Exercise load multi-objective optimization method based on anti-gravity rehabilitation treadmill

Abstract

The invention provides an exercise load multi-objective optimization method based on an anti-gravity rehabilitation treadmill. The exercise load multi-objective optimization method comprises the steps of obtaining gait kinematics data, plantar pressure distribution data, electromyographic signal data and pain scores of lower limbs of a patient; based on the data, a gait asymmetry index and a myoelectricity activation sequence consistency index are screened out, the gait asymmetry index and the myoelectricity activation sequence consistency index are combined with pain scores to serve as key features, and a function grading model is constructed to obtain function grades; establishing a multi-objective optimization problem by taking minimization of a pain score, maximization of joint activity and electromyographic activation sequence consistency and minimization of muscle fatigue as objectives and taking non-overrun of articular cartilage contact stress as a constraint; an NSGA-III algorithm is adopted to solve the problem, and a Pareto optimal solution set of the weight reduction proportion and the speed of the treadmill is obtained; through a PPO reinforcement learning strategy network, taking the real-time state of a patient as input, and selecting and dynamically adjusting treadmill parameters in a Pareto solution set; and issuing the adjusted parameters to a treadmill execution mechanism in real time.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of sports rehabilitation, and particularly relates to a sports load multi-objective optimization method based on an anti-gravity rehabilitation treadmill. BACKGROUND

[0002] The anti-gravity rehabilitation treadmill is an advanced rehabilitation training device for people with lower limb movement disorders, which can significantly improve the gait function, muscle strength and cardiopulmonary endurance of patients. At present, the setting of training parameters (such as the proportion of weight reduction support, the speed of the treadmill and other parameters) in rehabilitation training mainly depends on the experience of therapists and the subjective feedback of patients. Accurate control of load during rehabilitation training is the basis for precise rehabilitation, and therapists usually adjust the training intensity according to the apparent kinematic parameters (such as gait, speed) of patients and subjective pain perception. However, this rehabilitation load decision-making method has strong subjectivity: different therapists have different judgment standards, and patients' pain thresholds also vary from person to person, lacking unified and objective quantitative basis. Existing anti-gravity rehabilitation treadmills mainly focus on adjusting apparent kinematic parameters (such as speed, support force) to meet basic training needs, but ignore the important guiding role of patients' deep biological feedback signals (such as muscle strength changes, pain level) in load adjustment. In addition, due to the lack of systematic data analysis support, the adjustment of training parameters often presents an empirical open-loop control characteristic, and cannot be dynamically optimized according to the real-time physiological state changes of patients. This leads to a lack of scientific multi-index balance in the setting of rehabilitation training load: on the one hand, the training intensity may be insufficient, affecting the rehabilitation effect; on the other hand, if the load is too high, it may cause joint overload, pain aggravation and other risks.

[0003] Based on the above reasons, the existing anti-gravity rehabilitation treadmill has obvious deficiencies in sports rehabilitation load decision-making: there is no data-driven optimization method based on multi-modal biological signals to objectively quantify the functional status of patients and optimize the training parameters accordingly. Specifically, there is a lack of a comprehensive evaluation index system that integrates gait, biomechanics and pain feedback, making it difficult to timely detect the risk of excessive pain or joint stress during training; at the same time, there is no intelligent strategy to dynamically adjust load parameters such as weight reduction ratio and speed according to the rehabilitation progress of patients, and it is impossible to form a standardized and personalized rehabilitation training plan. SUMMARY

[0004] Therefore, the present application aims to solve the above technical problems and proposes a sports load multi-objective optimization method based on an anti-gravity rehabilitation treadmill, which objectively quantifies the functional status of patients by constructing an optimization framework driven by multi-modal biological signals, and optimizes the training load parameters in real time, to significantly reduce the subjective dependence of decision-making and improve the precision and safety of rehabilitation training.

[0005] The specific technical solution is: a sports load multi-objective optimization method based on an anti-gravity rehabilitation treadmill, comprising:

[0006] Obtaining gait kinematics data, plantar pressure distribution data, electromyography signal data and pain score of the lower extremity of the patient;

[0007] Based on the data, candidate variables such as step length, step frequency, gait asymmetry index, electromyography activation sequence consistency index, median frequency and COP trajectory feature are extracted, LASSO regression is used to screen out gait asymmetry index, electromyography activation sequence consistency index and pain score as key features in the comprehensive rehabilitation score prediction task, and a functional classification model is constructed to obtain a functional grade;

[0008] A multi-objective optimization problem is constructed with the goals of minimizing pain score, maximizing joint range of motion and electromyography activation sequence consistency, and minimizing muscle fatigue, and with the constraint of not exceeding the limit of articular cartilage contact stress;

[0009] NSGA-III algorithm is used to solve the problem, and a Pareto optimal solution set of treadmill weight reduction ratio and speed is obtained;

[0010] Through the PPO reinforcement learning strategy network, the real-time state of the patient is input, and the treadmill parameters are selected and dynamically adjusted in the Pareto solution set;

[0011] The adjusted parameters are real-time issued to the treadmill actuator.

[0012] The present application has the following beneficial effects:

[0013] The present application effectively solves the defects of strong subjectivity of traditional rehabilitation treadmill training load decision and single data dimension by constructing a multi-modal biological signal driven intelligent optimization system. By integrating inertial measurement, muscle strength evaluation and pain monitoring modules, a leapfrog expansion from single macro motion parameter to "gait biomechanics-muscle function-pain feedback" multidimensional information is realized. Such comprehensive sensing capability enables the system to quantitatively measure the muscle activation level, joint function status and pain degree of the patient in real time, breaking through the limitations of previous training adjustment relying only on surface parameters such as speed and support force. With the help of LASSO screening and functional classification model, the system extracts core indicators reflecting the patient's function, realizes objective quantitative evaluation of the patient's rehabilitation state, and lays a scientific foundation for load optimization.

