Self-adaptive training parameter adjusting method for hand rehabilitation exoskeleton

Abstract

The invention discloses a self-adaptive training parameter adjusting method for a hand rehabilitation exoskeleton, and relates to the technical field of rehabilitation medical instruments, and the method comprises the specific steps: deploying a multi-modal sensing system to synchronously collect full-dimensional original signals of a patient, and building a healthy hand motion rhythm feature library; a plurality of indexes are obtained through preprocessing and feature quantitative analysis; judging conflicts and types through intention-conflict decoding; constructing a self-adaptive parameter adjustment model, and executing a flexible solution strategy; a closed-loop adjustment period is set, all the steps are executed circularly, scene parameter adjustment is completed, and advanced rehabilitation training is implemented; according to the method, a comprehensive and accurate patient state sensing system is constructed, the intention and ability matching state is accurately judged, adaptive training parameters are generated, training risks are avoided, safety and adaptability are improved, a closed-loop adjustment mechanism is further constructed, scene-based and advanced training is combined, training is dynamically adjusted, the enthusiasm of a patient is aroused, hand function recovery is accelerated, and the patient experience is improved. The rehabilitation effect and efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation medical device technology, specifically to an adaptive training parameter adjustment method for a hand rehabilitation exoskeleton. Background Technology

[0002] Hand dysfunction severely impacts patients' daily living abilities and social participation, placing a heavy burden on individuals, families, and society. With the deep integration of rehabilitation medicine and engineering technology, hand rehabilitation exoskeletons, as an important assistive rehabilitation device, are gradually becoming a key means to improve hand motor function. These devices, through the combination of mechanical structures and sensing technology, provide patients with targeted motor assistance and training guidance, helping damaged neuromuscular function to gradually recover. Currently, rehabilitation medicine has an increasing demand for refined and personalized training. Exoskeleton systems that rely solely on fixed programs or manual adjustments can no longer meet clinical needs. The industry is gradually focusing on technologies that perceive patient status in multiple dimensions and adapt training parameters in real time. By integrating information from physiological signals, motor intentions, and physical abilities, dynamic optimization of rehabilitation training can be achieved, making the training process more tailored to the individual patient's situation and improving rehabilitation effectiveness and safety. This has become the core trend in the development of hand rehabilitation exoskeleton technology.

[0003] Traditional exoskeleton technologies for hand rehabilitation have many limitations and are difficult to adapt to the complex needs of rehabilitation training. At the signal perception level, most traditional systems only collect single-dimensional motion or physiological signals, failing to comprehensively capture key information such as the patient's physiological fatigue, pain state, movement rhythm, and rehabilitation intentions, leading to biases in the judgment of the patient's true condition. In terms of parameter adjustment, traditional technologies often use preset fixed parameters or simple linear adjustment modes, failing to fully consider the dynamic matching relationship between the patient's rehabilitation intentions and physical capabilities. It is difficult to flexibly adjust the training intensity, speed, and assistance methods according to the real-time status. Some systems lack effective conflict identification and resolution mechanisms, which can easily cause discomfort or even safety risks when the patient's rehabilitation intentions do not match their physical endurance. At the same time, traditional technologies generally lack closed-loop adjustment mechanisms, making it impossible to provide real-time feedback and optimization on training effects and changes in the patient's state, resulting in a relatively rigid training process. In addition, the training scenarios of most traditional systems are detached from real-life situations, making it difficult to effectively translate training results into practical hand functions for patients. Rehabilitation efficiency and patient compliance both need to be improved. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an adaptive training parameter adjustment method for a hand rehabilitation exoskeleton. This method uses a multimodal sensing system to simultaneously collect signals from the patient across all dimensions, including physiological, motor, and intentional data. After preprocessing and feature quantification to analyze core parameters, it relies on rehabilitation intention-conflict decoding logic to determine the matching status and conflict type between intention and ability. Then, it calculates adjustment coefficients through a multi-dimensional coupled adaptive parameter tuning model and optimizes training parameters using differentiated flexible resolution strategies. This achieves closed-loop adjustment throughout the entire process at a fixed cycle, dynamically adapting to changes in the patient's state. Combined with daily scenario-based and progressive training, this allows exoskeleton training to precisely meet individual needs, improving rehabilitation safety and adaptability, and effectively assisting patients with hand dysfunction in restoring motor function.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an adaptive training parameter adjustment method for a hand rehabilitation exoskeleton, the specific steps of which are as follows:

[0006] S1, Synchronous acquisition of multimodal sensor signals: Deploy a multimodal sensor system, configure a unified signal acquisition protocol, and synchronously acquire the full-dimensional raw signals of the patient's physiological fatigue, pain, pre-pain, rhythm of movement of the healthy and affected hands, intention of rehabilitation exercises, and physical abilities. Extract the rhythmic features of the healthy hand's daily movements to build a rhythmic feature library of the healthy hand.

[0007] S2, Multi-source signal preprocessing and feature quantization: Noise reduction, normalization and feature extraction and quantization analysis of the original signals in all dimensions are performed to obtain physiological fatigue degree, pain quantification value, pre-pain feature value, synchronous deviation value of the movement rhythm of healthy and affected hands, and intensity of rehabilitation movement intention. The comprehensive physical ability parameter is calculated by the comprehensive physical ability parameter quantification formula.

[0008] S3, Intention-Conflict Decoding and Type Determination: Input the comprehensive parameters of rehabilitation exercise intention intensity and physical ability into the rehabilitation intention-conflict decoding logic layer, obtain the matching degree through the rehabilitation intention-conflict matching degree calculation formula, and determine whether there is an instantaneous intention-ability conflict and the specific conflict type based on the preset threshold and risk classification standard.

