Elbow joint intelligent rehabilitation evaluation and training method based on multi-source sensing information fusion

Abstract

This invention discloses an intelligent rehabilitation assessment and training method for the elbow joint based on multi-source sensor information fusion, belonging to the field of intelligent elbow joint rehabilitation technology. The method includes: acquiring the patient's physiological and kinematic information during elbow joint rehabilitation training; constructing a two-layer assessment system of "macro-staging – micro-quantification" through multi-source information fusion; inputting muscle synergy index, motion smoothness characteristics, and interaction stiffness characteristics into a confidence rule base model; outputting the patient's current Brunnstrom rehabilitation stage; and then calculating a comprehensive rehabilitation progress score to sensitively reflect even minor progress in the patient's rehabilitation. Furthermore, based on the output Brunnstrom rehabilitation stage, the corresponding control mode is dynamically invoked, including a passive rehabilitation training mode based on S-shaped velocity planning, an assisted rehabilitation training mode based on nonlinear intention mapping, and an adaptive impedance training mode based on virtual trajectory prediction, to improve the safety, effectiveness, and scientific rigor of elbow joint rehabilitation training.

Description

Technical Field

[0001] This invention relates to the field of intelligent rehabilitation technology for elbow joints, and in particular to an intelligent rehabilitation assessment and training method for elbow joints based on multi-source sensor information fusion. Background Technology

[0002] As one of the most important joints in the upper limb, the elbow joint's flexion and extension capabilities significantly impact daily living activities. Patients with stroke, spinal cord injury, or orthopedic surgery often experience varying degrees of elbow joint dysfunction, requiring long-term, repetitive rehabilitation training using rehabilitation equipment.

[0003] Existing elbow joint rehabilitation devices mostly rely on single sensor information for feedback, such as acquiring joint movement status solely through angle encoders or force sensors. These devices struggle to accurately reflect a patient's active movement intentions and actual motor abilities. During rehabilitation training, they often cannot distinguish whether the patient is actively participating in the movement or passively performing actions guided by the device, making it difficult to personalize training strategies. Furthermore, some rehabilitation systems incorporate electromyography (EMG) signals to characterize muscle activation; however, using EMG signals alone is susceptible to noise, individual differences, and muscle fatigue, resulting in insufficient stability. Conversely, relying solely on force or angle information fails to reflect the recovery of neuromuscular control.

[0004] In terms of joint dynamics, current technologies for estimating elbow joint torque largely rely on simplified models or empirical formulas, failing to fully integrate multi-source physiological and kinematic information, resulting in limited prediction accuracy. Meanwhile, rehabilitation assessments often employ subjective scales or simple threshold methods, lacking continuous, quantitative, and objective assessment tools. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent rehabilitation assessment and training method for elbow joint based on multi-source sensor information fusion. The method integrates multi-source sensor information to predict elbow joint movement angle and joint torque, and performs quantitative assessment of the rehabilitation process, so as to improve the safety, effectiveness and scientific nature of elbow joint rehabilitation training.

[0006] In a first aspect, the present invention provides an intelligent rehabilitation assessment method for the elbow joint based on multi-source sensor information fusion, comprising: Multi-source sensor information acquisition and processing: Acquire physiological and motor information of the patient during elbow joint rehabilitation training, including surface electromyography signals, inertial posture signals, and human-computer interaction force signals; The surface electromyography signal matrix is ​​decomposed into nonnegative matrix to extract muscle coordination vectors and time activation coefficients; the similarity between the coordination vector of each muscle in the patient and the coordination vector of the muscles in the healthy person is calculated; based on the cumulative activation energy of each muscle coordination during the entire movement process, the corresponding coordination activation weight coefficient is calculated; the coordination vector similarity and coordination activation weight coefficient are fused to obtain the neural control coordination index. By performing dimensionless processing on the accelerometer signal of the elbow joint angular acceleration, motion smoothness features are constructed. Calculate the stiffness characteristics of human-computer interaction based on human-computer interaction force signals and elbow joint angle changes; Macro-stage process: The muscle synergy index, motion smoothness feature and interaction stiffness feature are input into the confidence rule base model. The confidence rule base model integrates prior knowledge and quantitative feature data to output the patient's current Brunnstrom rehabilitation stage. Micro-quantification process: Calculate the comprehensive score of rehabilitation progress, which is constructed based on the continuous numerical value obtained by weighted Euclidean distance calculation of multi-source features relative to the ideal value of the current stage.