[0014] Further, by establishing a closed-loop optimization control link between biofeedback and training equipment, the application successfully overcomes the lag and inaccuracy of artificial experience adjustment. The innovative multi-objective optimization and deep reinforcement learning decision mechanism directly converts the real-time collected quantitative indicators such as pain and gait into dynamic adjustment parameters of the treadmill reduction ratio and speed, enabling the treadmill to adaptively optimize the training load according to the actual state of the patient. This "perception-optimization-execution" closed-loop mechanism realizes the leap from passive response to active decision: instead of waiting for human judgment to adjust, it predicts in advance and adjusts in time to ensure that the patient is always at the most suitable training load level. Thus, the application significantly improves the individualization and precision of rehabilitation training, while avoiding overtraining injuries and maximizing functional recovery, providing a new intelligent control paradigm for the clinical application of anti-gravity rehabilitation treadmills. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a whole flowchart of the method of the application;

[0016] Figure 2 is a flowchart of muscle signal scoring;

[0017] Figure 3 is a pain score diagram;

[0018] Figure 4 is a multi-objective optimization flowchart;

[0019] Figure 5 is a reinforcement learning flowchart;

[0020] Figure 6 is a system integration flowchart. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other. In order to achieve the above purpose, the application adopts the following technical solutions.

[0022] The application proposes a multi-objective optimization method for exercise load based on an anti-gravity rehabilitation treadmill based on multi-source bio-signal fusion and intelligent optimization control, and the overall process is as follows Figure 1As shown, it covers the whole process from multi-source data collection, function evaluation, optimal decision to real-time training parameter regulation. Specifically, the system first collects multi-modal biological signals to obtain key movement and physiological indicators of patients during anti-gravity rehabilitation treadmill training; then extracts features and selects key indicators from the collected data to establish a quantitative evaluation model of the patient's functional status; next, based on the model, a multi-objective optimization problem is constructed and solved by an evolutionary algorithm to determine the optimized training load parameter set; finally, the optimization results are combined with a deep reinforcement learning strategy to train a real-time decision network, achieving adaptive control of treadmill weight reduction ratio, speed and other parameters. The whole system runs locally and seamlessly integrates with the treadmill hardware, forming a closed-loop intelligent rehabilitation training control system. The following steps are included:

[0023] Step 1, multi-source data collection and gait and physiological feature acquisition, the present application synchronously acquires kinematics, biomechanics and subjective feedback information of patients by using multiple sensors and evaluation devices, realizing comprehensive monitoring of the functional status of patients. As shown in the figure, Figure 2 First, during treadmill training, the patient wears an inertial measurement unit (IMU) on the lower limbs, and a plantar pressure sensor array is arranged on the treadmill bed to collect gait parameters in real time. The kinematics data of each joint of the lower limbs, including joint angle range, angular velocity, etc., are obtained through the IMU, and the force distribution of each area of the plantar, the trajectory of the center of pressure, and the time parameters of the gait cycle are collected through the plantar pressure sensor. In particular, the system focuses on extracting gait asymmetry indicators: such as the difference between the left and right foot support phase time and swing phase time, and calculating the gait asymmetry index to quantify the stability and balance of the patient's gait.

[0024] ;

[0025] ;

[0026] and represent the asymmetry coefficient of gait; and represent the shorter and longer values in the mean of the respective support phase time of the left and right feet during walking; and represent the shorter and longer values in the mean of the respective swing phase time.

[0027] In particular, to evaluate the muscle function status of the patient's lower limbs and the dynamic changes during training, the present application introduces surface electromyography (sEMG) signals to replace the traditional isokinetic muscle strength test method. sEMG sensors are arranged on muscle groups such as the quadriceps femoris, hamstrings, tibialis anterior, etc., and by extracting signal features such as RMS value, mean square root frequency (MF), activation sequence, etc., muscle activation consistency indicators (MCI) are calculated. ) and muscle fatigue spectrum index (MFSI ) to measure the muscle group coordination and fatigue trend.

[0028] ;

[0029] sEMG signal value of the i th sampling point; N: total number of sampling points; RMS measures the energy size of the electromyographic signal, which is proportional to the muscle contraction strength, and is one of the most commonly used amplitude characteristics of electromyographic signals.

[0030] ;

[0031] T: total number of gait cycles; : whether the activation sequence is consistent in the t th cycle (Boolean value); characterizes whether the muscle activation sequence is stable and consistent with the healthy template, and the higher the value, the more normal the sequence.

[0032] ;

[0033] : median frequency of sEMG signal at the beginning of training; : median frequency at the end of training; the median frequency of the electromyographic spectrum decreases during fatigue, so the larger the index, the higher the fatigue level.

[0034] In addition, for the subjective feeling of pain of the patient, the present application adopts the visual analogue scale (VAS) to quantify the pain intensity. During the training process, the patient is asked to evaluate the current pain level on the VAS scale of 0-10 at regular intervals (for example, after completing a certain time or distance), and the score is recorded as a pain feedback signal, as shown in Figure 3 The VAS score provides direct feedback on the patient's tolerance of the training load, and the higher the value, the stronger the pain. To ensure the objectivity of the data, the VAS score can be collected multiple times during the patient's training at different times, and the average value or the change trend can be focused on to reflect the change of pain with the progress of training.