[0009] S4, Multi-dimensional Coupling and Flexible Resolution: Construct a multi-dimensional coupling adaptive parameter tuning model, calculate the adjustment coefficient through the multi-dimensional coupling training parameter adjustment formula, complete the calculation of exoskeleton training parameters, and execute a differentiated flexible resolution strategy based on the conflict determination results;

[0010] S5, Closed-loop adjustment and advanced training: Set a closed-loop adjustment cycle of 500ms, and execute the above steps in a loop to drive the exoskeleton drive module to respond synchronously to the adjustment parameters, complete the scenario-based coupling parameter adjustment of eating, dressing, writing and washing scenarios, and implement advanced rehabilitation training in combination with the recovery of the affected hand function.

[0011] Furthermore, the deployment of the multimodal sensing system specifically involves: attaching surface electromyography (EMG) sensors to the core muscle groups of the patient's fingers, wearing a heart rate variability sensor on the wrist, embedding a thin-film pressure sensor in the key joints of the exoskeleton, fixing a facial expression recognition module in front of the patient's face, wearing a flexible IMU wristband on the healthy hand, and integrating a motion sensing unit into the exoskeleton on the affected hand side. A unified millisecond-level triggered signal acquisition protocol is configured to synchronously acquire EMG and heart rate variability signals representing the patient's physiological fatigue, joint impedance and facial micro-expression signals representing pain and pre-pain, movement trajectory and timing signals representing the rhythm of movement in both the healthy and affected hands, EMG intensity and trend signals representing the intention to perform rehabilitation exercises, and joint force and impedance feedback signals representing physical capabilities.

[0012] Furthermore, the noise reduction process employs three methods: adaptive filtering, moving average filtering, and wavelet denoising, adapted to different original signals. Specifically, signals related to physiological fatigue and rehabilitation exercise intention use 5Hz-500Hz bandpass filtering, signals related to pain and pre-pain use wavelet denoising with 3-5 decomposition layers, and signals related to the rhythm of healthy and affected hands use moving average filtering with a sliding window size of 5-10 sampling points. The normalization process uses linear normalization to uniformly map all denoised signals to the [0, 1] value range. During feature extraction and quantification, physiological fatigue is obtained by weighted fusion of the mean decay rate of the electromyographic signal MPF and the time-domain features of heart rate variability. Pain quantification and pre-pain feature values ​​are obtained by quantifying the peak value of human-computer interaction force and the change features of micro-expression amplitude. The intensity of rehabilitation exercise intention is obtained by analyzing the peak value and rising slope of the electromyographic signal intensity.

[0013] Furthermore, the mathematical expression of the formula for quantifying the comprehensive parameters of physical capabilities is as follows: Where S is the comprehensive parameter of physical ability, and k Z W is the joint impedance correction factor. Z W F W Pp W Ff These are the weighting coefficients for each indicator. Z represents the real-time joint impedance value. max Z min The maximum and minimum joint resistance values ​​are individually calibrated for each patient; F represents the normalized physiological fatigue level; P... p F represents the normalized pre-pain characteristic value. f F is the real-time feedback value of limb force exertion. fmax The maximum value of limb force feedback is individually calibrated for each patient.

[0014] Furthermore, the rehabilitation intention-conflict decoding logic layer is divided into a signal fusion layer and a conflict determination layer. The signal fusion layer performs data alignment and feature weighted fusion on the comprehensive parameters of rehabilitation exercise intention intensity and physical ability, with both fusion weights ranging from 0.4 to 0.5. The preset threshold and risk grading standard are specifically as follows: the preset intention-ability conflict matching degree threshold is 0.8, and the intention-ability conflict matching degree is set to M, where... The time was determined to be an unintentional - momentary conflict of ability. The moment was determined to be a transient conflict between intention and ability; the risk grading criteria combined the intensity of rehabilitation exercise intention (I), the comprehensive physical ability parameter (S), and the risk correction factor (k). R Delineation, specifically This is a conflict characterized by high intent, low capability, and high risk. The conflict is characterized by moderate intent, low capability, and medium risk. This was a time of extremely high-risk conflict.

[0015] Furthermore, the mathematical expression for the formula for calculating the rehabilitation intention-conflict matching degree is: Where M represents the rehabilitation intention-conflict matching degree; k R As a risk correction factor, it is derived from the pre-pain characteristic value P. p Calibration; T a α is the movement type adaptation coefficient, which is determined by the specific movement type of the rehabilitation exercise intention; I is the normalized intensity of the rehabilitation exercise intention; S is the comprehensive parameter of physical ability; α is the intention-ability deviation correction coefficient.

[0016] Furthermore, the multi-dimensional coupled adaptive parameter tuning model is a three-layer neural network structure, including an input layer, a hidden layer, and an output layer. The input layer receives six types of parameters: physiological fatigue level, pain quantification value, synchronous deviation value of movement rhythm of healthy and patient hands, intensity of rehabilitation exercise intention, comprehensive physical ability parameters, and matching degree. The hidden layer has 2-3 neuron layers, each containing 10-15 neurons, and uses the ReLU activation function. The output layer outputs the exoskeleton training parameter adjustment values. In the differentiated flexible resolution strategy, when there is a conflict between high intention, low ability, and high risk, the patient's intention... In the initial stage of the movement trajectory initiated by the figure, a guiding auxiliary force with an amplitude of 5°-15° and a strength of 3N-8N is applied for 0.8s-1.2s to safely guide the range of motion of the joint. Then, the assistance is increased by 5%-10% / 100ms step length, and the movement speed is increased by 3%-5% / 100ms step length. When there is a conflict between intention, low ability, and medium risk, the basic assistance is increased by 20%-30% and the movement speed is reduced by 15%-25%. When there is no conflict, the exoskeleton training parameters are adjusted according to normal rehabilitation standards, and the interval time is set to 10s-15s / set.