[0007] Secondly, this invention provides an intelligent rehabilitation training method for the elbow joint based on multi-source sensor information fusion, which dynamically calls the corresponding control mode according to the Brunnstrom rehabilitation stage output in the first aspect: When Brunnstrom Phase 1-2 is output, the passive rehabilitation training mode based on S-shaped velocity planning is invoked. The passive rehabilitation training mode uses a fifth-order polynomial S-shaped velocity curve to construct the target motion trajectory of the elbow joint, driving the rehabilitation actuator to move the affected limb to perform passive flexion and extension training. When Brunnstrom Phase 3-4 is output, the assisted rehabilitation training mode based on nonlinear intention mapping is invoked. The assisted rehabilitation training mode adopts a three-level cascaded control architecture of "intention prediction - nonlinear mapping - closed-loop force control", which enables the rehabilitation equipment to dynamically adjust the amount of assistance according to the changes in the patient's active movement ability. When Brunnstrom stages 5-6 are output, the adaptive impedance training mode based on virtual trajectory prediction is invoked. This adaptive impedance training mode uses a deep learning model to mine the correlation between the temporal features of electromyography signals and the current motion state, and predicts the joint angles and angular velocities in real time. The predicted joint angles and angular velocities are used as "virtual reference trajectories" to achieve follow-up impedance training. Then, based on the deviation between the "virtual reference trajectory" and the actual motion state, the compliant impedance torque that the rehabilitation equipment should apply is calculated, and an resistance adjustment factor is introduced to fine-tune the intensity of the impedance training.

[0008] The technical solutions provided in the embodiments of the present invention have at least the following technical effects: 1. Breaking through the limitations of traditional electromyography (EMG) assessments that only focus on "intensity," this paper proposes a neuromuscular control assessment method based on muscle synergy. By introducing nonnegative matrix factorization (NMF) technology, muscle synergy patterns are decoupled from cluttered EMG signals. Through similarity comparison with healthy modalities, a quantitative evaluation of neuromuscular control coordination in stroke patients is achieved. This allows the system to not only determine whether a patient "has strength" but also "knows how to exert force," providing a deeper physiological basis for correcting abnormal synergy patterns.

[0009] 2. A two-tiered, refined assessment system combining qualitative staging and quantitative scoring has been constructed, addressing the issue of insufficient sensitivity in assessments using single scales. Existing technologies often only output Brunnstrom staging, which is insufficient to reflect short-term training effects. This invention combines the macroscopic staging capabilities of the BRB model with microscopic quantitative scoring based on motion smoothness. This not only guides the switching of major training modes but also captures subtle progress made by patients within the same stage through continuous scoring, enhancing patients' confidence in rehabilitation and providing therapists with precise references for adjusting training details.

[0010] 3. This invention achieves "knowledge-guided online evolution" of the assessment model, overcoming the modeling challenges of scarce rehabilitation medical data and significant individual differences. The invention employs a Confidence Rule Base (BRB) as the core inference engine. Its mechanism of "expert knowledge initialization + online correction based on individual data" effectively avoids the dependence on large datasets and the "black box" uninterpretability of pure neural network models. The model can automatically calibrate rule weights as patient training data accumulates, enabling the assessment system to adaptively evolve from "general" to "personalized."

[0011] 4. An assistive control algorithm based on intention prediction and nonlinear safety adjustment was developed, solving the problems of "zero-point jitter" and human-machine aggression in traditional electromyography (EMG) control. In assisted rehabilitation mode, this invention abandons the simple threshold triggering mechanism and instead uses an offline-trained joint torque estimation model to map the patient's active movement intention in real time. In particular, this invention designs a nonlinear safety control module, which effectively suppresses noise interference from EMG signals at rest or during weak contractions by setting dynamic dead zones and saturation thresholds, eliminating abnormal jitter of the actuator. Combined with a torque tracking closed loop based on PD control, the system can output smooth and stable assistive torque according to the patient's real-time weak intentions, significantly enhancing the coordination of human-machine interaction and allowing the patient to maintain a sense of active participation while receiving assistance.

