Personalized precision training method based on human lower limb muscle model

Abstract

The present invention relates to the field of medical rehabilitation technology, and in particular to a personalized precision training method based on a human lower limb muscle model. Through the collection and processing of relevant data after clustering of experimenters, the muscle characteristics of the experimenters in the whole training process are analyzed, and a second clustering is performed with similarity as a target. The analyzed muscle characteristics are combined with a training program to design an autonomous learning optimization network, and the coefficients are continuously corrected until a training program configuration with the highest similarity in training effects is generated among experimenters of the same category, and the training effect is identified as an expected curve under the training program configuration; new trainees are trained using the classified training program, and a weighted least squares optimization process is performed on an actual muscle force change curve in the whole process and an expected muscle force change curve under the category, and the result is regarded as an error and the variable coefficient is corrected by the autonomous learning optimization network through feedback control until the error is infinitely close to 0, thereby providing reference value for personalized muscle training.

Description

Technical Field

[0001] The present invention relates to the field of medical rehabilitation technology, and in particular to a personalized precision training method based on a human lower limb muscle model. Background Art

[0002] With the improvement of quality of life, physical fitness is becoming more and more popular, from teenagers to middle-aged and elderly people to strengthen muscle training to avoid or alleviate the occurrence of some diseases. In addition, the rehabilitation of people with limb movement disorders due to various reasons has become a social problem that needs to be solved urgently. According to the relevant literature and patents, there is a lack of special training designed for muscle changes caused by human individuality in both sports fitness and muscle rehabilitation training. These situations are also common in life. For example, when exercising muscle strength, sports students do not observe the changes in their own muscle strength during training more intuitively, which leads to muscle injuries.

[0003] Due to the individual differences in human movement, different people show different muscle characteristics during training. The same training method is not suitable for every trainee. Even if the same training program is used for the same type of people, the training results will be different or deviate from the expected results. To eliminate this deviation, a personalized training program needs to be designed. However, there is almost no accurate training method based on the human lower limb muscle model. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a personalized and precise training method based on the human lower limb muscle model. By designing a personalized training plan, the trainee can not only reduce the choice of training plans and save a lot of time, but also greatly improve the effectiveness of training.

[0005] The present invention solves the above technical problems by the following technical means:

[0006] The personalized precision training method based on the human lower limb muscle model includes the following steps:

[0007] S10. Select subjects for muscle training and perform the first clustering based on gender, age, height, and weight;

[0008] S20. Use multi-axis inertial sensors and plantar pressure sensors to collect and process relevant data during training;

[0009] S30. Establishing a human lower limb musculoskeletal model that conforms to human kinematics and dynamics, and performing muscle characteristic analysis on the human lower limb muscle model;

[0010] S40. Perform a second clustering based on muscle characteristics analysis among subjects of the same category;

[0011] S50. Based on the second clustering result and in combination with the training scheme, the autonomous learning optimization network is designed, and the coefficients are continuously modified until a training scheme configuration with the highest similarity in training effects is produced among the experimenters of the same category, and the training effect is identified as the expected curve under the training scheme configuration;

[0012] S60. After the new trainees are classified, they are trained using the training plan of the corresponding category after classification, and the actual muscle force change curve in the whole process is optimized by weighted least squares with the expected muscle force change curve under this category. The result is regarded as the error and the variable coefficient is corrected by the autonomous learning optimization network through feedback control until the error is infinitely close to 0.

[0013] Further, the first clustering comprises the following steps:

[0014] S11. Generate a certain number N of quasi-random numbers in the area where the cluster samples are located;

[0015] S12. Find k neighboring values ​​of all quasi-random number points;

[0016] S13. Calculate the weight coefficient of each quasi-random point based on the k-neighbor distance, as follows:

[0017]

[0018] Among them, r i represents the average k-nearest neighbor distance of the i-th quasi-random point;

[0019] S14. Perform weighted k-means clustering on the quasi-random number points to obtain an initial cluster center u;

[0020] S15. According to the initial cluster center mean vector u={u 1 ,u 2 ,...,u k}, calculate the sample x j With each mean vector u i The distance is as follows:

[0021] d ji =||x j -u i ||, (1≤i≤k);

[0022] S16. Determine x based on the nearest mean vector j Cluster markers:

[0023] γ j =argmin i∈{1,2,...,k} d ji ;

[0024] The sample xj Assign to the corresponding clusters:

[0025]

[0026] Calculate the new mean vector for the cluster samples divided for the first time:

[0027]

[0028] If u i '≠u i , then update the current mean vector to u i ', otherwise, keep the current mean vector unchanged, repeat steps S15 and S16 until the current mean vector is not updated, and divide the clusters:

[0029] C={C 1 ,C 2 ,...,C k}.

[0030] Further, the step S20 includes the following steps:

[0031] S21. The clustered subjects wear inertial sensors and stand on the three-dimensional plantar pressure sensor to collect relevant data;

[0032] S22. Collect plantar pressure data using a three-dimensional plantar pressure sensor and process the data using an SG filtering algorithm;

[0033] S23. Use multi-axis inertial sensors to collect kinematic data, convert the kinematic data into quaternions and perform Kalman filtering. Update the system's time and state information through cyclic iteration of the equation, and then estimate the optimal estimate value at each moment.

[0034] Furthermore, in step S22, the plantar pressure is mainly divided by region, and the plantar pressure data of different regions are subjected to the same data processing, and the 2n+1 data before and after a certain moment of all the data in the region are filtered and fitted with a k-1 order polynomial.

[0035] Furthermore, the kinematic data in step S23 includes acceleration, angular velocity and magnetometer data.

[0036] Further, the step S30 includes the following steps:

[0037] S31. Establish a human lower limb musculoskeletal model with virtual sensors that conforms to human kinematics and dynamics;

[0038] S32. The multi-axis inertial sensor is used to collect and process the quaternion Q = (q 0 ,q1 ,q 2 ,q 3 ), converting the sensor coordinate system in quaternion form and represented by direction cosine matrix into the rotation matrix of the reference coordinate system;

[0039] S33. Perform residual reduction processing on the human lower limb musculoskeletal model, and apply a set of weighting factors w to all degrees of freedom of the human lower limb musculoskeletal model. q ;

[0040] S34. A set of muscle excitation points are calculated by combining proportional differential control and static optimization to drive the generalized coordinates of the dynamic human lower limb musculoskeletal model toward the desired motion trajectory.

[0041] Furthermore, the residual reduction applies a set of weighting factors w to all degrees of freedom of the human lower limb musculoskeletal model. q The steps are as follows:

[0042] Limiting the amount of residual forces and moments applied to the human lower extremity musculoskeletal model is achieved by the following formula:

[0043] Min:J=K err +R err

[0044] In the formula, J represents the modified objective function, K err ,R err represents the kinematic measurement errors and residuals,

[0045]

[0046]

[0047] In the formula, w J ,w K ,w L ,w M They represent the root mean square error between the experimental and simulated pelvis, pelvis and lower limb rotation, waist and upper limb rotation, and residual force and moment, respectively. i ,R j represent the kinematic degrees of freedom and residual forces / torques, respectively.

[0048] Furthermore, the step S40 is to perform a second clustering based on the muscle characteristic parameters of the subject during the entire training process and taking the similarity of all curves compared pairwise as a target.

[0049] Further, the step S50 is as follows:

[0050] Select a set of initial training plans for all subjects and configure a set of variable coefficients for the training plans. Design an autonomous learning optimization network to optimize the variable coefficients until the muscle force change curves detected by subjects of the same type under the same training plan configuration are almost the same, that is, the training effect of this training plan configuration on subjects of the same type is the most similar, and the obtained change curve is identified as the optimal expected curve F under this training plan configuration. E .

[0051] Further, the step S60 is as follows:

[0052] For new trainees who have never been trained and tested, the data processing and analysis are carried out in the same way as the experimental subjects, and the trainees are classified into the existing experimental subject categories. According to the training program designed for this category, the trainees are trained and the actual muscle force change curve F is detected. R ;

[0053] The actual muscle force change curve F of the trainee collected R The expected muscle force change curve F obtained for this category E Do weighted least squares processing, that is:

[0054]

[0055] Consider e as the error between actual and expected, and modify the variable coefficient of the training scheme through feedback until the error e→0.

