) on the back of each finger of the glove to collect the joint bending angle, and a pressure sensor on the palm to collect the gripping force.

[0035] The intent decoding and conversion control layer includes a brain intent quality evaluator 4, an adaptive assist decision-maker 5, and a progressive conversion strategy engine 6. It is used to receive multimodal signals from the signal acquisition layer, calculate a comprehensive brain intent quality score including event-related desynchronization strength, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency, map the quality score to an assist level, and combine electromyographic inhibition correction and motion error compensation correction to output an assist command. The execution feedback layer includes a pneumatic flexible glove execution module 8, a multimodal feedback module 9, and a rehabilitation assessment module 10. It is used to receive the assist command and target air pressure value output by the intention decoding and conversion control layer, drive the execution pneumatic glove, and send the motion feedback and assessment results back to the intention decoding and conversion control layer to achieve real-time closed loop. In this embodiment, the pneumatic flexible glove execution module includes five independent pneumatic telescopic tubes (bellows actuators) corresponding to the five fingers respectively. Each tube is connected to a micro air pump and a proportional valve through an independent air circuit. The air pressure is continuously adjustable in the range of 0–200 kPa and can be driven independently by each finger. The multimodal feedback module includes fingertip vibration motors, LED indicator lights, and voice prompts; the rehabilitation assessment module calculates the completion rate of movements, trajectory smoothness, and active force ratio online, and regularly connects to the Fugl-Meyer Upper Limb Motor Assessment Scale (FMA-UE) and the Action Research Arm Test (ARAT).

[0036] like Figure 2 As shown, the brain intent quality evaluator upgrades the traditional binary intent detection of brain-computer interfaces to continuous quality scoring. Specifically, it includes the following steps: Preprocessing: The raw EEG was subjected to baseline correction, 4th order Butterworth bandpass filtering (1-40Hz), independent component analysis to remove EEG / ECG artifacts, and time window extraction in sequence; Feature extraction: Four sub-components normalized to [0,1] were extracted, including the event-related desynchronization intensity. Frequency band energy ratio shift Classifier confidence Consistency with spatial side-ization The specific calculation process is as follows; Calculate the power drop rate of the contralateral motor cortex leads during the mission period (t=0.5s-3s) in the μ and β bands relative to the baseline period (t=-0.2s-0s). : ,in and These represent the power values ​​of the contralateral motor cortex leads during the mission and baseline periods in the μ and β bands, respectively; and the power decay rate is expressed by the following formula. Normalization is performed to obtain the event-related desynchronization strength. : ,in The reference desynchronization amplitude (e.g., the healthy side or rehabilitation target level) was defined for this patient. This represents a limiting function, used to restrict the value within the parentheses to a certain value. Within range; reaching the reference level That is, 1, indicating that no effective desynchronization has occurred. Then it is 0; Frequency band energy ratio shift Reflecting the degree of shift of the activation mode relative to the baseline, it is the ratio of the μ-band power to the β-band power during the mission period and the baseline period. It is obtained by mapping using the sigmoid function, and the calculation formula is as follows: ,in This represents the ratio of μ-band power to β-band power during the mission period. This represents the ratio of μ-band power to β-band power during the baseline period; , The further away from 1 The closer to the boundary, the closer to 1. ≈0.5; After spatial filtering using the Common Spatial Pattern (CSP), Support Vector Machine (SVM) or Linear Discriminant Analysis (LDA) is used to perform binary classification of the motor imagery or resting state on the affected side. The posterior probability of the output motor imagery is then used to obtain the classifier confidence. , The classifier is trained offline and updated online using patient historical labeled data. The specific implementation process is as follows: Offline Training Phase: During system initialization, labeled EEG data of the patient in resting state and affected side motor imagery state are collected. The covariance matrix is ​​calculated for each class of data. A spatial filter is obtained by solving the generalized eigenvalue problem. This filter projects the multi-channel EEG data in a direction that maximizes the variance ratio between the two classes. For the real-time acquired multi-channel segmented EEG signals, after projection through this spatial filter, the logarithmic variance of each component's projected signal is calculated as a feature vector. The feature vectors from the labeled data are used to train an SVM or LDA binary classifier. For SVM, Platt scaling (Sigmoid probability calibration) is used to map the classifier's decision function values ​​to posterior probabilities in the [0,1] interval. For LDA, assuming that each feature class follows a Gaussian distribution, the posterior probability of belonging to the motor imagery state is directly calculated using Bayes' theorem. After training, an initial classifier model is obtained.