[0012] 5. This invention achieves an active resistance training mode that "de-presets the trajectory," enhancing the compliance and immersion of rehabilitation training through deep temporal prediction and dual-loop control. Targeting the advanced training needs of the later stages of rehabilitation (Brunnstrom stages 5-6), this invention changes the traditional rigid pattern of "movement along a fixed trajectory." It utilizes a Long Short-Term Memory (LSTM) network to deeply mine the temporal characteristics of multi-source signals, predicting the patient's desired joint angles and angular velocities in real time, thus replacing the preset trajectory. Combining a dual closed-loop structure of "outer loop impedance model + inner loop force control," and introducing an resistance adjustment factor in the outer loop, the system not only provides a compliant resistance sensation but also steplessly adjusts the resistance intensity according to muscle strength recovery. This design greatly replicates the real limb movement environment, stimulates the patient's active movement potential, and effectively promotes the reshaping of neuromuscular control circuits.

[0013] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Figure 1 This is a flowchart illustrating the overall evaluation method in Embodiment 1 of the present invention; Figure 2 This is a detailed flowchart of the evaluation method in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of rehabilitation assessment and training equipment; Figure 4 This is a schematic diagram illustrating the working principle of the rehabilitation training mode in the training method of Embodiment 2 of the present invention. Figure 5 This is a schematic diagram illustrating the working principle of the adaptive impedance training mode in the training method of Embodiment 2 of the present invention. Detailed Implementation

[0016] This invention provides an intelligent rehabilitation assessment and training method for the elbow joint based on multi-source sensor information fusion. By integrating multi-source sensor information, the method can predict the elbow joint motion angle and joint torque, and quantitatively assess the rehabilitation process, thereby improving the safety, effectiveness, and scientific nature of elbow joint rehabilitation training.

[0017] Before introducing specific embodiments, let's first introduce one type of rehabilitation assessment and training device corresponding to the method of this embodiment, such as... Figure 3As shown, the device includes at least a surface electromyography (SEMG) signal acquisition module, an inertial sensor acquisition module, a pressure sensor acquisition module, a control board, a hand support base, and a transmission device. The SEMG signal acquisition module is positioned on the surface of the relevant muscles involved in elbow flexion and extension movements to acquire SEMG signals reflecting the patient's muscle activation state. The inertial sensor acquisition module includes an accelerometer and a gyroscope, fixed to the end of the rehabilitation actuator, and is used to acquire information on posture changes during elbow joint movement, including angle changes, angular velocity, and angular acceleration. The pressure sensor acquisition module is located at the human-machine interface between the patient's wrist and the rehabilitation actuator, and is used to acquire interactive force information between the patient and the rehabilitation actuator to reflect the force experienced by the patient during rehabilitation training.

[0018] Example 1 This embodiment provides an intelligent rehabilitation assessment method for the elbow joint based on multi-source sensor information fusion, such as... Figure 1 and Figure 2 As shown, it includes: S1. Multi-source sensor information acquisition and processing: Acquire physiological and kinematic information of the patient during elbow joint rehabilitation training, including surface electromyography signals, inertial posture signals (including elbow joint angular acceleration signals and elbow joint angle changes), and human-computer interaction force signals; filter the acquired signals to remove environmental noise and high-frequency interference; detect and remove outliers, and normalize the processed signals to reduce the impact of differences in the dimensions of different sensor signals on subsequent analysis. In one possible implementation, the information processing process includes: S1-1, Perform nonnegative matrix decomposition on the surface electromyography signal matrix. Muscle co-vectors were extracted. With time activation coefficient ( h (representing the number of cooperative vectors); The extracted muscle coordination vectors from the patient are matched against pre-established muscle coordination norms from healthy individuals. Specifically, the cosine similarity between the coordination vector of each muscle in the patient and the coordination vector of the muscles in the healthy individual is calculated using the following formula: ; in, Indicates the patient's first k One collaborative vector, The first number corresponding to a healthy person k A number of collaborative vectors.

[0019] Based on the cumulative activation energy of each muscle synergy throughout the entire exercise process, a corresponding synergistic activation weight coefficient is calculated, so that muscle synergies with higher activation energy have a higher weight in the overall evaluation. The formula for calculating the weight is as follows: ; in, The time activation coefficient represents the activation coefficient at the th... k A collaborative mode in t Activation intensity at any given moment Indicates the first i Each co-vector in t Activation intensity at any given moment N Represents the number of times. h This represents the number of cooperative vectors.