[0035] During data collection, the system ensures the alignment of multi-source data in time and space by combining hardware synchronization and software calibration. For example, the IMU and plantar pressure sensors are synchronized by triggering the sampling with a unified clock, and the dynamic time warping algorithm is used to match gait events (such as the heel strike moment and the peak joint angular velocity). This ensures that the gait parameters and mechanical data are analyzed on the same time axis, which is beneficial for subsequent feature fusion. In addition, by filtering the joint movement signals measured by the IMU and converting them to the global coordinate system, the comparability between different sensor data can be improved; the pressure sensor data is calibrated to correspond to the actual force value. After all the raw data is preprocessed as described above, it enters the feature extraction and analysis stage.

[0036] The dynamic time warping algorithm uses the time series of pressure events as the reference signal, and the time series of joint angular velocity and the peak time series of electromyographic signals as the signals to be aligned. By calculating the Euclidean distance matrix of each sensor signal in the time domain, and using the cumulative minimum distance path for elastic time matching, the local time scale is nonlinearly stretched and registered in the presence of small differences in gait cycle, ensuring synchronization of multi-source signals on a unified time axis.

[0037] Step 2, key indicator selection and functional grading model establishment: Based on the collected lower limb kinematic data, plantar pressure distribution data, electromyographic signal data, and pain score, the candidate variables such as step length, step frequency, gait asymmetry index, electromyographic activation consistency index, median frequency, and COP trajectory features are extracted through feature engineering. The present application uses LASSO regression (Least Absolute Shrinkage and Selection Operator Regression) algorithm for feature selection and functional evaluation modeling. First, the multi-dimensional features obtained, including gait parameters, electromyographic signals, pain scores, etc., are associated with the patient's functional status indicators to build a regression model to predict or evaluate the patient's functional performance. Due to the high initial feature dimension and the possible correlation between different features, directly modeling with all features can easily lead to overfitting and is difficult to interpret. LASSO regression can compress the regression coefficients of unimportant features to zero by introducing an L1 norm penalty term in the regression process, thereby achieving feature selection. By training the LASSO model, the present application can automatically select a few key indicators that have the greatest impact on the patient's functional status. LASSO is a regularization regression method for feature selection and overfitting prevention, and its optimization objective is as follows:

[0038] ;

[0039] : target variable of the i-th sample; : feature vector of the i-th sample; : regression coefficient vector; : regularization hyperparameter.

[0040] Based on a large amount of rehabilitation training data, the LASSO feature selection results show that, through LASSO regression, the gait asymmetry index and electromyographic activation sequence consistency were ultimately selected. Using gait asymmetry index and VAS pain score as core indicators, a functional grading model is constructed. This model can use methods such as multiple linear regression or decision trees to map the combination of gait asymmetry index, EMG activation sequence consistency, and pain score into a functional score or grade. Functional grading helps the system determine the patient's current rehabilitation stage (early, stable, or intensive). The grading results directly affect the weighting coefficients in the multi-objective optimization model. For example, early rehabilitation focuses on controlling pain (higher α) and reducing joint load (higher θ); mid-stage rehabilitation focuses on expanding ROM and restoring muscle coordination (increased β and γ); late-stage rehabilitation focuses on endurance and fatigue recovery (increasing the weight of muscle fatigue indicators). Furthermore, during NSGA-III optimization, the initial solution space for patients with lower functional grades can be more conservatively restricted in the initial search population or decision space. In long-term system operation, this score also serves as a quantitative assessment of therapeutic efficacy. When the gait asymmetry index is close to 0 (ideal symmetry), EMG activation sequence consistency is high, and the VAS score is low, the model assigns a high functional level; conversely, it assigns a low functional level. The functional level output by the model can be a discrete level (such as level I, II, III) or a continuous comprehensive score. The higher the value, the better the patient's functional status.

[0041] This functional grading model enables the training system to objectively quantify the patient's functional recovery level. Compared to traditional methods that rely solely on subjective descriptions or single-indicator scoring, this model integrates information from multiple aspects, including gait, physiology, and pain, resulting in a more comprehensive and accurate assessment. The functional grading model also provides a foundation for subsequent load optimization. When determining the training plan, different initial training load ranges of varying intensities can be selected based on the patient's current functional level. During the optimization process, the functional level is used as a reference to ensure that the proposed training parameters are appropriate for the patient's tolerance and rehabilitation stage.

[0042] Step 3: Construction and Solution of the Multi-Objective Optimization Model: Based on the patient's current functional assessment, this invention defines the optimization of rehabilitation training load parameters as a multi-objective optimization problem. For example... Figure 4As shown, the first goal is to minimize the pain level of the patient, i.e. to reduce the VAS score as much as possible; the second goal is to maximize the joint mobility of the patient, i.e. to improve the flexibility and range of motion of the lower limb joints. It is noted that the joint mobility can be improved by increasing the training intensity, but this often increases the pain and joint stress, so there is a conflict between the two goals. In addition, in order to ensure the safety of the training, the key constraint condition is set in the optimization problem: the cartilage contact stress of the lower limb joints of the patient should not exceed the safety threshold. The cartilage stress threshold can be set according to medical research or clinical standards, and the maximum safe pressure per square millimeter of the joint is estimated according to the body weight and cartilage condition of the patient, and exceeding the threshold may cause the risk of joint cartilage damage. Therefore, any training parameter combination that causes excessive joint impact force is considered as an infeasible solution and is excluded. After determining the goals and constraints, the NSGA-III (Non-dominated Sorting Genetic Algorithm III) is used to solve the above multi-objective optimization problem. NSGA-III is a global optimization algorithm based on evolutionary ideas, which is particularly suitable for multi-objective optimization and can generate a set of Pareto optimal solutions rather than a single solution. The algorithm takes a series of candidate training schemes as the initial population, and each scheme is composed of a set of decision variables. In the present invention, the decision variables mainly include the weight reduction support ratio of the anti-gravity rehabilitation treadmill (for example, the percentage of the patient's body weight) and the speed of the treadmill, and if necessary, the training duration or intermittent mode and other parameters can also be extended. Each individual in the population represents an example of a training load scheme, and the system estimates the pain index, joint mobility index and joint cartilage stress value corresponding to the scheme according to the biomechanical model or empirical formula. The pain index can be directly taken as the predicted change of the patient's latest VAS score under the load condition, the joint mobility index can be estimated in combination with the patient's current functional level and training parameters, and the cartilage stress can be estimated by the biomechanical model. The multi-objective optimization function is to minimize the VAS score, maximize the joint mobility (ROM), muscle activation consistency index (MAI) and minimize the fatigue index:

[0043]

[0044]

[0045] x = [r, v] is the training parameter combination to be optimized. Wherein: r represents the weight reduction ratio of the treadmill (such as 30% means only 70% body weight is supported), v represents the speed of the treadmill; is the weighted coefficient of the target term, which can be manually set according to the rehabilitation stage and clinical goals; : subjective pain score, the goal is to minimize; : joint mobility (Range of Motion), the goal is to maximize; ​​​Muscle activation sequence consistency index, the goal is to maximize; Muscle fatigue spectrum index, the goal is to minimize it; : Cartilage contact stress under the selected parameters (which can be estimated using a biomechanical model); : The maximum safety threshold set.

[0046] The NSGA-III algorithm evaluates solutions in the population through a non-dominated ranking mechanism: it prioritizes solutions that are not inferior to others in multiple objectives (i.e., Pareto superior) for the next generation. Simultaneously, it introduces methods such as crowding distance to maintain solution diversity, thereby obtaining a set of optimal solutions—the Pareto front solution set—after iterative optimization, balancing factors such as pain, activity level, muscle fatigue, and consistency of muscle activation sequence. Each solution in this set corresponds to a combination of weight loss percentage and treadmill speed. Through this multi-objective optimization search, the system obtains a variety of selectable training parameter combinations for clinical use, allowing for trade-offs and selections based on individual needs.

[0047] During the optimization process, all candidate solutions must satisfy cartilage stress constraints—the algorithm automatically eliminates any combination where the estimated joint stress exceeds a safe threshold. Therefore, solutions at the Pareto frontier are safe, feasible, and achieve an optimal trade-off between pain and function. Compared to traditional methods where therapists adjust training parameters with a single objective, this multi-objective optimization model simultaneously considers rehabilitation effectiveness and safety, ensuring that training effectively improves function without causing excessive pain or injury risk. This optimization module typically runs offline before each patient training session, calculating the recommended set of parameter combinations for that session based on the patient's condition that day.

[0048] Step 4: Deep Reinforcement Learning Policy Network and Real-Time Decision Making: Although multi-objective optimization can provide a set of optimal solutions, the patient's state during actual training is dynamically changing. Relying solely on fixed offline calculations is insufficient to adapt to unexpected situations and gradual changes during training. Therefore, this invention introduces deep reinforcement learning technology on top of multi-objective optimization to train an intelligent decision-making network, enabling real-time adjustment and optimization of training load parameters. Specifically, the system employs advanced deep reinforcement learning algorithms such as Proximal Policy Optimization (PPO) to construct a policy network that takes the patient's current state as input and outputs the optimal treadmill parameter control decision.

[0049] First, the environment and elements of reinforcement learning need to be defined. In the present scenario, the state space is the training process of the patient on the anti-gravity treadmill, and its state includes the aforementioned functional assessment indicators and training feedback, such as the current gait asymmetry index, real-time electromyographic signal changes, the current VAS pain score or pain alarm signal, and the current weight reduction ratio and speed of the treadmill, etc. The action space is defined as the adjustment of the treadmill parameters by the system, including the speed of the treadmill, the weight reduction ratio, the training duration and the interval time, etc. For simplicity, the actions can be discretized, with the weight reduction ratio adjusted by ±2%, the speed adjusted by ±0.5km / h, the training time adjusted by ±5min, and the interval time adjusted by ±5min. The design of the reward function integrates the multi-objective requirements of rehabilitation: during the training process, a given reward value r is given at each time step, which is set to encourage the reduction of pain and the improvement of joint range of motion, while punishing situations that exceed the safe stress threshold. Specifically, let the reward be:

[0050] ;

[0051] where VAS represents the pain score at the current time step, ROM represents the estimated joint range of motion, StressPenalty is a large penalty given when the cartilage stress exceeds the threshold, is the weight coefficient used to balance the dimension. Through such a reward setting, the agent will learn in training to both control pain at a low level and maximize the range of motion as much as possible, while avoiding any behavior that may harm the joint.

[0052] In the training phase, real patient data or a simulated patient model based on the model is used to interact with the policy network repeatedly for reinforcement learning: the agent (policy network) observes the current state, selects an adjustment action to act on the environment (i.e., changes the weight reduction or speed of the treadmill), and the environment feeds back the new patient state (such as a slight increase in pain or an increase in ROM due to a decrease in weight reduction) and the immediate reward according to the biomechanical model. Through the PPO algorithm, the policy parameters are constantly iterated, and the agent gradually learns to take actions that can obtain higher long-term returns in different states. Unlike traditional Q-learning, PPO is an on-policy policy optimization method that can steadily improve the policy and avoid drastic fluctuations. After sufficient training, the resulting policy network can be deployed in the actual system for real-time decision-making. The maximization of the long-term expected return is:

[0053] ;

[0054] where represents the expected cumulative reward in the future k steps. k represents the time k steps after the current time t. γ is the discount factor, and 0 < γ ≤ 1. γ is used to control the weight of the future reward in the current value. The smaller γ is, the more short-sighted the algorithm is, and the more emphasis it puts on recent rewards. The larger γ is, the more emphasis it puts on long-term benefits. represents the immediate reward value obtained at the future k step.