[0037] Online operation and update phase: EEG data is collected in real time, and after feature vector extraction through the same CSP spatial filtering, it is input into the classifier. The output is the posterior probability of the current motor imagery state. After each training unit is completed, the system collects the newly labeled data within that unit (based on the actual motion performance scores provided by the motion sensing module). The corresponding state labels are used to update the classifier model through incremental learning or parameter fine-tuning (such as SVM support vector set update or LDA mean / covariance update), so that the classifier continuously adapts to the changes in the patient's EEG signal as the rehabilitation process progresses, ensuring... The accuracy of the assessment.

[0038] Spatial side-ization consistency The normalized value is obtained by the ratio of the desynchronization amplitude of the contralateral to that of the ipsilateral motor cortex. The calculation formula is as follows: , ,in The original spatial lateralization consistency index is dimensionless and has a theoretical value range of [0.5, 1]. It is used to initially quantify the relative proportion of the desynchronization amplitude of the contralateral and ipsilateral motor cortex. This represents the event-related desynchronization (ERD) amplitude of the contralateral motor cortex leads relative to baseline during the mission period. It is dimensionless, ranges from [0,1], and is determined by the power decay rate. Normalization yields the result; The magnitude of event-related desynchronization in ipsilateral motor cortex leads during the task period relative to baseline; in healthy individuals, desynchronization in the contralateral hemisphere is stronger than that in the ipsilateral hemisphere during motor imagery, while in stroke patients, lateralization is impaired and spatial lateralization consistency is reduced.

[0039] The overall quality score is calculated based on the following formula. : Weight , , and The initial values ​​were (0.35, 0.20, 0.30, 0.15). Weights were assigned based on the patient's actual motor performance. Perform online adaptive updates: After each training unit is completed, the actual motion performance score provided by the motion sensing module is updated. To achieve the goal, track Using loss function Calculate the loss, where This represents the value of the loss function. This represents the total number of consecutive action samples contained within a training unit, i.e., the number of samples collected. The total number of data pairs; and They represent the first The overall quality score and actual athletic performance score of each sample As the target value for supervised learning, it represents the first... The actual quality of the patient's affected hand movement during the second action; based on the loss function. Update weights using gradient descent. , , and And project back to the non-negative constraints that sum to 1.

[0040] Actual athletic performance score The normalized fusion of action completion rate, trajectory smoothness, and active force ratio The specific calculation process includes: after each training unit is completed (such as a set of 10 grasping movements), the data collected by the motion sensing module and the electromyography module are used to calculate three sub-indicators, which are then normalized to [0,1] and weighted and fused. Action completion rate For each single movement, divide it into time frames and compare the actual joint bending angle in each frame. Angle relative to the target trajectory If the absolute error ( If the default angle is 5°, then the frame is considered a valid completion. This is expressed as the ratio of the number of valid completed frames to the total number of frames in the action, and ; trajectory smoothness Extract the curves showing the change of joint angles over time during the movement, and calculate the root mean square value of the acceleration derivative. , ,in The reference smoothness benchmark established for this patient (e.g., maximum smoothness at the beginning of recovery) (or data from healthy individuals); The smaller The closer it is to 1, the smoother the trajectory; The larger The closer to 0; Proactive efforts account for a certain percentage : Root mean square of each channel of surface electromyography With maximum spontaneous contraction The residual muscle strength is obtained by comparing the ratio of the root mean square value of the surface electromyography signals collected when the patient performs maximal force contraction during the system initialization and calibration phase. This reflects the degree to which the patient's agonist muscles are involved in the movement.