[0020] By weighting and fusing the various synergistic similarities according to the aforementioned activation weights, a neural control synergy index is constructed to quantify the degree of improvement in the patient's muscle group synergistic control ability and abnormal synergistic patterns during elbow joint movement. The formula for calculating the neural control synergy index is as follows: ; S1-2. Statistical feature extraction is performed on pressure sensor signals and inertial sensor signals to characterize the changes in force and kinematic properties during elbow joint rehabilitation training. Specifically, the jerk is dimensionlessly processed based on the elbow joint angular acceleration signal acquired by the inertial sensor. ; in, This indicates the acceleration value at that moment. This represents the acceleration value at the previous moment. This represents the time difference between two moments when calculating jerk; A motion smoothness feature is constructed based on the dimensionless jerk, which is used to quantitatively characterize the fineness of the patient's elbow joint motion control; the formula for the motion smoothness feature is as follows: ; in, T Indicates the cycle of motion. Indicates the range of motion angles within the motion cycle. M Indicates the number of sampling points within the motion cycle. It indicates jerk.

[0021] S1-3. Calculate the global equivalent human-computer interaction stiffness characteristics based on the human-computer interaction force signal and the change in elbow joint angle. This reflects the change in interaction force corresponding to a unit joint displacement and is used to quantitatively assess the degree of limb spasticity and passive compliance level in patients. The calculation formula for the interaction stiffness characteristics is as follows: ; in, This represents the difference between the maximum and minimum values ​​of the human-computer interaction force measured by the pressure sensor during the motion cycle; It indicates the range of motion angles within the motion cycle.

[0022] Subsequently, through the fusion of multi-source information, a two-layer assessment system of "macro-staging – micro-quantification" is constructed: This invention abandons the single classification assessment model and adopts a two-layer architecture to conduct a three-dimensional evaluation of rehabilitation effects.

[0023] S2. Macro-stage process: The muscle synergy index, motion smoothness features and interaction stiffness features are input into the confidence rule base model. The confidence rule base (BRB) model integrates prior knowledge and quantitative feature data to output the patient's current Brunnstrom rehabilitation stage (stages 1-6), which is used to guide the switching of major rehabilitation modes (passive / assisted / resistance).

[0024] S3. Micro-quantification process: Based on the determination of the rehabilitation stage, the Rehabilitation Progress Score (RPS) is further calculated. The RPS is constructed based on the weighted Euclidean distance of multi-source features relative to the ideal value of the current stage. It is used to sensitively reflect the patient's small progress in muscle endurance, motor fluency and neural control within the same rehabilitation stage.

[0025] In one possible implementation, the formula for calculating the comprehensive score of the rehabilitation process is as follows: ; in, Indicates the first i The weights of each feature and ; m Indicates the number of features. Indicates the first i The actual value of each feature, Indicates the first i The ideal value of each feature and The values ​​are set separately for different stages of rehabilitation, and their ranges are determined by clinical rehabilitation experts based on long-term rehabilitation practice experience and statistical data of patients at different stages. This represents the scoring sensitivity adjustment factor.

[0026] Preferably, individual adaptation of the assessment model can be achieved through a parameter optimization mechanism: given the small sample size of rehabilitation data, the training of the BRB model adopts a "knowledge-data dual-driven" strategy. Initial model parameters are derived from clinical rehabilitation guidelines. During rehabilitation training, a covariance matrix adaptive evolution strategy (CMA-ES) is used to update the rule weights and confidence levels of the confidence rule base model online based on newly generated training data from patients. This mechanism enables the assessment model to gradually adapt to and "learn" the specific movement patterns of individual patients as training progresses, thereby improving the accuracy of assessment results at the individual level.

[0027] Example 2 This embodiment provides an intelligent elbow joint rehabilitation training method based on multi-source sensor information fusion. Targeting the characteristics of the entire elbow joint rehabilitation cycle, it proposes a multi-mode closed-loop control strategy based on adaptive rehabilitation stages. According to the rehabilitation stage (Brunnstrom staging) identified in Embodiment 1, the corresponding control mode is dynamically invoked. When Brunnstrom Phase 1-2 is output, the passive rehabilitation training mode based on S-shaped velocity planning is invoked. The passive rehabilitation training mode uses a fifth-order polynomial S-shaped velocity curve to construct the target motion trajectory of the elbow joint, driving the rehabilitation actuator to move the affected limb to perform passive flexion and extension training. When Brunnstrom Phase 3-4 is output, the assisted rehabilitation training mode based on nonlinear intention mapping is invoked. The assisted rehabilitation training mode adopts a three-level cascaded control architecture of "intention prediction - nonlinear mapping - closed-loop force control", which enables the rehabilitation equipment to dynamically adjust the amount of assistance according to the changes in the patient's active movement ability. When Brunnstrom stages 5-6 are output, the adaptive impedance training mode based on virtual trajectory prediction is invoked. This adaptive impedance training mode uses a deep learning model to mine the correlation between the temporal features of electromyography signals and the current motion state, and predicts the joint angles and angular velocities in real time. The predicted joint angles and angular velocities are used as "virtual reference trajectories" to achieve follow-up impedance training. Then, based on the deviation between the "virtual reference trajectory" and the actual motion state, the compliant impedance torque that the rehabilitation equipment should apply is calculated, and an resistance adjustment factor is introduced to fine-tune the intensity of the impedance training.