[0055] In actual operation, as shown in Figure 5 , the reinforcement learning strategy network evaluates the patient's state according to the latest sensor data every few seconds (for example, every gait cycle) and outputs adjustment instructions. For example, if the patient's gait symmetry suddenly deteriorates and the VAS increases, the strategy network may output the action instruction "increase the weight reduction by 5% and reduce the speed by 0.5 km / h", so as to reduce the patient's load, relieve discomfort and stabilize the gait; on the contrary, if the patient's state is stable and the pain is very low, the strategy may gradually reduce the weight reduction ratio or increase the speed to ensure that the training load is sufficient to stimulate the improvement of joint range of motion. Through continuous fine-grained adjustment, the system realizes closed-loop adaptive optimization of the training load: it not only avoids harm caused by excessive load when the patient's state deteriorates, but also continuously challenges and improves the training intensity within the patient's tolerance range, ultimately achieving the purpose of individualized optimization of rehabilitation. This deep reinforcement learning driven control strategy effectively makes up for the limitations of offline optimization, and upgrades the anti-gravity rehabilitation treadmill from a passive execution of a predetermined program to an active intelligent decision-making device.

[0056] Step 5, system integration and local deployment: as shown in Figure 6 , the optimization decision algorithm described in the present application is run entirely on a local computing device, without relying on cloud services, achieving the unity of real-time performance and data security. A high-performance embedded processor or industrial computer is configured in the control system of the anti-gravity rehabilitation treadmill to deploy the LASSO model, the multi-objective optimization algorithm and the deep reinforcement learning network. Since real-time control requires millisecond-level response, all algorithms are optimized for local hardware. For example, the multi-objective optimization module can calculate the Pareto solution set in advance for reference during the training gap, and the reinforcement learning inference part uses a model-compressed neural network to improve inference speed. In terms of system architecture, the data acquisition module, the optimization decision module and the execution control module are tightly integrated: sensor data is transmitted in real time to the processor through the interface circuit, and after being calculated by the functional evaluation model and the strategy network, the generated control instructions are directly sent to the treadmill drive unit (including the weight reduction air pressure controller and the treadmill motor controller). The entire process takes no more than a few tens of milliseconds, which is much lower than the human reaction time and the physiological index change rate, thus achieving a control experience with no perceptual delay.

[0057] All data are stored and processed locally, avoiding privacy leakage and latency problems caused by cloud transmission, and meeting the requirements of data security and real-time reliability in rehabilitation medical scenarios. The software architecture of the system provides a friendly interface to the outside, which can record the key data and decision results in the training process for doctors to view, but will not upload the original biosignal to the cloud. In terms of hardware implementation, the present application makes full use of the existing platform of the anti-gravity rehabilitation treadmill: the sensor components, multi-objective optimization and RL algorithm modules are integrated as software upgrades, and the remaining physical components can run without major changes. This means that the present application has good engineering feasibility and practical value, and can significantly improve the intelligent optimization function of training load by combining software and hardware on the basis of the existing rehabilitation treadmill.

[0058] In order to facilitate the understanding of the technical solutions of the present application, the key links are described from the perspective of system architecture and workflow as follows, highlighting the closed-loop data link of perception-evaluation-optimization-decision-execution:

[0059] Multi-modal data fusion in the perception layer: The perception layer of the present system integrates multiple sensors and testing methods to achieve a three-dimensional perception of the patient's movement state. For example, we use IMU sensors (Inertial Measurement Unit) to capture the movement acceleration and angular velocity of the patient's lower limbs, use plantar pressure sensors to collect the force distribution of each part of the foot during gait, and combine muscle strength test results to quantify the patient's muscle strength level. A single sensor often cannot fully reflect the motor function, and multi-modal data fusion can make up for this deficiency. By fusing the kinematic data of IMU, plantar pressure information and muscle strength evaluation results into a unified input, the present system obtains more comprehensive and accurate patient state data. These multi-source data provide a reliable basis for subsequent functional evaluation and decision control.

[0060] Function evaluation model and key indicator selection: The multi-modal data obtained by the perception layer is input into the function evaluation model to assess the patient's current functional status and rehabilitation progress. Due to the large number of original features collected by the sensors, we introduce the LASSO method (Least Absolute Shrinkage and Selection Operator) to select important indicators. LASSO is a regression algorithm with L1 regularization that can automatically shrink the weights of irrelevant features to zero, thus selecting a small number of key features. Through LASSO feature selection, the evaluation model extracts indicators closely related to rehabilitation outcomes (such as gait symmetry, joint range of motion, muscle fatigue, etc.) from the vast amount of data, avoiding interference from secondary data affecting the evaluation results. Based on the selected key indicators, a rehabilitation function grading system is established: that is, the patient's rehabilitation status is divided into grades or scores according to their abilities. This grading system is like giving the patient a score on their current functional performance, which helps therapists and the system intuitively understand the patient's rehabilitation stage and determine the initial setting and adjustment range of training load accordingly.