[0056] The personalized precision training method based on the human lower limb muscle model of the present invention collects and processes the relevant data of the experimenter after clustering, analyzes the muscle characteristics (i.e. muscle strength) of the experimenter during the whole training process, and analyzes the muscle characteristics (i.e. muscle strength) of the experimenter during the whole training process. ), and the similarity after comparing the curves pairwise is used as the target for the second clustering. According to the analyzed muscle characteristics and combined with the training plan, the autonomous learning optimization network is designed, and the coefficients are continuously corrected until the training plan configuration with the highest similarity in training effects is produced among the experimenters of the same category, and its training effect is identified as the expected curve under the training plan configuration; after the new trainees are classified, they are trained using the training plan of the classified category, and the actual muscle force change curve in the whole process is optimized by weighted least squares with the expected muscle force change curve under the category, and the result is regarded as an error and the autonomous learning optimization network is affected by feedback control to correct the variable coefficient until the error is infinitely close to 0. The personalized training plan designed by the present invention can not only reduce the choice of training plans for trainees and save a lot of time for trainees, but also greatly improve the effectiveness of training, providing reference value for personalized muscle training. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1It is a training flow chart of the personalized precision training method based on the human lower limb muscle model of the present invention;

[0058] Figure 2 It is a human lower limb muscle characteristic analysis diagram of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the accompanying drawings of the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The personalized precision training method based on the human lower limb muscle model of the present invention collects and processes the relevant data of the experimenter after clustering, analyzes the muscle characteristics (i.e. muscle strength) of the experimenter during the whole training process, and analyzes the muscle characteristics (i.e. muscle strength) of the experimenter during the whole training process. ), and the similarity after comparing the curves pairwise is used as the target for the second clustering. According to the analyzed muscle characteristics and combined with the training plan, an autonomous learning optimization network is designed, and the coefficient is continuously corrected until a training plan configuration with the highest similarity in training effects is produced among the experimenters of the same category, and its training effect is identified as the expected curve under the training plan configuration; after the new trainees are classified, they are trained using the training plan of the classified category, and the actual muscle force change curve in the whole process is optimized by weighted least squares with the expected muscle force change curve under the category. The result is regarded as an error and affects the autonomous learning optimization network to correct the variable coefficient through feedback control until the error is infinitely close to 0.

[0061] The personalized precision training method based on the human lower limb muscle model of the present invention comprises the following steps:

[0062] S10. Select subjects for muscle training and perform the first clustering based on gender, age, height, and weight;

[0063] S20. Use multi-axis inertial sensors and plantar pressure sensors to collect and process relevant data during training;

[0064] S30. Establishing a human lower limb musculoskeletal model that conforms to human kinematics and dynamics, and performing muscle characteristic analysis on the human lower limb muscle model;

[0065] S40. Perform a second clustering based on muscle characteristics analysis among subjects of the same category;

[0066] S50. Based on the second clustering result and in combination with the training scheme, the autonomous learning optimization network is designed, and the coefficients are continuously modified until a training scheme configuration with the highest similarity in training effects is produced among the experimenters of the same category, and the training effect is identified as the expected curve under the training scheme configuration;

[0067] S60. After the new trainees are classified, they are trained using the training plan of the category after classification, and the actual muscle force change curve in the whole process is optimized by weighted least squares with the expected muscle force change curve under this category. The result is regarded as the error and the variable coefficient is corrected by the autonomous learning optimization network through feedback control until the error is infinitely close to 0.

[0068] Furthermore, the personalized precision training method based on the human lower limb muscle model of the present invention is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0069] The subjects were selected for muscle training. Before training, they were classified by gender and then divided according to their height (cm) x 1 、Age (years) x 2 , body weight (kg) x 3 The experimenters are clustered for the first time to reduce the training error. The clustering algorithm K_means is used, and the data set for clustering the experimenters is D = {X 1 ,X 2 ,...,X n},X i ={x 1 ,x 2 ,x 3},From the literature review, combined with factors such as algorithm complexity, number of clustering algorithm iterations, and clustering error, the Quasi-Monte Carlo algorithm (QMC) is used to select the initial clustering center u.