[0041] The actual athletic performance score is obtained by weighting and integrating the above three sub-indicators. ,in Initial weights =0.4、 =0.3、 =0.3; This process makes It comprehensively reflects whether the movement is performed correctly, whether the movement is smooth, and how much active force is used, thus fully quantifying the patient's actual motor performance.

[0042] like Figure 3 As shown, the adaptive assist decision-maker will integrate the quality score within the allowable assist interval of each stage. Continuous mapping is used to determine the level of assistance, and corrections are made using both electromyography (EMG) and motion error signals; level of assistance , =1 means completely passive (fully driven gloves). =0 indicates fully active (gloves provide no assistance). Current stage and its assistance level range. Given by the progressive transition strategy engine, the intervals of the three stages shift downwards sequentially, with adjacent intervals partially overlapping. The overlapping segments serve as buffers for the transition between stages. Specifically, the process includes the following steps: Continuous mapping of assist levels within the interval: Within the assist level interval given by the progressive transition strategy engine for the current stage. Internally, using an S-shaped function to... Mapped to baseline boost level :

[0043] in, and These represent the lower and upper limits of the current training stage's support level range, respectively. The steepness of the transition of the sigmoid function. The midpoint of the auxiliary area and This determines the center position of the S-curve, that is, at this point When the value is reached, the benchmark assist level is the midpoint of the interval. ,in and The stage division threshold set for the progressive transition strategy engine; Electromyographic inhibition correction: residual muscle strength is obtained by root mean square normalization of each channel of surface electromyography. Calculate the electromyographic inhibition correction amount ,in The suppression coefficient is, and ; Motion error compensation correction: The normalized error is calculated by comparing the actual angle obtained from the bending sensor with the target trajectory. ,in To determine the maximum angle variation range of this joint in the current training task (i.e., the angle normalization baseline), the motion error compensation correction amount is further obtained. ,in For compensation coefficient, ; The final assist level is obtained by additively fusing the above three data sources. :

[0044] in, This represents a limiting function, used to restrict the value within the parentheses to a certain value. Within the range; and converted into the target air pressure for the proportional valve to drive the pneumatic telescopic tube to perform bending and stretching. ,in This is the system's maximum air pressure.

[0045] Figure 4 Displayed brain intention quality score Assistance Level The relationship and interval division between the passive, auxiliary, and active stages. For example... Figure 4 As shown, the progressive transition strategy engine manages the judgment, interval setting, and threshold adaptation of the passive, auxiliary, and active stages, specifically including the following steps: Three stages and assist interval: Initial threshold in this embodiment =0.3、 =0.7. When the overall quality score... Less than the preset first threshold When the time is right, it is determined to be the passive stage, and the assistance level range is set to [0.7, 1.0]. At this time, the rehabilitation glove is the main driving force, and the goal is to maintain joint range of motion, establish an initial connection between intention-driven pneumatic movement and receiving proprioceptive feedback. When the comprehensive quality score Greater than or equal to the preset first threshold And less than the preset second threshold When the patient actively exerts force, the rehabilitation gloves are adjusted to the final assistance level, with the goal of strengthening the muscle pathway between the cortex and muscles. When the comprehensive quality score Greater than or equal to the preset second threshold When the time is right, it is determined to be the active phase, and the assistance level range is set to [0, 0.3], mainly for monitoring or mild resistance, with the goal of consolidating autonomous movement. The assistance level ranges of adjacent phases partially overlap as a transition buffer.

[0046] Threshold dynamic adaptation: maintenance length is A sliding window is used to calculate the overall quality score within the window. mean with standard deviation ,right and Perform adaptive adjustments separately; when The value remains consistently above a preset average threshold (in this embodiment, the threshold value adjusted accordingly). )and When the value is less than the preset standard deviation threshold (default is 0.1), the formula is used. Raise the threshold The updated threshold is obtained. ,in Stage division thresholds set for the progressive transition strategy engine and , The default value is 0.02. The recovery progress factor is calculated using the following formula:

[0047] and These represent the current and previous scores on the Fugl-Meyer Upper Limb Motor Assessment Scale, respectively. The full score of the Fugl-Meyer Upper Limb Motor Assessment Scale is given. This function limits the values ​​within the brackets to the range [0,1]; it includes a backlash mechanism to prevent values ​​from rolling back only when the values ​​within the window are consecutive. Second-rate Value is lower than the current threshold Only then was it lowered; a safety lower limit was set. ≥0.15 ≥0.5.