[0061] Load optimization and NSGA-III multi-objective algorithm: After determining the patient's functional classification and baseline state, the invention uses a multi-objective optimization algorithm to customize the load scheme for rehabilitation training. The determination of rehabilitation training load involves the trade-off of multiple objectives: for example, it is desirable to increase training intensity to promote functional recovery, while it is also necessary to control the load to avoid pain or secondary injury, and it may also consider training efficiency and patient subjective comfort, etc. These objectives often conflict with each other and are difficult to optimize with a single indicator. Therefore, we use the Non-dominated Sorting Genetic Algorithm III (NSGA-III) to generate a set of load optimization solutions. NSGA-III is an efficient multi-objective optimization method based on evolutionary algorithms that can consider multiple conflicting objectives simultaneously and find a set of optimal solutions that balance each objective, i.e. the Pareto front solution set. Simply put, NSGA-III simulates the "survival of the fittest" evolution process, searching for different training parameter combinations (such as treadmill weight reduction ratio, speed, training time, etc.), and selecting a number of solutions that achieve a balanced compromise on different objectives. The solution set formed by these solutions has no absolute advantage or disadvantage - each solution performs better in some indicators, but may sacrifice other indicators. For example, one solution may have a higher load, good training effect, but is relatively hard, while another solution has a moderate load and is more comfortable but has a slower effect. NSGA-III provides these candidate solutions for selection, forming a Pareto optimal solution set for the decision module to refer to. In this way, we do not need to subjectively choose a single solution, but let the algorithm give multiple alternatives, and let the system select the most suitable one in the subsequent steps according to the real-time situation.

[0062] PPO policy network closed-loop control: After obtaining the optimized candidate solution set, the present application makes closed-loop control decisions on the training load through a deep reinforcement learning policy network. We use the PPO algorithm (Proximal Policy Optimization) to let the agent (policy network) learn to dynamically adjust the treadmill load according to the real-time state of the patient. PPO is a reinforcement learning algorithm of the _actor-critic_ architecture: the "actor" (policy network) outputs an action decision (such as increasing or decreasing the treadmill weight reduction ratio, adjusting the speed, etc.) according to the current environmental state input (patient sensor data and functional assessment results); at the same time, the "critic" (value network) will evaluate the effect of the action, calculate the reward signal, and evaluate whether this adjustment is conducive to the rehabilitation goal. As the training progresses, the actor network continuously optimizes the decision-making strategy under the feedback guidance of the critic, so that the cumulative reward (corresponding to the comprehensive evaluation of rehabilitation effect) is maximized. It needs to be emphasized that the PPO algorithm limits the change amplitude of the new and old strategies when updating the strategy, and introduces a clipping mechanism to avoid excessive adjustment each time, so as to ensure the stability of the training process. Intuitively, the PPO reinforcement learning policy network is equivalent to an experienced "intelligent coach": it continuously perceives patient feedback (sensor data changes), and then fine-tunes the support force or speed of the treadmill and other parameters. If the adjustment improves the rehabilitation indicators, it will receive a positive reward, otherwise it will be punished, so as to continuously learn the best control strategy through trial and error. Closed-loop control means that this process is a continuous cycle: after the policy network performs a load adjustment, it immediately sees the patient's response through the sensor, such as whether the gait is more stable, whether the muscles are excessively fatigued, etc., and then decides the next action, and so on. This repeated process realizes real-time and adaptive control of the treadmill load. This closed-loop reinforcement learning control ensures that the training load is always optimized in real time according to the patient's state, avoiding the situation of being too light and ineffective or too heavy and harmful, so that rehabilitation training is both safe and efficient.

[0063] System hardware integration and local deployment operation: The method of the present application fully considers the practicability in actual application, can directly combine the existing anti-gravity rehabilitation treadmill hardware for deployment, and all models are locally run without relying on the cloud. The anti-gravity rehabilitation treadmill is a rehabilitation training device that realizes weight reduction support through air pressure difference technology. The sealed cavity provides upward bearing support force for the user by adjusting air pressure, and can gradually reduce the actual bearing of the patient from 100% body weight to partial body weight. For example, a commercial device can reduce the body weight load from 100% to 20% without affecting the natural gait of the patient, and can finely adjust the bearing ratio in steps of 1%. The system utilizes this controllable weight reduction characteristic of the treadmill to directly act on the treadmill control unit through the software and hardware interface: when the strategy network decides to adjust the load, the corresponding instruction is locally sent to the treadmill (for example, modify the air pressure intensity to change the weight reduction ratio, adjust the running belt speed, etc.), and the treadmill immediately executes the new parameter setting, realizing seamless connection of decision-making and execution. The whole process does not need to upload data to the cloud or rely on remote servers, and all calculations (including the function evaluation model, NSGA-III optimization algorithm and PPO control algorithm) are completed on the local computing device matched with the treadmill. The advantage of this is fast response, high real-time performance, while avoiding the risk of network delay or interruption, and ensuring the privacy and security of patient data. The locally deployed model is fully trained and verified, and can run efficiently on ordinary industrial computers or embedded computing platforms, real-time processing of sensor data and output of control decisions to ensure smooth implementation of system closed-loop control.

[0064] In summary, the present application constructs a complete closed-loop rehabilitation training control system, whose data link contains five links of perception -> evaluation -> optimization -> decision -> execution in turn, and circulates: first, the multi-modal perception layer obtains patient state information, the key indicators are refined and the rehabilitation level is judged by the function evaluation model; then the NSGA-III algorithm provides multiple optimized training load schemes; then the PPO strategy network selects and implements the best scheme, and the specific load adjustment is executed through the treadmill hardware; the response of the patient under the new load is captured by the perception layer and fed back to the system, entering the next cycle. This closed-loop control process enables the training load to be adjusted adaptively according to the real-time state of the patient, ensuring that the entire rehabilitation training process approaches the optimal rehabilitation effect under the premise of safety and controllability. The system is analogous to an "intelligent rehabilitation coach": it monitors the state of the student (patient) at all times, assesses his ability, intelligently plans the training program, and constantly adjusts the program according to the student's performance during training, ultimately realizing scientific, efficient and personalized rehabilitation training management.

[0065] IMU (Inertial Measurement Unit): A miniature device containing accelerometers, gyroscopes, and magnetometers that measures linear acceleration, angular velocity, and orientation of the human body or objects, widely used in sports biomechanics analysis to measure joint angles and movement trajectories.