[0070] S11. Generate a certain number N=2000 of quasi-random numbers in the area where the cluster samples are located;

[0071] S12. Find k neighboring values ​​of all quasi-random number points;

[0072] S13. Calculate the weight coefficient of each quasi-random point based on k neighbor distances;

[0073]

[0074] Among them, r i represents the average k-nearest neighbor distance of the i-th quasi-random point.

[0075] S14. Perform weighted k-means clustering on the quasi-random number points to obtain the initial cluster center u.

[0076] S15. According to the initial cluster center mean vector u={u 1 ,u 2 ,...,u k}, calculate the sample x j With each mean vector u i Distance:

[0077] d ji =||x j -u i ||, (1≤i≤k)

[0078] S16. Determine x based on the nearest mean vector j Cluster markers:

[0079] γ j =argmin i∈{1,2,...,k} d ji

[0080] The sample x j Assign to the corresponding clusters:

[0081]

[0082] Calculate the new mean vector for the cluster samples divided for the first time:

[0083]

[0084] If u′ i ≠u i , then update the current mean vector to u′ i Otherwise, keep the current mean vector unchanged, repeat steps S15 and S16 until the current mean vector is not updated, and divide the clusters:

[0085] C={C 1 ,C 2 ,...,C k}.

[0086] S21. The clustered subjects wear inertial sensors and stand on the three-dimensional plantar pressure sensor to collect relevant data.

[0087] S22. Use a three-dimensional plantar pressure sensor to collect plantar pressure data and use the SG filtering algorithm to process the data. Plantar pressure is mainly divided by region, and the plantar pressure data of different regions are processed in the same way. The 2n+1 data before and after a certain moment in all the data in the region are filtered and fitted with a k-1 order polynomial.

[0088] For the data at a certain time t, the following formula is used for fitting:

[0089] xt =a 0 +a 1 *t+a 2 t 2 +...+a k-1 t k-1

[0090] Similarly, all the data before and after the moment are calculated using the above formula, and 2n+1 formulas are obtained to form a matrix:

[0091]

[0092] in, represents the minimum value that satisfies the above polynomial equation, which conforms to the least squares optimization problem.

[0093] In order to make the entire matrix have a solution, it must satisfy 2n+1>k, so that the parameter a can be determined by the least squares method. 0 ,a 1 ,a 2 ,...,a k-1 . Simplify the above matrix to get:

[0094] X (2n+1)×1 =T (2n+1)×k A k×1 +E (2n+1)×1

[0095] The subscripts of each parameter represent their respective dimensions. By using the least squares method, A can be obtained. k×1 The solution is:

[0096] A=(T T ·T) -1 ·T T ·X

[0097] Therefore, the data after model filtering is:

[0098] P=T·A=T·(T T ·T) -1 ·T T ·X=B·X

[0099] Where B = T·(T T ·T) -1 ·T T .

[0100] S23. Use a multi-axis inertial sensor to collect kinematic data, convert the kinematic data into quaternion and perform Kalman filtering.

[0101] According to strapdown inertial navigation theory, quaternion is defined as:

[0102]

[0103] In the formula, q 0 ,q 1 ,q 2 ,q 3 is a real number, i, j, k are mutually orthogonal unit vectors. The rotation matrix from the sensor coordinate system b to the reference coordinate system n in quaternion form is:

[0104]

[0105] Update the quaternion in real time through the quaternion differential equation:

[0106]

[0107] Where T is the sampling interval; w x ,w y ,w z It is the three-axis output value of the gyroscope.