[0028] 1. Passive rehabilitation training model based on S-curve velocity planning: This model addresses the abnormal muscle tone and spasticity characteristics of Brunnstrom stage 1-2 patients. It abandons the traditional trapezoidal velocity planning and uses a fifth-order polynomial S-curve velocity curve to construct the target motion trajectory of the elbow joint; specifically including: First, set the starting angle, ending angle, maximum angular velocity, and maximum angular acceleration; then use fifth-order polynomial interpolation to generate a smooth position command sequence with continuously changing angular acceleration. The generated smooth position command sequence is differentially compared with the actual angle fed back by the inertial sensor and input to the position controller (PID) to drive the rehabilitation actuator to perform ultra-compliant passive flexion and extension training of the affected limb. This trajectory ensures that the acceleration at the beginning and end of the movement is zero, avoiding limb spasticity induced by the shock of motor start-stop.

[0029] 2. Assisted Rehabilitation Training Mode Based on Nonlinear Intent Mapping: This mode aims to address the problems of control instability and "zero-point jitter" under weak electromyographic signals, employing a three-level cascaded control architecture of "intent prediction – nonlinear mapping – closed-loop force control"; such as... Figure 4As shown, the rehabilitation training mode specifically includes: Surface electromyography (EMG) signals and inertial sensing signals are collected, input into a pre-trained joint torque estimation model, and the predicted torque is output. To eliminate static noise and achieve smooth assist, a nonlinear mapping function incorporating dynamic dead zone and saturation characteristics is designed to shape the original predicted torque. Specifically, a two-sided dead zone threshold is set; when the absolute value of the predicted torque T is less than a first threshold (T... min When the output target auxiliary torque is forced to zero (to eliminate jitter), the absolute value of the predicted torque is greater than the first threshold (T). min And less than the second threshold (T) max When the predicted torque is greater than the second threshold (T), an exponential gain scheduling function is used to calculate the target assist torque, so that the magnitude of the assistance increases smoothly and non-linearly with the increase of the patient's effort intention; when the absolute value of the predicted torque is greater than the second threshold (T), max When the target auxiliary torque is saturated and truncated at the safety limit, the target auxiliary torque is truncated.

[0030] The shaped target auxiliary torque is compared with the human-machine interaction force measured by the pressure sensor, and the motor output is adjusted by the PD torque controller. The control rate is as follows: ; in, For proportional gain, For differential gain, To predict torque, For interactive force, The arm length.

[0031] Through the above control process, the rehabilitation equipment can dynamically adjust the level of assistance according to changes in the patient's active movement ability, thereby achieving human-machine collaborative assisted rehabilitation training.

[0032] 3. Adaptive impedance training mode based on virtual trajectory prediction: In view of the characteristics of patients at this stage who need to enhance muscle strength and nerve control ability, an active rehabilitation strategy without preset trajectory is proposed. Deep learning is used to generate "virtual guide points" to realize dynamic impedance training, thereby overcoming the defects of traditional impedance control that rely on preset fixed trajectories.

[0033] This embodiment utilizes Long Short-Term Memory (LSTM) networks to deeply mine the correlation between the temporal characteristics of electromyographic signals and the current motor state. The model predicts joint angles and angular velocities in real time, using these predicted angles and angular velocities as a "virtual reference trajectory" (i.e., the motor state the patient's brain desires to achieve), rather than a mechanically set forced trajectory. A dual closed-loop structure of "outer loop impedance + inner loop force control" is adopted. The outer loop calculates the compliant resistive torque that the rehabilitation device should apply based on the deviation between the "virtual reference trajectory" and the actual motor state, combined with set stiffness and damping coefficients. An resistance adjustment factor is introduced. This resistance adjustment factor is set according to the patient's muscle strength recovery level, with different adjustment factors corresponding to different intensities of resistance training. In the outer loop controller, the desired joint position and desired joint velocity required by the control system are not preset, but are predicted in real time based on the patient's active movement intention.