[0066] Plantar pressure sensor: A sensor used to measure the pressure distribution of each area of the human foot in real time, usually arranged in an array on the insole or ground, providing information such as gait parameters (such as step length, step frequency, support phase time), and pressure center trajectory.

[0067] VAS (Visual Analogue Scale): A scale used for subjective evaluation of pain intensity, usually from 0 to 10, patients mark on the scale according to their own feelings, 0 represents no pain, and 10 represents unbearable pain.

[0068] GAI (Gait Asymmetry Index): An index used to quantitatively describe the degree of asymmetry between the left and right sides of the patient's gait, the formula is the absolute value of the difference between the two sides of the step length or support phase time divided by the average value of the two sides, the larger the value, the more serious the asymmetry.

[0069] LASSO regression (Least Absolute Shrinkage and Selection Operator): A feature selection and regularization method that adds an L1 regularization penalty term to the regression coefficient, which can effectively compress the coefficient to zero and filter out key variables.

[0070] NSGA-III (Non-dominated Sorting Genetic Algorithm III): An advanced multi-objective optimization evolutionary algorithm that can efficiently find a series of Pareto optimal solution sets when dealing with multiple conflicting objectives, widely used in engineering optimization problems.

[0071] Pareto front: In multi-objective optimization problems, it represents the set of all non-dominated (i.e. cannot be improved by improving one objective without reducing other objectives) optimal solutions, providing multiple compromise options for decision-makers.

[0072] Cartilage contact stress: The mechanical pressure that the human joint cartilage bears during weight-bearing activities, exceeding the safety threshold may cause cartilage damage or accelerate degeneration, used as a safety constraint condition in the optimization of training parameters in this invention.

[0073] Deep Reinforcement Learning (DRL): A machine learning method that combines deep learning neural networks with reinforcement learning theory, enabling artificial intelligence systems to automatically learn optimal decision-making strategies in interactions with the environment. It is widely used in automated control and decision-making scenarios.

[0074] Proximal Policy Optimization (PPO): A deep reinforcement learning algorithm that limits the magnitude of policy updates to ensure training stability. It is suitable for control tasks in continuous or discrete decision spaces and is commonly used in robot control and real-time optimization decision-making problems.

[0075] Closed-loop Control: A feedback control system that monitors output in real-time and adjusts input based on feedback, achieving adaptive dynamic adjustment to make actual output always approach the target state.

[0076] The invention has the following alternatives:

[0077] 1. Multimodal acquisition and integration of IMU, plantar pressure, and muscle strength data

[0078] Alternative: In terms of gait parameters, if there is a lack of plantar pressure pads, dual-foot posture sensors such as shoe pad pressure switches or camera vision analysis can be used to obtain step frequency, step length, and symmetry information. Pain scores can be directly input by patients through real-time feedback buttons on the treadmill combined with a numerical scale. In data fusion processing, replace hardware synchronization with mutual correlation alignment of transmission signals. Determine the correspondence between gait events by calculating the cross-correlation between IMU acceleration peaks and plantar contact signals to ensure data collaborative analysis. This alternative reduces dependence on professional equipment and uses general sensing and estimation methods to obtain main rehabilitation evaluation information, achieving multi-source data acquisition and fusion under simplified conditions.

[0079] 2. Key indicator screening and functional grading model based on LASSO regression

[0080] Alternative approach: Principal Component Analysis (PCA) and other dimensionality reduction methods can be used to replace LASSO screening for selecting key features. PCA reduces the dimensionality of gait and physiological multidimensional data, extracting the top principal components as comprehensive indicators, avoiding biases introduced by manual feature selection. Simultaneously, the functional assessment model can employ expert-weighted scoring methods: for example, rehabilitation experts pre-set weight coefficients for gait symmetry, muscle strength ratio, and pain indicators, linearly combining them into a functional score, rather than using machine learning for automatic modeling. In this approach, the score decreases accordingly when the gait asymmetry index exceeds a certain threshold, the muscle strength ratio is below the normal range, or the pain is above a certain level, thus reflecting the degree of functional limitation. This alternative approach utilizes traditional statistical analysis and expert knowledge to achieve functional assessment, avoiding complex regularized regression algorithms, and offers advantages such as simplicity and easy interpretation of results when data is limited.

[0081] 3. NSGA-III Multi-Objective Optimization Model

[0082] Alternative approach: Replace multi-objective evolutionary algorithms with weighted and single-objective optimization. By assigning appropriate weights to the two objectives of pain and joint range of motion, the multi-objective problem is transformed into a single objective. The optimization problem (where For weights, minimizing VAS and maximizing ROM is equivalent to minimizing the negative. Then, traditional single-objective optimization algorithms (such as particle swarm optimization or sequential quadratic programming) can be applied to find the comprehensive objective. The minimum solution. Simultaneously, to obtain solutions with different preferences, the weight ratios can be adjusted for repeated optimization, resulting in a series of solutions approximating the Pareto front. Another alternative is to use earlier generation multi-objective genetic algorithms such as NSGA-II. Although the convergence efficiency is slightly lower than NSGA-III, it can still produce a Pareto optimal solution set. The above alternatives use simpler or more mature optimization algorithms to implement multi-objective trade-offs, achieving similar effects to NSGA-III to some extent, while avoiding the potential problem of high implementation complexity of NSGA-III.