[0017] Furthermore, the adjustment coefficient, calculated using the multi-dimensional coupling training parameter adjustment coefficient formula, has a value range of [0.1, 1.0]. Based on the adjustment coefficient, specific values ​​for exoskeleton assistance, movement speed, training amplitude, interval time, assistance torque, and force triggering sequence are determined. Then, combined with the conflict determination result of step S3, a differentiated flexible resolution strategy is executed. If it is a high intention-low ability-high risk conflict, the exoskeleton is first controlled to perform a joint preparatory extension movement with an amplitude of 5°-15°, a force of 3N-8N, and a duration of 0.8s-1.2s. Then, the exoskeleton parameters are gradually adjusted to follow the patient's movement intention to complete the movement. If it is a medium intention-low ability-medium risk conflict, the basic assistance and movement speed of the exoskeleton are first adjusted, and then the movement is completed according to the patient's movement intention. If there is no conflict, the exoskeleton is directly controlled to complete the assisted training movement according to the parameters corresponding to the adjustment coefficient.

[0018] Furthermore, the mathematical expression for the multi-dimensional coupling training parameter adjustment coefficient formula is as follows: Where K is the adjustment coefficient of the multi-dimensional coupling training parameters; W M W S W F W P W D is the weight coefficient for each dimension, and the sum of all weight coefficients is 1; M is the calculated rehabilitation intention-conflict matching degree; S is the calculated comprehensive physical ability parameter; F is the normalized physiological fatigue degree; P is the normalized real-time pain quantification value, with a value range of [0,1]; D is the synchronous deviation value of the movement rhythm of the healthy and affected hands; D max The maximum rhythm synchronicity deviation calibrated for individualized patients.

[0019] Furthermore, the specific process of the full-process real-time closed-loop adjustment is as follows: Within each 500ms closed-loop cycle, the following operations are performed sequentially: 0-100ms: The multimodal sensing system synchronously acquires the patient's full-dimensional raw signals; 100-200ms: The acquired signals are denoised, normalized, feature extracted, and quantitatively analyzed to generate standardized core feature parameters; 200-350ms: The comprehensive physical ability parameters, intent-conflict matching degree, and multi-dimensional coupled training parameter adjustment coefficients are calculated respectively to complete the conflict determination; 350-450ms: The exoskeleton drive module receives adjustment parameters and strategy instructions, and synchronously executes parameter adjustment and flexible resolution strategies; 450-500ms: The exoskeleton parameter adjustment status, patient physiological status, and action status are monitored in real time, and the adjustment results and monitoring data are recorded to provide a basis for the next closed-loop cycle; the steady-state error of the closed-loop adjustment does not exceed 3%.

[0020] Compared with existing technologies, this adaptive training parameter adjustment method for a hand rehabilitation exoskeleton has the following beneficial effects:

[0021] I. This invention deploys a multimodal sensing system to simultaneously collect signals from all dimensions of the patient, including physiological fatigue, pain, movement rhythm, and rehabilitation intention. Through professional preprocessing and feature quantification analysis, a comprehensive and accurate patient state perception system is constructed. Relying on the quantitative calculation of comprehensive physical ability parameters and intention-conflict decoding logic, it achieves accurate determination of the matching state between intention and ability, clearly identifying various conflict types and risk levels. Based on a multi-dimensional coupled adaptive parameter tuning model, it deeply integrates multiple feature parameters to generate training parameters adapted to the patient's real-time state. Combined with differentiated and flexible resolution strategies, this allows exoskeleton training to both meet the patient's rehabilitation wishes and adapt to their physical endurance, effectively avoiding potential discomfort or risks during training. This significantly improves the safety and adaptability of rehabilitation training, providing patients with personalized rehabilitation support tailored to their individual needs.

[0022] Second, this invention constructs a closed-loop adjustment mechanism with a fixed cycle, cyclically executing the entire process of signal acquisition, processing, judgment, and parameter adjustment. This ensures that the exoskeleton training parameters can respond in real time to the dynamic changes in the patient's physiological state and motor ability, achieving continuous optimization of the training process. Combining scenario-based training with progressive training based on daily needs such as eating and dressing, it makes rehabilitation training more relevant to real-life applications, helping patients quickly transform training results into practical hand functions. Through the coupling and adjustment of multi-dimensional feature parameters and the precise execution of flexible mitigation strategies, it ensures the scientific nature and continuity of training while dynamically adjusting the intensity and method of training according to the patient's functional recovery progress. This fully mobilizes the patient's enthusiasm for rehabilitation, avoids ineffective or over-training, accelerates the patient's hand function recovery process, and improves the overall rehabilitation effect and efficiency.

[0023] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0025] Figure 1 A flowchart of an adaptive training parameter adjustment method for a hand rehabilitation exoskeleton;

[0026] Figure 2This is a schematic diagram illustrating the data transmission between steps in an adaptive training parameter adjustment method for a hand rehabilitation exoskeleton. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below. Example 1:

[0028] Example of adaptive training parameter adjustment for hand rehabilitation exoskeleton in a feeding scenario.