[0034] like Figure 5 As shown, the adaptive impedance training specifically includes: Simultaneously, the patient's surface electromyography signals and elbow joint motion information acquired by inertial sensors are collected. The multimodal signals are preprocessed and feature extracted to construct a multi-source feature vector. The multi-source feature vector is then input into a trained LSTM motion state prediction model to predict the patient's desired elbow joint angular displacement and desired elbow joint angular velocity. The motion state prediction model is trained offline using a long short-term memory network structure and deployed in the rehabilitation device as model parameters for online inference.

[0035] In the outer loop controller, the desired elbow joint angular displacement and desired elbow joint angular velocity are substituted into the impedance model for calculation to obtain the desired impedance joint torque. ; in, These are the expected joint angular displacement and expected joint angular velocity output by the prediction model; These are the actual joint angles and actual joint angular velocities measured in real time by inertial sensors; These are the joint space damping parameters and stiffness parameters.

[0036] Because the movement speed during elbow joint rehabilitation training is relatively slow and stable, and to ensure training safety, the control system limits both the movement speed and output torque to avoid sudden changes in force and speed. Under these conditions, the change in joint angular acceleration is small, and its impact on the impedance model output is negligible. Therefore, in this embodiment, no acceleration term is introduced, thereby reducing computational complexity and improving system real-time performance while ensuring control stability. Based on the above parameters, the corresponding joint space impedance torque is calculated through the impedance model, which is used to characterize the impedance effect that the rehabilitation equipment should apply to the patient's elbow joint.

[0037] To achieve adjustable resistance training intensity, this embodiment introduces an resistance adjustment factor into resistance-based rehabilitation training. This resistance adjustment factor is set according to the patient's muscle strength recovery level; different adjustment factors correspond to different intensities of resistance training, thereby achieving personalized adjustment of the rehabilitation training intensity. Specifically, the joint auxiliary torque obtained from the joint torque prediction model is multiplied by the resistance adjustment factor to generate the resistance load torque. ; After calculating the desired impedance joint torque and generating the resistive load torque, the desired impedance joint torque output by the impedance model is compared with the no-load torque and the resistive load torque to calculate the torque deviation. The torque deviation is input to the inner loop controller, which adopts a proportional-integral-derivative (PID) control method. The controller outputs the corresponding motor control quantity according to the torque deviation, thereby driving the motor to adjust the output and realizing the precise control of the elbow joint impedance torque by the rehabilitation equipment.

[0038] Through the above control process, the rehabilitation equipment can apply a continuous, stable and adjustable impedance load during the patient's active movement, thereby effectively promoting the further recovery of the patient's muscle strength and motor control ability.

[0039] The technical solutions provided in the embodiments of the present invention have at least the following technical effects: 1. Breaking through the limitations of traditional electromyography (EMG) assessments that only focus on "intensity," this paper proposes a neuromuscular control assessment method based on muscle synergy. By introducing nonnegative matrix factorization (NMF) technology, muscle synergy patterns are decoupled from cluttered EMG signals. Through similarity comparison with healthy modalities, a quantitative evaluation of neuromuscular control coordination in stroke patients is achieved. This allows the system to not only determine whether a patient "has strength" but also "knows how to exert force," providing a deeper physiological basis for correcting abnormal synergy patterns.

[0040] 2. A two-tiered, refined assessment system combining qualitative staging and quantitative scoring has been constructed, addressing the issue of insufficient sensitivity in assessments using single scales. Existing technologies often only output Brunnstrom staging, which is insufficient to reflect short-term training effects. This invention combines the macroscopic staging capabilities of the BRB model with microscopic quantitative scoring based on motion smoothness. This not only guides the switching of major training modes but also captures subtle progress made by patients within the same stage through continuous scoring, enhancing patients' confidence in rehabilitation and providing therapists with precise references for adjusting training details.

[0041] 3. This invention achieves "knowledge-guided online evolution" of the assessment model, overcoming the modeling challenges of scarce rehabilitation medical data and significant individual differences. The invention employs a Confidence Rule Base (BRB) as the core inference engine. Its mechanism of "expert knowledge initialization + online correction based on individual data" effectively avoids the dependence on large datasets and the "black box" uninterpretability of pure neural network models. The model can automatically calibrate rule weights as patient training data accumulates, enabling the assessment system to adaptively evolve from "general" to "personalized."