[0048] Smooth transition between stages: When crossing a stage threshold, a transition window is set, and within the window, the target is gradually changed from the old stage value to the new stage value through linear interpolation.

[0049] like Figure 1 As shown, the system's intent decoding and conversion control layer also includes a Heblin coupling window optimization unit 7. Figure 5 The timing diagram for Hebbian coupling window optimization consists of two subplots. The upper subplot represents time t on the horizontal axis and intent probability p on the vertical axis, showing an S-shaped upward curve of intent probability, and is labeled with the pre-inflation threshold (0.6) and intent confirmation threshold (0.85). The lower subplot also represents time t on the horizontal axis and air pressure (kPa) on the vertical axis, showing a comparison of air pressure changes with and without pre-inflation (solid line), and is labeled with the standby pressure (20 kPa), effective torque threshold (60 kPa), and target pressure. .

[0050] This timing diagram fully illustrates the entire process from the patient's initial motor imagery to the establishment of effective torque in the pneumatic telescopic tube. The specific timing steps are as follows: Initially, the patient receives motor imagery cues and begins to attempt them; the intention probability curve in the subgraph above shows an S-shaped rise from the baseline. When the probability curve of the moment intention rises to the pre-inflation threshold (0.6), the system triggers the pre-inflation strategy; in the corresponding sub-figure, the air pressure rises rapidly to the standby pressure of 20 kPa and is maintained. At this time, the pneumatic telescopic tube is in the pre-expansion state and does not generate driving displacement. As the probability curve of the intent continues to rise and reaches the intent confirmation threshold (0.85), the system officially confirms the brain intent; corresponding to the sub-graph below, the air pressure rapidly climbs from the standby pressure of 20 kPa to the target pressure. .

[0051] Delay Definition and optimization: The effective torque threshold (60 kPa) is marked in the sub-figure below. The system defines it from the moment of intent confirmation. The time interval until the air pressure rises to the effective torque threshold (60 kPa) is the intentional action delay. The comparison in the figure shows the delay when using a pre-inflation strategy. Significantly less delay than without pre-inflation strategy (The dashed line shows the pressure climbing directly from zero to the target pressure). Through a pre-inflation strategy, the system hides the mechanical response delay of the pneumatic lines and telescopic tubes within... to During the waiting period for confirmation of intent, the time was significantly reduced. The subsequent delay in establishing effective torque ensures that the central intention and peripheral action feedback are closely locked in time.

[0052] Through the above timing process. Figure 5 The mechanism by which the present invention utilizes a pre-inflated strategy to optimize the Heb coupling window and shorten the system response delay is fully disclosed.

[0053] The system also includes security protection mechanisms, specifically including: Abnormal signal protection: If the overall quality score changes by more than 0.5 within 1 second, it is judged as an abnormal signal, the assist is paused and a recalibration prompt is given; Pressure limits: Pressure not exceeding 200 kPa, rate of change not exceeding 50 kPa / s; Spasm protection: The bending sensor immediately releases pressure and pauses when it detects involuntary rapid flexion; Fatigue monitoring: When the overall quality score of multiple consecutive movements shows a significant downward trend, it is considered fatigue, and the intensity is reduced or a rest is suggested.