[0029] S1, Synchronous Acquisition of Multimodal Sensing Signals: A multimodal sensing system is built according to deployment requirements. Surface electromyography (EMG) sensors are attached to the core muscle groups of the patient's fingers, a heart rate variability sensor is worn on the wrist, thin-film pressure sensors are embedded in key joints of the exoskeleton, a facial expression recognition module is fixed in front of the patient's face, a flexible IMU wristband is worn on the healthy hand, and a motion sensing unit is integrated into the exoskeleton on the affected hand. A unified millisecond-level triggered signal acquisition protocol is configured, and raw signals of all dimensions are synchronously acquired according to this protocol, including EMG and heart rate variability signals representing physiological fatigue, joint impedance and facial micro-expression signals representing pain and pre-pain, movement trajectory and timing signals representing the rhythm of movement in both the healthy and affected hands, the healthy hand mainly performs eating-related actions such as grasping and moving utensils to bring food to the mouth, EMG intensity and trend signals representing the intention to perform rehabilitation exercises, the patient's intention to use the affected hand to grasp a spoon to scoop food, and joint force and impedance feedback signals representing physical capabilities. Simultaneously, the rhythmic features of daily movements related to eating with the healthy hand were extracted, and a rhythmic feature database of the healthy hand was built. This provides a reference benchmark for the rhythmic movements of the affected hand that fits the daily eating scenario, ensuring that the collected signals comprehensively cover the patient's physiological and motor state during the eating process. This lays a data foundation for subsequent parameter adjustments. Figure 1 As shown.

[0030] S2, Multi-source signal preprocessing and feature quantization: Targeted noise reduction is performed on the original signals across all dimensions. Signals related to physiological fatigue and rehabilitation exercise intention are filtered using a 5Hz-500Hz bandpass filter; signals related to pain and pre-pain are denoised using 3-5 layers of wavelet denoising; and signals related to the rhythm of movement in the healthy and affected hands are filtered using a sliding average filter with a sliding window size of 5-10 sampling points. This effectively filters out interference components in the signals, making the data more accurate. Linear normalization is used to uniformly map all denoised signals to the [0, 1] value range, giving different types of signals a unified comparison dimension. Subsequently, feature extraction and quantification are performed: Physiological fatigue is obtained by weighted fusion of the mean decay rate of the electromyographic signal MPF and the temporal features of heart rate variability; pain quantification and pre-pain feature values ​​are obtained by quantifying the peak value of human-computer interaction force and the amplitude change of micro-expression; the intensity of rehabilitation exercise intention is obtained by analyzing the peak value and rising slope of the electromyographic signal intensity; finally, the comprehensive physical ability parameter is calculated using the comprehensive physical ability parameter quantification formula. The mathematical expression of the comprehensive physical ability parameter quantification formula is: Where S is the comprehensive parameter of physical ability, and k Z W is the joint impedance correction factor. Z W F W Pp W Ff These are the weighting coefficients for each indicator. Z represents the real-time joint impedance value. max Z min The maximum and minimum joint resistance values ​​are individually calibrated for each patient; F represents the normalized physiological fatigue level; P... p F represents the normalized pre-pain characteristic value. f F is the real-time feedback value of limb force exertion. fmax The maximum value of limb force feedback is calibrated for individual patients, and multi-source signals are converted into standardized feature parameters that can be directly used for analysis, comprehensively reflecting the patient's physiological state, pain status, movement intention and physical support ability in the context of eating.

[0031] S3, Intent-Conflict Decoding and Type Determination: The comprehensive parameters of rehabilitation exercise intention intensity and physical ability are input into the rehabilitation intention-conflict decoding logic layer, which includes a signal fusion layer and a conflict determination layer. The signal fusion layer first performs data alignment and feature weighted fusion on the two parameters, with the fusion weights set between 0.4 and 0.5 to make the fusion of the two key parameters more balanced and reasonable. The intention-ability conflict matching degree M is obtained through the rehabilitation intention-conflict matching degree calculation formula. The mathematical expression of the rehabilitation intention-conflict matching degree calculation formula is: Where M represents the rehabilitation intention-conflict matching degree; k R As a risk correction factor, it is derived from the pre-pain characteristic value P. p Calibration; T aThe movement type adaptation coefficient is defined by the specific movement type of the rehabilitation exercise intention; I is the normalized intensity of the rehabilitation exercise intention; S is the comprehensive parameter of physical ability; α is the intention-ability deviation correction coefficient, which is determined based on a preset threshold of 0.8. M ≥ 0.8 indicates no momentary intention-ability conflict, and M < 0.8 indicates a conflict, clearly defining the adaptation between intention and ability. This is achieved by combining the intensity of the rehabilitation exercise intention I, the comprehensive parameter of physical ability S, and the risk correction factor k. R Delineate risk levels A conflict characterized by high intent, low capability, and high risk. For conflicts with moderate intent, low capability, and medium risk. For extremely high-risk conflicts, it is necessary to accurately distinguish different levels of conflict to provide a clear basis for the implementation of subsequent differentiated strategies.