[0042] 4. An assistive control algorithm based on intention prediction and nonlinear safety adjustment was developed, solving the problems of "zero-point jitter" and human-machine aggression in traditional electromyography (EMG) control. In assisted rehabilitation mode, this invention abandons the simple threshold triggering mechanism and instead uses an offline-trained joint torque estimation model to map the patient's active movement intention in real time. In particular, this invention designs a nonlinear safety control module, which effectively suppresses noise interference from EMG signals at rest or during weak contractions by setting dynamic dead zones and saturation thresholds, eliminating abnormal jitter of the actuator. Combined with a torque tracking closed loop based on PD control, the system can output smooth and stable assistive torque according to the patient's real-time weak intentions, significantly enhancing the coordination of human-machine interaction and allowing the patient to maintain a sense of active participation while receiving assistance.

[0043] 5. This invention achieves an active resistance training mode that "de-presets the trajectory," enhancing the compliance and immersion of rehabilitation training through deep temporal prediction and dual-loop control. Targeting the advanced training needs of the later stages of rehabilitation (Brunnstrom stages 5-6), this invention changes the traditional rigid pattern of "movement along a fixed trajectory." It utilizes a Long Short-Term Memory (LSTM) network to deeply mine the temporal characteristics of multi-source signals, predicting the patient's desired joint angles and angular velocities in real time, thus replacing the preset trajectory. Combining a dual closed-loop structure of "outer loop impedance model + inner loop force control," and introducing an resistance adjustment factor in the outer loop, the system not only provides a compliant resistance sensation but also steplessly adjusts the resistance intensity according to muscle strength recovery. This design greatly replicates the real limb movement environment, stimulates the patient's active movement potential, and effectively promotes the reshaping of neuromuscular control circuits.

[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A smart rehabilitation assessment method for the elbow joint based on multi-source sensor information fusion, characterized in that, include: Multi-source sensor information acquisition and processing: Acquire physiological and motor information of the patient during elbow joint rehabilitation training, including surface electromyography signals, inertial posture signals, and human-computer interaction force signals; The surface electromyography signal matrix is ​​subjected to nonnegative matrix decomposition to extract the muscle coordination vector and time activation coefficient; Calculate the similarity between the co-vector of each muscle in the patient and the co-vector of the muscles in the healthy person; Based on the cumulative activation energy of each muscle synergistically throughout the entire movement, the corresponding synergistic activation weight coefficient is calculated. The neural control synergy index is obtained by fusing synergy vector similarity and synergy activation weight coefficients. By performing dimensionless processing on the accelerometer signal of the elbow joint angular acceleration, motion smoothness features are constructed. Calculate the stiffness characteristics of human-computer interaction based on human-computer interaction force signals and elbow joint angle changes; Macro-stage process: The muscle synergy index, motion smoothness feature and interaction stiffness feature are input into the confidence rule base model. The confidence rule base model integrates prior knowledge and quantitative feature data to output the patient's current Brunnstrom rehabilitation stage. Micro-quantification process: Calculate the comprehensive score of rehabilitation progress, which is constructed based on the continuous numerical value obtained by weighted Euclidean distance calculation of multi-source features relative to the ideal value of the current stage.

2. The method according to claim 1, characterized in that, The formula for calculating the collaborative vector similarity is: ; in, Indicates the patient's first k One collaborative vector, The first number corresponding to a healthy person k One collaborative vector; The formula for calculating the collaborative activation weight coefficient is as follows: ; in, The time activation coefficient represents the activation coefficient at the t-th k A collaborative mode in t Activation intensity at any given moment Indicates the first i The activation intensity of each collaborative vector at time t. N Represents the number of times. h Represents the number of cooperative vectors; The formula for constructing the neural control synergy index is as follows: .

3. The method according to claim 1, characterized in that, The formula for the motion smoothness feature is as follows: , ; in, T Indicates the cycle of motion. Indicates the range of motion angles within the motion cycle. This represents the time difference between two moments when calculating jerk. M Indicates the number of sampling points within the motion cycle. Indicates jerk. This represents the acceleration value at that moment. This represents the acceleration value at the previous moment.

4. The method according to claim 1, characterized in that: The formula for calculating the equivalent human-computer interaction stiffness characteristics is as follows: ; in, This represents the difference between the maximum and minimum values ​​of the human-computer interaction force measured by the pressure sensor during the motion cycle; It indicates the range of motion angles within the motion cycle.