[0054] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0055] Example 2: This application provides an intention-driven active-passive switching pneumatic rehabilitation glove control method, implemented based on the intention-driven active-passive switching pneumatic rehabilitation glove system described in Example 1, as follows... Figure 6 As shown, the method includes: Step S1: Initialization, loading patient historical data and current parameters, including assistance level range thresholds. and , Steepness of transition of sigmoid function Weighting coefficients in the overall quality score calculation formula , , and Each stage has a range of support levels; Step S2: Simultaneously collect EEG, surface electromyography, and motion sensing signals; Step S3: Preprocess the EEG and extract event-related desynchronization intensity, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency; calculate the overall quality score. ; Step S4, by Determine the current training phase and identify the assistance level range. ; Step S5: Calculate the baseline assist level within the interval using S-shaped mapping. ; Step S6, calculate residual muscle strength With electromyographic inhibition ; Step S7, calculate the normalized motion error With compensation amount ; Step S8: Additively integrate the baseline assistance level, electromyographic inhibition, and compensation level, and limit the amplitude to obtain the final assistance level. Further conversion to target air pressure ; Step S9: Implement the pre-inflation strategy, monitor and optimize the intention action delay. ; In step S10, the proportional valve drives the pneumatic telescopic tube to perform bending and stretching, and smoothly transitions across the threshold according to the transition window. Step S11: Collect motion and evaluation feedback, and update the sliding window. statistics; Step S12, dynamically adjust according to rehabilitation assessment and and weight , , and ; Step S13: Determine if the process has ended. If not, return to step S2 and repeat.

[0056] In this invention, to eliminate dimensional differences between different physiological signals and facilitate multi-source information fusion and closed-loop control, most intermediate calculation variables were normalized. The dimensional relationships in the core formula are clearly defined as follows: Dimensionless variable: Brain intention quality score (including) - ), Assistance Level (including) , , , ), residual muscle strength Normalization error Classifier confidence, threshold ( , , and loss function Sample score in , All values ​​have been normalized and are dimensionless pure numerical values, ranging from [0,1], and directly participate in weighted, mapping, and additive fusion calculations.

[0057] Variables with physical units: target air pressure , by dimensionless boost level Multiply by the system's maximum air pressure The dimension is pressure unit (kilopascal, kPa). This embodiment... The value is 200 kPa; joint flexion angle Data is collected by a bending sensor or inertial measurement unit, with dimensions in degrees (°) or radians (rad); intended action delay. The dimension is time (milliseconds ms), and the optimization target for this embodiment is within 100ms~300ms; the pre-inflation standby pressure is pressure (kPa), and the example value is 20 kPa.

[0058] Consistency of dimensions in formula operations: In the additive fusion formula of the adaptive assist decision-maker, the baseline assist, electromyographic inhibition correction, and motion error compensation correction are all dimensionless values, which can be directly added or subtracted to ensure dimensional consistency.

[0059] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An intention-driven active-passive switching pneumatic rehabilitation glove system, characterized in that, The system comprises a three-layer closed-loop architecture: a signal acquisition layer, an intent decoding and conversion control layer, and an execution feedback layer. The signal acquisition layer includes an EEG signal acquisition module, a surface electromyography (EMG) acquisition module, and a motion sensing module, which are used to acquire EEG signals, surface EMG signals, and joint flexion angle data in real time. The intention decoding and conversion control layer includes a brain intention quality evaluator, an adaptive assist decision-maker, and a progressive conversion strategy engine. It is used to receive multimodal signals from the signal acquisition layer, calculate a comprehensive brain intention quality score including event-related desynchronization strength, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency, map the quality score to an assist level, and combine electromyographic inhibition correction and motion error compensation correction to output an assist command. The execution feedback layer includes a pneumatic flexible glove execution module, a multimodal feedback module, and a rehabilitation assessment module. It is used to receive the assistance commands and target air pressure values ​​output by the intention decoding and conversion control layer, drive the execution pneumatic glove, and transmit the motion feedback and assessment results back to the intention decoding and conversion control layer to achieve real-time closed loop.

2. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 1, characterized in that, The brain intent quality evaluator is used to convert the binary intent detection of traditional brain-computer interfaces into a continuous comprehensive quality score. The comprehensive quality score It is obtained by weighted fusion of the following four sub-components normalized to [0,1]: Event-related desynchronization strength Calculate the power decline rate of the contralateral motor cortex leads during the task period relative to the baseline period in the μ and β bands, and normalize the power decline rate; 1 is obtained when the healthy side or rehabilitation target level is reached, and 0 is obtained when no effective desynchronization occurs; Frequency band energy ratio shift : Reflects the degree of shift of the activation mode relative to the baseline, and is obtained by mapping the ratio of μ-band power to β-band power during the mission period and the baseline period through the sigmoid function; Classifier confidence After spatial filtering using the Common Spatial Pattern (CSP), the motor imagery state or resting state on the affected side is classified into two categories using Support Vector Machine (SVM) or Linear Discriminant Analysis (LDA). The posterior probability of the output motor imagery state is then obtained. The classifier is trained offline and updated online using the patient's historical labeled data. Spatial side-ization consistency The value was obtained by normalizing the ratio of the desynchronization amplitude of the contralateral and ipsilateral motor cortexes. In healthy individuals, desynchronization of the contralateral hemisphere was stronger than that of the ipsilateral hemisphere during motor imagery. In stroke patients, desynchronization of the contralateral hemisphere was impaired, and the level of spatial lateralization consistency was reduced.

3. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 2, characterized in that, The overall quality score The calculation process uses the following formula: in, , , and These represent the event-related desynchronization strength, band energy ratio shift, classifier confidence, and spatial desynchronization consistency, respectively. , , and This represents the weighting coefficient.

4. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 3, characterized in that, The weighting coefficients are based on the patient's actual athletic performance. Perform online adaptive updates, including: After each training unit is completed, the actual motion performance score given by the motion sensing module will be used as the benchmark. To achieve the goal of improving the overall quality score track The actual motion performance is divided into The normalized fusion of action completion rate, trajectory smoothness, and active force ratio ; Using the mean square error loss function The goal of online adaptive updating is to quantify the deviation between the output of the current brain intention quality assessor and the patient's actual motor performance, and to adjust the weighting coefficients accordingly. Minimize; based on calculation Mean squared error loss function value of the next action The fusion weights of the four sub-components in the brain intent quality evaluator were adjusted using the gradient descent method. , , and After updating, the projection back is non-negative and the sum is 1, thus improving the overall quality score. The correlation with the patient's actual motor ability is gradually maximized.

5. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 1, characterized in that, The adaptive assist decision-maker continuously maps the comprehensive quality score to an assist level within the allowable assist range of the current training phase, and corrects it using two signals: electromyography and motion error. The progressive transition strategy engine provides a current stage support level range. Internally, the overall quality score is calculated using an S-shaped function. Mapped to baseline boost level : in, and These represent the lower and upper limits of the current training stage's support level range, respectively. The steepness of the transition of the sigmoid function. The midpoint of the auxiliary area and This determines the center position of the S-curve, that is, at this point When the value is reached, the benchmark assist level is the midpoint of the interval. ,in and The stage division threshold set for the progressive transition strategy engine; Residual muscle strength was obtained by root mean square normalization of each channel of surface electromyography. Calculate the electromyographic inhibition correction amount ,in The suppression coefficient is calculated by comparing the actual angle obtained from the bending sensor with the target trajectory, and then calculating the normalized error. The motion error compensation correction amount is expressed as ,in This is the compensation coefficient; For the aforementioned benchmark assist level The electromyographic inhibition correction amount and the motion error compensation correction amount The final boost level is obtained by performing additive fusion. : in, This represents a limiting function, used to restrict the value within the parentheses to a certain value. Within the range; and converted to target air pressure. ,in This is the system's maximum air pressure.

6. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 1, characterized in that, The progressive transition strategy engine manages the determination of passive, auxiliary, and active stages, divides the assistance level range for each stage, and performs threshold adaptation and smooth transition. The determination of passive, auxiliary, and active stages and the corresponding assistance level range division strategy include: When the comprehensive quality score Less than the preset first threshold When the time is right, it is determined to be the passive stage, and the assistance level range is set to [0.7, 1.0]. At this time, the rehabilitation glove is the main driving force, and the goal is to maintain joint range of motion, establish an initial connection between intention-driven pneumatic movement and receiving proprioceptive feedback. When the comprehensive quality score Greater than or equal to the preset first threshold And less than the preset second threshold When the patient actively exerts force, the rehabilitation gloves are adjusted to the final assistance level, with the goal of strengthening the muscle pathway between the cortex and muscles. When the comprehensive quality score Greater than or equal to the preset second threshold When the time is right, it is determined to be the active phase, and the assistance level range is set to [0, 0.3], which mainly focuses on monitoring or mild resistance, with the goal of consolidating autonomous movement.

7. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 6, characterized in that, The threshold adaptation includes: Maintenance length is A sliding window is used to calculate the overall quality score within the window. mean with standard deviation ;when Continuously above the preset average threshold and When it is less than the preset standard deviation threshold, according to the formula Raise the threshold The updated threshold is obtained. ,in Stage division thresholds set for the progressive transition strategy engine and , The default value is 0.

02. The recovery progress factor is calculated using the following formula: and These represent the current and previous scores on the Fugl-Meyer Upper Limb Motor Assessment Scale, respectively. The full score of the Fugl-Meyer Upper Limb Motor Assessment Scale is given. This function limits the values ​​within the brackets to the range [0,1]; it includes a backlash mechanism to prevent values ​​from rolling back only when the values ​​within the window are consecutive. Second-rate Value is lower than the current threshold Only then was it lowered; a safety lower limit was set. ≥0.15 ≥0.

5.

8. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 1, characterized in that, The system's intent decoding and conversion control layer also includes a Hebbian coupling window optimization unit, used to monitor in real time the delay from the moment the brain intent is confirmed to the moment the pneumatic telescopic tube establishes effective torque. Set target delay windows at different training stages, passive stage ≤300ms, auxiliary phase 150≤ ≤200ms, active phase ≤100ms; The Herb coupling window optimization unit adopts a pre-inflation strategy. Before the comprehensive quality score is confirmed, the probability of intention occurrence is estimated based on the trend of EEG characteristics. When the estimated probability exceeds the pre-inflation threshold, the pneumatic telescopic tube is pre-inflated to a low-pressure standby state. Once the comprehensive quality score is confirmed, the pressure is increased from the standby pressure to the target pressure, shortening the effective torque establishment time.

9. The intention-driven active-passive switching pneumatic rehabilitation glove system as described in claim 1, characterized in that, The system also includes a security protection mechanism: Abnormal signal protection: If the overall quality score changes by more than 0.5 within 1 second, it is judged as an abnormal signal, the assist is paused and a recalibration prompt is given; Pressure limits: Pressure not exceeding 200 kPa, rate of change not exceeding 50 kPa / s; Spasm protection: The bending sensor immediately releases pressure and pauses when it detects involuntary rapid flexion; Fatigue monitoring: When the overall quality score of multiple consecutive movements shows a significant downward trend, it is considered fatigue, and the intensity is reduced or a rest is suggested.

10. A method for controlling an intention-driven active-passive switching pneumatic rehabilitation glove, characterized in that, The method is implemented based on an intention-driven active-passive switching pneumatic rehabilitation glove system according to any one of claims 1-9, and the method includes: Step S1: Initialization, loading patient historical data and current parameters, including assistance level range thresholds. and , Steepness of transition of sigmoid function Weighting coefficients in the overall quality score calculation formula , , and Each stage has a range of support levels; Step S2: Simultaneously collect EEG, surface electromyography, and motion sensing signals; Step S3: Preprocess the EEG and extract event-related desynchronization intensity, frequency band energy ratio shift, classification confidence, and spatial lateralization consistency; calculate the overall quality score. ; Step S4, by Determine the current training phase and identify the assistance level range. ; Step S5: Calculate the baseline assist level within the interval using S-shaped mapping. ; Step S6, calculate residual muscle strength With electromyographic inhibition ; Step S7, calculate the normalized motion error With compensation amount ; Step S8: Additively integrate the baseline assistance level, electromyographic inhibition, and compensation level, and limit the amplitude to obtain the final assistance level. Further conversion to target air pressure ; Step S9: Implement the pre-inflation strategy, monitor and optimize the intention action delay. ; In step S10, the proportional valve drives the pneumatic telescopic tube to perform bending and stretching, and smoothly transitions across the threshold according to the transition window. Step S11: Collect motion and evaluation feedback, and update the sliding window. statistics; Step S12, dynamically adjust according to rehabilitation assessment and and weight , , and ; Step S13: Determine if the process has ended. If not, return to step S2 and repeat.

Images