[0032] S4, Multi-dimensional Coupling and Flexible Parametric Tuning: A three-layer neural network structure is constructed for multi-dimensional coupling and adaptive parametric tuning. The input layer receives six types of parameters: physiological fatigue level, pain quantification value, synchronous deviation value of healthy and affected hand movement rhythm, intensity of rehabilitation movement intention, comprehensive physical ability parameters, and matching degree. The hidden layer has 2-3 neurons per layer with 10-15 neurons, using the ReLU activation function. The output layer outputs the adjustment values ​​of exoskeleton training parameters, comprehensively integrating multi-dimensional features to achieve accurate parametric tuning calculation. The adjustment coefficient in the range of [0.1, 1.0] is calculated using the multi-dimensional coupling training parameter adjustment coefficient formula. The mathematical expression of the multi-dimensional coupling training parameter adjustment coefficient formula is as follows: Where K is the adjustment coefficient of the multi-dimensional coupling training parameters; W M W S W F W P W D is the weight coefficient for each dimension, and the sum of all weight coefficients is 1; M is the calculated rehabilitation intention-conflict matching degree; S is the calculated comprehensive physical ability parameter; F is the normalized physiological fatigue degree; P is the normalized real-time pain quantification value, with a value range of [0,1]; D is the synchronous deviation value of the movement rhythm of the healthy and affected hands; D maxThe maximum value of rhythmic synchronicity deviation is individually calibrated for each patient, and based on this, specific parameters such as the amount of exoskeleton assistance and movement speed are determined to ensure that the adjustment parameters match the patient's real-time state. A differentiated and flexible mitigation strategy is implemented: In cases of high-risk conflict, a preparatory joint extension movement with an amplitude of 5°-15°, a force of 3N-8N, and a duration of 0.8s-1.2s is performed first, followed by the application of the same specification of guiding assistance force at the beginning of the movement, and then increasing the assistance by 5%-10% / 100ms step length and increasing the speed by 3%-5% / 100ms step length, ensuring joint movement safety while gradually aligning with the patient's eating intentions; In cases of medium-risk conflict, the basic assistance is increased by 20%-30% and the movement speed is reduced by 15%-25%, and then the movement follows the patient's intended movement, balancing assistance support and movement safety; In cases of no conflict, training is conducted according to the adjustment coefficient parameters, with an interval of 10s-15s per set, maintaining a stable and efficient rehabilitation training rhythm for eating scenarios.

[0033] S5, Closed-Loop Adjustment and Advanced Training: The aforementioned steps are executed cyclically with a closed-loop adjustment cycle of 500ms. Within each cycle, raw signals are acquired from 0-100ms; signal preprocessing and feature quantization are completed from 100-200ms; core parameters are calculated and conflict determination is performed from 200-350ms; the exoskeleton drive module responds and adjusts parameters and strategies from 350-450ms; and state data is monitored and recorded from 450-500ms, providing a basis for the next round. The closed-loop steady-state error is ≤3%, ensuring real-time tracking of patient state changes and rapid response adjustments. The exoskeleton is driven to complete scenario-based coupling parameter tuning, making the training movements highly compatible with the eating scenario. Advanced rehabilitation training is implemented based on the recovery of the affected hand function. As the patient's ability to grasp and deliver food improves, training parameters are gradually optimized, continuously improving rehabilitation outcomes.

[0034] In summary, in the context of eating, the adaptive training parameter adjustment of the hand rehabilitation exoskeleton achieves precise rehabilitation through a complete process. First, a multimodal sensing system is deployed to simultaneously collect raw signals related to eating from all dimensions and build a feature library of the healthy hand's movement rhythm, providing a reference benchmark that fits daily life for training. After targeted noise reduction, normalization, and feature quantization, these signals are transformed into standardized parameters using a comprehensive physical ability parameter quantification formula. The type and risk level of conflict are determined through a rehabilitation intention-conflict decoding logic layer, and then adjustment coefficients are calculated based on a multi-dimensional coupled adaptive parameter tuning model to execute differentiated and flexible resolution strategies. Continuous iteration with a 500ms closed-loop cycle enables scenario-based coupled parameter tuning for eating, advancing training based on the recovery status of the affected hand, balancing safety and adaptability, and efficiently assisting in the recovery of eating-related functions of the affected hand. Example 2:

[0035] An example of adjusting adaptive training parameters for a hand rehabilitation exoskeleton in a washing and grooming scenario.

[0036] S1, Multimodal Sensor Signal Synchronous Acquisition: Deploy a multimodal sensing system as required. Surface electromyography (EMG) sensors are attached to the core muscle groups of the patient's fingers, a heart rate variability sensor is worn on the wrist, thin-film pressure sensors are embedded in key joints of the exoskeleton, a facial expression recognition module is fixed in front of the face, a flexible IMU wristband is worn on the healthy hand, and a motion sensing unit is integrated into the exoskeleton on the affected hand. A unified millisecond-level triggered signal acquisition protocol is configured to synchronously acquire full-dimensional raw signals: EMG and heart rate variability signals representing physiological fatigue; joint impedance and facial micro-expression signals representing pain and pre-pain; movement trajectory and timing signals representing the rhythm of movement in both the healthy and affected hands; actions such as grasping a towel, wiping the face, wringing out the towel, and turning on the faucet performed by the healthy hand; EMG intensity and trend signals representing the intention to perform rehabilitation exercises; the patient's intention to use the affected hand to perform actions such as brushing teeth, wiping the face, and operating the faucet; and joint force and impedance feedback signals representing physical capabilities. By extracting the rhythmic characteristics of the healthy hand's movements related to washing and grooming, and constructing a rhythmic characteristic database of the healthy hand, rehabilitation training for the affected hand can be tailored to the movement patterns of the washing and grooming scenario. This ensures that the collected signals fully capture the patient's physiological responses, motor state, and ability level during the washing and grooming process, providing comprehensive data support for subsequent precise adjustments. Figure 1 As shown.