5. The method according to claim 1, characterized in that: The formula for calculating the comprehensive score of the rehabilitation process is as follows: ； in, Indicates the first i The weights of each feature and ; m Indicates the number of features. Indicates the first i The actual value of each feature Indicates the first i The ideal value of each feature and The values ​​are set separately for different stages of rehabilitation, and their ranges are determined by clinical rehabilitation experts based on long-term rehabilitation practice experience and statistical data of patients at different stages. This represents the scoring sensitivity adjustment factor.

6. The method according to claim 1, characterized in that: By employing an adaptive evolution strategy based on the covariance matrix, the rule weights and confidence levels of the confidence rule base model are updated according to newly generated patient assessment data.

7. A smart rehabilitation training method for the elbow joint based on multi-source sensor information fusion, characterized in that, During the Brunnstrom recovery phase output according to any one of claims 1-6, the corresponding control mode is dynamically invoked: When Brunnstrom Phase 1-2 is output, the passive rehabilitation training mode based on S-shaped velocity planning is invoked. The passive rehabilitation training mode uses a fifth-order polynomial S-shaped velocity curve to construct the target motion trajectory of the elbow joint, driving the rehabilitation actuator to move the affected limb to perform passive flexion and extension training. When Brunnstrom Phase 3-4 is output, the assisted rehabilitation training mode based on nonlinear intention mapping is invoked. The assisted rehabilitation training mode adopts a three-level cascaded control architecture of "intention prediction - nonlinear mapping - closed-loop force control", which enables the rehabilitation equipment to dynamically adjust the amount of assistance according to the changes in the patient's active movement ability. When Brunnstrom stages 5-6 are output, the adaptive impedance training mode based on virtual trajectory prediction is invoked. This adaptive impedance training mode uses a deep learning model to mine the correlation between the temporal features of electromyography signals and the current motion state, and predicts the joint angles and angular velocities in real time. The predicted joint angles and angular velocities are used as "virtual reference trajectories" to achieve follow-up impedance training. Then, based on the deviation between the "virtual reference trajectory" and the actual motion state, the compliant impedance torque that the rehabilitation equipment should apply is calculated, and an resistance adjustment factor is introduced to fine-tune the intensity of the impedance training.

8. The method according to claim 7, characterized in that, The passive rehabilitation training mode specifically includes: First, set the starting angle, ending angle, maximum angular velocity, and maximum angular acceleration; then use fifth-order polynomial interpolation to generate a smooth position command sequence with continuously changing angular acceleration. The generated smooth position command sequence is differentially compared with the actual angle fed back by the inertial sensor and input to the position controller to drive the rehabilitation actuator to perform passive flexion and extension training of the affected limb.

9. The method according to claim 7, characterized in that: The rehabilitation training assistive model specifically includes: Surface electromyography (EMG) signals and inertial sensing signals are collected, input into a pre-trained joint torque estimation model, and the predicted torque is output. A double-sided dead zone threshold is set. When the absolute value of the predicted torque is less than the first threshold, the output target auxiliary torque is forced to zero. When the absolute value of the predicted torque is greater than the first threshold and less than the second threshold, an exponential gain scheduling function is used to calculate the target auxiliary torque, so that the magnitude of assistance increases smoothly and non-linearly with the increase of the patient's intention to exert force. When the absolute value of the predicted torque is greater than the second threshold, the target auxiliary torque is saturated and truncated. The shaped target auxiliary torque is compared with the human-machine interaction force measured by the pressure sensor, and the motor output is adjusted by the PD torque controller.

10. The method according to claim 7, characterized in that, The adaptive impedance training specifically includes: Simultaneously, the patient's surface electromyography signals and elbow joint motion information obtained by the inertial sensor are collected. The multimodal signals are preprocessed and feature extracted to construct a multi-source feature vector. The multi-source feature vector is then input into the trained LSTM motion state prediction model to predict the patient's desired elbow joint angular displacement and desired elbow joint angular velocity. In the outer loop controller, the desired elbow joint angular displacement and desired elbow joint angular velocity are substituted into the impedance model for calculation to obtain the desired impedance joint torque. The joint auxiliary torque obtained from the joint torque prediction model is multiplied by the resistance adjustment factor to generate the resistance load torque; then the torque output by the impedance model is compared with the no-load torque and the resistance load torque to calculate the torque deviation; the torque deviation is input to the inner loop controller, and the corresponding motor control quantity is output according to the torque deviation to drive the motor to adjust the output.

Images