[0083] 4. Intelligent decision-making for real-time training load in deep reinforcement learning

[0084] Alternative: Use fuzzy control + classic PID regulation combined rule control instead of deep reinforcement learning strategy. First, design a set of fuzzy control rules, for example, with "pain increases and gait instability" as the condition to trigger the control output of "increase weight reduction and reduce speed", and predefine the corresponding adjustment strategy for different physiological state combinations. The fuzzy controller will fuzz the gait symmetry deviation, pain score, etc. as input, and output the adjustment suggestion for the treadmill parameters through rule base reasoning. Then combine with PID control for fine adjustment: for example, set the target of gait asymmetry index to 0, use PID to dynamically adjust the weight reduction ratio according to the current asymmetry error; to maintain the pain score in a safe range, use another PID controller to adjust the treadmill speed or pause the training according to the pain error. Feedforward-feedback control strategy can also be introduced: such as reducing the load in advance according to the predicted fatigue trend (feedforward), and then fine-tuning according to real-time error feedback. This scheme is completely based on deterministic control algorithms and expert experience rules, realizes the adaptive adjustment of training load, avoids the large amount of data and computing resources required for deep reinforcement learning model training, and reduces the complexity of system implementation.

[0085] In summary, the present application constructs a complete anti-gravity rehabilitation treadmill rehabilitation load optimization decision framework through the four stages of "data collection - index modeling - multi-objective optimization - reinforcement learning control". Figure 1 The flowchart fully embodies the synergy of multi-source information fusion and intelligent control: first, use multiple sensors to obtain patient movement and physiological data, then condense key indicators for functional evaluation, then use evolutionary algorithms to develop a scheme that meets rehabilitation goals and safety constraints, and finally use a reinforcement learning agent to execute real-time fine-tuning. This scheme significantly improves the scientificity, timeliness and individual adaptability of rehabilitation training parameter adjustment, and provides a set of objective, efficient and safe training load optimization method for patients with lower extremity motor disorders.

Claims

1. A multi-objective optimization method for exercise load based on an anti-gravity rehabilitation treadmill, characterized in that, include: Acquire gait kinematics data, plantar pressure distribution data, electromyographic signal data, and pain scores of the patient's lower limbs; Based on the data, candidate variables such as stride length, stride frequency, gait asymmetry index, electromyographic activation sequence consistency index, median frequency, and COP trajectory features were extracted. LASSO regression was used to screen out gait asymmetry index, electromyographic activation sequence consistency index, and pain score as key features in the comprehensive rehabilitation score prediction task, and a functional grading model was constructed to obtain the functional level. A multi-objective optimization problem is constructed with the objectives of minimizing pain score, maximizing joint range of motion and consistency of electromyographic activation sequence, and minimizing muscle fatigue, while constraining the articular cartilage contact stress to not exceed the limit. The problem was solved using the NSGA-III algorithm to obtain the Pareto optimal solution set for the treadmill weight reduction ratio and speed. Using the real-time patient status as input, the PPO strategy network selects and dynamically adjusts the treadmill parameters in the Pareto solution set. The adjusted parameters are sent to the runway actuators in real time.

2. The method according to claim 1, characterized in that, The gait kinematics data is collected by an IMU unit worn on the patient's lower limbs, and the plantar pressure distribution data is collected by an array of pressure sensors laid on the treadmill floor.

3. The method according to claim 1, characterized in that, The electromyographic signal data was acquired using surface electromyography sensors deployed on the quadriceps femoris, hamstrings, and tibialis anterior muscles. The RMS value, root mean square frequency (MF), and activation sequence were extracted to calculate a muscle activation consistency index. Muscle fatigue spectrum index .

4. The method according to claim 1, characterized in that, The pain score mentioned above is a subjective score of 0-10 points given by the patient in real time during the training process using the visual analog scale, and is collected at preset time intervals.

5. The method according to claim 1, characterized in that, The gait asymmetry index is specifically: ； ； and The asymmetry coefficient representing gait; and These represent the shorter and longer average support phase times for the left and right feet during walking, respectively. and These represent the shorter and longer values ​​in the mean time of each oscillation phase, respectively.

6. The method according to any one of claims 1-5, characterized in that, The functional grading model maps a gait asymmetry index close to zero, high consistency of electromyographic activation sequence, and low pain score to a high functional level, and vice versa.

7. The method according to claim 1, characterized in that, To minimize VAS pain score, maximize joint range of motion, and muscle activation consistency index Minimizing the fatigue index is a multi-objective optimization function: ； ； x = [r, v] represents the combination of training parameters to be optimized, where: r represents the weight reduction ratio on the treadmill, and v represents the treadmill speed; The weighting coefficients for the target term; Subjective pain score, with the goal of minimizing it; The goal is to maximize the range of motion of the joint. Muscle activation sequence consistency index, the goal is to maximize; Muscle fatigue spectrum index, the goal is to minimize it; : Cartilage contact stress under the selected parameters; : The maximum safety threshold set.

8. The method according to any one of claims 1-7, characterized in that, The reward function for the PPO policy network is: ； Where VAS represents the pain score at the current step, and ROM represents the estimated range of motion of the joint. Indicators representing the consistency of muscle activation sequence Stress Penalty indicates the degree of muscle fatigue; it is a severe penalty inflicted when cartilage stress exceeds a certain threshold. The weighting coefficients are used to balance the dimensions.

9. The method according to claim 1, characterized in that, The treadmill parameter adjustment actions are discretized as: weight loss percentage ±2%, speed ±0.5 km / h, training time ±5 min, and rest interval ±5 min.

10. The method according to claim 1, characterized in that, The maximum long-term expected return of the PPO reinforcement learning strategy is: ； in, This means that the goal is to maximize the expected cumulative reward by adjusting the weight parameters θ of the policy network, where k represents the k-th step from the current time t, and γ is the discount factor, ranging from 0 < γ ≤ 1. This represents the instantaneous reward value obtained at the k-th step in the future.

Images