[0037] S2, Multi-source signal preprocessing and feature quantization: The original signals are classified and denoised. Signals related to physiological fatigue and rehabilitation exercise intention are filtered using a 5Hz-500Hz bandpass filter. Signals related to pain and pre-pain are denoised using 3-5 layers of wavelet denoising. Signals related to the rhythm of movement in the healthy and affected hands are filtered using a sliding window with 5-10 sampling points of moving average filtering, effectively removing irrelevant interference and making the signals more reflective of the true state. The denoised signals are mapped to the [0,1] interval by linear normalization to achieve a unified quantization standard for different signal types. Then, feature extraction and quantization analysis are performed: Physiological fatigue is obtained by weighted fusion of the mean decay rate of the electromyographic signal MPF and the temporal features of heart rate variability. Pain quantification and pre-pain feature values ​​are obtained by quantifying the peak force of human-computer interaction and the amplitude change of micro-expression. The intensity of rehabilitation exercise intention is obtained by analyzing the peak intensity and rising slope of the electromyographic signal. The comprehensive physical ability parameter is calculated using the formula for quantifying the comprehensive physical ability parameter. The mathematical expression of the formula for quantifying the comprehensive physical ability parameter is: Where S is the comprehensive parameter of physical ability, and k Z W is the joint impedance correction factor. Z W F W Pp W Ff These are the weighting coefficients for each indicator. Z represents the real-time joint impedance value. max Z min The maximum and minimum joint resistance values ​​are individually calibrated for each patient; F represents the normalized physiological fatigue level; P... pF represents the normalized pre-pain characteristic value. f F is the real-time feedback value of limb force exertion. fmax The maximum value of limb force feedback is individually calibrated for each patient, and the raw signal is transformed into core features that can be used for analysis and decision-making, comprehensively presenting the patient's physiological fatigue level, pain status, intensity of movement intention, and physical support ability in the washing and grooming scenario.

[0038] S3, Intent-Conflict Decoding and Type Determination: The comprehensive parameters of rehabilitation exercise intention intensity and physical ability are input into the rehabilitation intention-conflict decoding logic layer. The signal fusion layer performs data alignment and feature weighted fusion (0.4-0.5) to make the combination of the two key parameters more scientific and reasonable. The conflict determination layer obtains the matching degree M through the rehabilitation intention-conflict matching degree calculation formula. The mathematical expression of the rehabilitation intention-conflict matching degree calculation formula is: Where M represents the rehabilitation intention-conflict matching degree; k R As a risk correction factor, it is derived from the pre-pain characteristic value P. p Calibration; T a The movement type adaptation coefficient is defined by the specific movement type of the rehabilitation exercise intention; I is the normalized intensity of the rehabilitation exercise intention; S is the comprehensive parameter of physical ability; α is the intention-ability deviation correction coefficient, which is determined by a preset threshold of 0.8. M ≥ 0.8 indicates no conflict, and M < 0.8 indicates a conflict, thus clarifying the adaptation relationship between intention and ability. Combining I, S, and k... R Delineate risk levels High-intention, low-capability, high-risk conflict; The conflict is characterized by moderate intent, low capability, and medium risk. To accurately identify conflict types with different risk levels for extremely high-risk conflicts, providing a clear direction for subsequent targeted resolution strategies.

[0039] S4, Multi-dimensional Coupling and Flexible Parametric Adjustment: A three-layer neural network structure is constructed using a multi-dimensional coupling adaptive parameter tuning model. The input layer receives six types of parameters: physiological fatigue level, pain quantification value, synchronous deviation value of the movement rhythm of the healthy and affected hands, intensity of rehabilitation exercise intention, comprehensive physical ability parameters, and matching degree. The hidden layer has 2-3 neurons per layer with 10-15 neurons each, using the ReLU activation function. The output layer outputs the adjustment values ​​of the exoskeleton training parameters. By integrating multi-dimensional information, the training parameters can be accurately calculated. The adjustment coefficient [0.1, 1.0] is calculated using the multi-dimensional coupling training parameter adjustment coefficient formula. The mathematical expression of the multi-dimensional coupling training parameter adjustment coefficient formula is as follows: Where K is the adjustment coefficient of the multi-dimensional coupling training parameters; W M W S W F W P W Dis the weight coefficient for each dimension, and the sum of all weight coefficients is 1; M is the calculated rehabilitation intention-conflict matching degree; S is the calculated comprehensive physical ability parameter; F is the normalized physiological fatigue degree; P is the normalized real-time pain quantification value, with a value range of [0,1]; D is the synchronous deviation value of the movement rhythm of the healthy and affected hands; D max To determine the maximum value of rhythmic synchronicity deviation for each patient, specific parameters such as exoskeleton assistance and speed are established to ensure that the parameter settings match the patient's real-time state and the needs of the washing and grooming scenario. A differentiated strategy is implemented: In cases of high-risk conflict, a preparatory joint extension with an amplitude of 5°-15°, a force of 3N-8N, and a duration of 0.8s-1.2s is performed first, followed by the same guiding assistance force at the beginning of the washing and grooming movement. Subsequently, assistance and speed are increased according to the corresponding step length, balancing joint safety and the completion of the washing and grooming movement. In cases of medium-risk conflict, basic assistance is increased by 20%-30% and movement speed is reduced by 15%-25%, then the movement follows the patient's intended movement, balancing assistance support and movement stability. In cases without conflict, training is conducted according to the adjustment coefficient parameters, with an interval of 10s-15s per set, maintaining a stable and efficient rehabilitation training process in the washing and grooming scenario.

[0040] S5, Closed-Loop Adjustment and Advanced Training: The aforementioned steps are executed cyclically with a closed-loop cycle of 500ms. Within each cycle, signals are acquired from 0-100ms; preprocessing and feature quantization are performed from 100-200ms; core parameters are calculated and conflicts are identified from 200-350ms; the exoskeleton responds to adjustment commands from 350-450ms; and steady-state error of data is monitored and recorded from 450-500ms, ensuring real-time capture of changes in the patient's state and rapid adjustment of training parameters. The exoskeleton is driven to complete scenario-based coupling parameter tuning for washing, allowing training movements to deeply adapt to the washing scenario. Based on the functional recovery of the affected hand's washing movements, progressive rehabilitation training is gradually advanced. As the affected hand's ability to brush teeth, wipe the face, and operate the faucet improves, the training program is continuously optimized, steadily improving the rehabilitation effect.

[0041] In summary, the adaptive training parameter adjustment for the washing and grooming scenario is structured around the entire process based on the scenario's requirements. A multimodal sensing system comprehensively captures physiological, motor, and intention-related signals from the patient during the washing and grooming process, extracting the rhythmic features of the healthy hand's washing and grooming movements to form a reference. After classification, noise reduction, normalization, and feature quantification analysis, core parameters are generated using a comprehensive physical ability parameter quantification formula. The rehabilitation intention-conflict matching degree calculation formula and risk grading standards are used to accurately determine conflicts, and then a multi-dimensional coupled adaptive parameter tuning model outputs adjustment parameters to execute corresponding flexible resolution strategies. Real-time adjustments are made with a 500ms closed-loop cycle to achieve deep adaptation between the washing and grooming scenario and the training parameters. Through advanced training, the patient's hand's washing and grooming-related motor abilities are steadily improved, ensuring the accuracy and efficiency of rehabilitation training.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for adjusting adaptive training parameters of a hand rehabilitation exoskeleton, characterized in that, The specific steps of this method are as follows: S1, Synchronous acquisition of multimodal sensor signals: Deploy a multimodal sensor system, configure a unified signal acquisition protocol, and synchronously acquire the full-dimensional raw signals of the patient's physiological fatigue, pain, pre-pain, rhythm of movement of the healthy and affected hands, intention of rehabilitation exercises, and physical abilities. Extract the rhythmic features of the healthy hand's daily movements to build a rhythmic feature library of the healthy hand. S2, Multi-source signal preprocessing and feature quantization: Noise reduction, normalization and feature extraction and quantization analysis of the original signals in all dimensions are performed to obtain physiological fatigue degree, pain quantification value, pre-pain feature value, synchronous deviation value of the movement rhythm of healthy and affected hands, and intensity of rehabilitation movement intention. The comprehensive physical ability parameter is calculated by the comprehensive physical ability parameter quantification formula. S3, Intention-Conflict Decoding and Type Determination: Input the comprehensive parameters of rehabilitation exercise intention intensity and physical ability into the rehabilitation intention-conflict decoding logic layer, obtain the matching degree through the rehabilitation intention-conflict matching degree calculation formula, and determine whether there is an instantaneous intention-ability conflict and the specific conflict type based on the preset threshold and risk classification standard. S4, Multi-dimensional Coupling and Flexible Resolution: Construct a multi-dimensional coupling adaptive parameter tuning model, calculate the adjustment coefficient through the multi-dimensional coupling training parameter adjustment formula, complete the calculation of exoskeleton training parameters, and execute a differentiated flexible resolution strategy based on the conflict determination results; S5, Closed-loop adjustment and advanced training: Set a closed-loop adjustment cycle of 500ms, and execute the above steps in a loop to drive the exoskeleton drive module to respond synchronously to the adjustment parameters, complete the scenario-based coupling parameter adjustment of eating, dressing, writing and washing scenarios, and implement advanced rehabilitation training in combination with the recovery of the affected hand function.

2. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S1, the deployment of the multimodal sensing system specifically involves: attaching surface electromyography (EMG) sensors to the core muscle groups of the patient's fingers, wearing a heart rate variability sensor on the wrist, embedding a thin-film pressure sensor in the key joints of the exoskeleton, fixing a facial expression recognition module in front of the patient's face, wearing a flexible IMU wristband on the healthy hand, and integrating a motion sensing unit into the exoskeleton on the affected hand side. A unified millisecond-level triggered signal acquisition protocol is configured, and the EMG and heart rate variability signals representing the patient's physiological fatigue, joint impedance and facial micro-expression signals representing pain and pre-pain, movement trajectory and timing signals representing the rhythm of movement in both the healthy and affected hands, EMG intensity and trend signals representing the intention of rehabilitation exercises, and joint force and impedance feedback signals representing physical capabilities are acquired synchronously according to the signal acquisition protocol.

3. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S2, the noise reduction process employs three methods: adaptive filtering, moving average filtering, and wavelet denoising, which are adapted to different original signals. Specifically, signals related to physiological fatigue and rehabilitation exercise intentions are filtered using a 5Hz-500Hz bandpass filter, signals related to pain and pre-pain are denoised using wavelet denoising with a decomposition layer of 3-5 layers, and signals related to the rhythm of healthy and affected hands are filtered using a moving average filter with a sliding window size of 5-10 sampling points. Normalization was performed using a linear normalization method, mapping all denoised signals to the [0, 1] value range. During feature extraction and quantification, physiological fatigue was obtained by weighted fusion of the mean decay rate of electromyographic signal MPF and the temporal features of heart rate variability. Pain quantification and pre-pain feature values ​​were obtained by quantification of the peak value of human-computer interaction force and the change features of micro-expression amplitude. The intensity of rehabilitation exercise intention was obtained by analyzing the peak value and rising slope of electromyographic signal intensity.

4. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S2, the mathematical expression of the formula for quantifying the comprehensive physical ability parameters is: Where S is the comprehensive parameter of physical ability, and k Z W is the joint impedance correction factor. Z W F W Pp W Ff These are the weighting coefficients for each indicator. Z represents the real-time joint impedance value. max Z min The maximum and minimum joint resistance values ​​are individually calibrated for each patient; F represents the normalized physiological fatigue level; P... p F represents the normalized pre-pain characteristic value. f F is the real-time feedback value of limb force exertion. fmax The maximum value of limb force feedback is individually calibrated for each patient.

5. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S3, the rehabilitation intention-conflict decoding logic layer is divided into a signal fusion layer and a conflict determination layer. The signal fusion layer performs data alignment and feature weighted fusion on the comprehensive parameters of rehabilitation exercise intention intensity and physical ability, with both fusion weights being 0.4-0.

5. The preset threshold and risk grading standard are as follows: the preset intention-ability conflict matching degree threshold is 0.8, and the intention-ability conflict matching degree is set to M, where... The time was determined to be an unintentional - momentary conflict of ability. The time frame indicates a momentary conflict between intent and capability. The risk grading criteria combine the intensity of rehabilitation exercise intention (I), the comprehensive physical ability parameter (S), and the risk correction factor (k). R Delineation, specifically This is a conflict characterized by high intent, low capability, and high risk. The conflict is characterized by moderate intent, low capability, and medium risk. This was a time of extremely high-risk conflict.

6. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 5, characterized in that, In step S3, the mathematical expression for the rehabilitation intention-conflict matching degree calculation formula is: ;in, M represents the degree of match between rehabilitation intention and conflict; k R As a risk correction factor, it is derived from the pre-pain characteristic value P. p Calibration; T a α is the movement type adaptation coefficient, which is determined by the specific movement type of the rehabilitation exercise intention; I is the normalized intensity of the rehabilitation exercise intention; S is the comprehensive parameter of physical ability; α is the intention-ability deviation correction coefficient.

7. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S4, the multi-dimensional coupled adaptive parameter tuning model is a three-layer neural network structure, including an input layer, a hidden layer, and an output layer. The input layer receives six types of parameters: physiological fatigue level, pain quantification value, synchronous deviation value of movement rhythm of healthy and patient hands, intensity of rehabilitation exercise intention, comprehensive parameters of physical ability, and matching degree. The hidden layer has 2-3 neuron layers, each containing 10-15 neurons, and uses the ReLU activation function. The output layer outputs the exoskeleton training parameter adjustment values. In the differentiated flexible resolution strategy, when there is a conflict between high intention, low ability, and high risk, the patient's intention... In the initial stage of the movement trajectory initiated by the figure, a guiding auxiliary force with an amplitude of 5°-15° and a strength of 3N-8N is applied for 0.8s-1.2s to safely guide the range of motion of the joint. Then, the assistance is increased by 5%-10% / 100ms step length, and the movement speed is increased by 3%-5% / 100ms step length. When there is a conflict between intention, low ability, and medium risk, the basic assistance is increased by 20%-30% and the movement speed is reduced by 15%-25%. When there is no conflict, the exoskeleton training parameters are adjusted according to normal rehabilitation standards, and the interval time is set to 10s-15s / set.

8. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S4, the adjustment coefficient, with a value range of [0.1, 1.0], is calculated using the multi-dimensional coupling training parameter adjustment coefficient formula. Based on the adjustment coefficient, the specific values ​​of exoskeleton assistance, movement speed, training amplitude, interval time, assistance torque, and force triggering sequence are determined. Then, combined with the conflict determination result of step S3, a differentiated flexible resolution strategy is executed. If it is a high intention-low ability-high risk conflict, the exoskeleton is first controlled to perform a joint preparatory extension movement with an amplitude of 5°-15°, a force of 3N-8N, and a duration of 0.8s-1.2s. Then, the exoskeleton parameters are gradually adjusted to follow the patient's movement intention to complete the movement. If it is a medium intention-low ability-medium risk conflict, the basic assistance and movement speed of the exoskeleton are first adjusted. Then, the movement is completed according to the patient's movement intention. If there is no conflict, the exoskeleton is directly controlled to complete the assisted training movement according to the parameters corresponding to the adjustment coefficient.

9. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S4, the mathematical expression for the multi-dimensional coupling training parameter adjustment coefficient formula is: Where K is the adjustment coefficient of the multi-dimensional coupling training parameters; W M W S W F W P W D represents the weight coefficients for each dimension, and the sum of all weight coefficients is 1; M is the calculated rehabilitation intention-conflict matching degree. S represents the calculated comprehensive physical capabilities parameter; F represents the normalized physiological fatigue level; P represents the normalized real-time pain quantification value, ranging from [0,1]; D represents the synchronous deviation value of the movement rhythm of the healthy and affected hands; D max The maximum rhythm synchronicity deviation calibrated for individualized patients.

10. The adaptive training parameter adjustment method for a hand rehabilitation exoskeleton according to claim 1, characterized in that, In step S5, the specific process of the real-time closed-loop adjustment of the entire process is as follows: the following operations are performed sequentially within each 500ms closed-loop cycle: 0-100ms, the multimodal sensing system synchronously collects the patient's full-dimensional raw signals; 100-200ms, the collected signals are denoised, normalized, feature extracted and quantized to generate standardized core feature parameters. From 200 to 350 ms, the comprehensive parameters of physical ability, intention-conflict matching degree, and adjustment coefficients of multi-dimensional coupled training parameters are calculated to complete the conflict determination; from 350 to 450 ms, the exoskeleton drive module receives adjustment parameters and strategy instructions, and executes parameter adjustment and flexible resolution strategies simultaneously; from 450 to 500 ms, the exoskeleton parameter adjustment status, patient physiological status, and movement status are monitored in real time, and the adjustment results and monitoring data are recorded to provide a basis for the next closed-loop cycle; the steady-state error of closed-loop adjustment does not exceed 3%.

Images