[0108] The acceleration, angular velocity and magnetometer data collected by the inertial sensor are fused by extended Kalman filtering, and the process of calculating the quaternion based on the angular velocity data is used as the state, that is, X = [q 0t q 1t q 2t q 3t ] T , the state space model is designed as:

[0109] X t+1 =f t (X t ,y w,t ,e w,t )

[0110] Z t =h t (X t )+e t

[0111] Among them, f t (·) indicates that it depends on the quaternion state at time t and the gyroscope measurement value y w,t , and gyro measurement noise e w,t ~N(0,∑ w ) nonlinear time-varying model; h t (·) represents the measurement model that depends on the quaternion state at time t; e t ~N(0,R) represents the measurement noise of the accelerometer and magnetometer.

[0112] Time update process:

[0113]

[0114] Where T is the sampling time, exp represents the matrix exponent, and w t Represents the gyroscope measurement value. Since the sampling time is very short, the above formula can be equivalently converted to:

[0115]

[0116] Among them, I 4 represents the 4th-order identity matrix.

[0117]

[0118] P t+1|t =A t P t|t A t T +G t QG t T

[0119] in, Q=∑ w

[0120] Measurement Update:

[0121] X t|t =X t|t-1 +K g (Z t -H t Z t|t-1 )

[0122] in,

[0123]

[0124] R t|t-1 Indicates the rotation of the coordinate system from time t-1 to time t, g n ,m n They represent the gravitational acceleration and magnetometer measurement data in the global coordinate system respectively.

[0125] P t|t =(IK g (t)H t ) t|t-1

[0126] The system's time and state information are updated through the iterative loop of the equation, and then the optimal estimate value at each moment is estimated.

[0127] According to the relevant literature on human musculoskeletal system, a personalized human lower limb musculoskeletal model with virtual sensors and in line with human kinematics and dynamics is established, such as Figure 2As shown. The inertial sensor (IMU) collects and transforms it into a quaternion Q = (q 0 ,q 1 ,q 2 ,q 3 ) data, positioning the model in the pose that "best matches" the experimental IMU, and the "best match" represents the pose with the smallest weighted squared direction error. Since the sensor coordinate system and the reference system can be rotated to coincide, assuming the rotation order is ZXY, rotate in this way, and convert the sensor coordinate system represented by the direction cosine matrix into the rotation matrix of the reference coordinate system:

[0128]

[0129] When performing attitude calculation based on quaternions, assign initial values ​​to the quaternions:

[0130]

[0131]

[0132]

[0133]

[0134] Since the direction cosine matrix is ​​equal to the rotation matrix represented by the quaternion and the attitude angle is calculated according to the quaternion after Kalman filtering by the following formula:

[0135]

[0136] Among them, the weighted least squares problem solved by inertial sensors (IMU) combined with inverse kinematics (IK) is:

[0137]

[0138] In the formula, q represents the quaternion, w i is the weight corresponding to each IMU orientation, θ i is the angular component of the orientation error represented by an axis angle.

[0139] Often, modeling assumptions, noise, and other errors in the data from motion capture result in dynamic inconsistencies, essentially ground reaction forces and accelerations estimated from the measured object kinematics that do not satisfy Newton’s second law, which states:

[0140] F=ma

[0141] Due to the noise of the model sensor and joint angle data and the inaccuracy of the model geometry and mass distribution, there are a lot of non-physical forces in the model. To satisfy Newton's second law of motion, the residual needs to be corrected:

[0142] F+F residual =ma

[0143] In the formula, F residual Represents the force added by the model to eliminate residuals.

[0144] Therefore, the model needs to be subjected to residual reduction. The purpose of residual reduction is to minimize the effects of modeling and model sensor data processing errors that aggregate and result in large non-physical compensation forces called residuals. Specifically, residual reduction can change the torso center of mass of the custom model and allow the kinematics from the inverse kinematics model based on inertial sensor data to change to be more dynamically consistent with the ground reaction force data on the body.

[0145] Residual reduction (RRA) applies a set of weighting factors w to all degrees of freedom (DOF) of the musculoskeletal model. q , to reduce the difference between metrology and model kinematics and to limit the amount of residual forces and moments applied to the model, as achieved by the following formula:

[0146] Min:J=K err +R err

[0147] In the formula, J represents the modified objective function, K err ,R err Represents kinematic measurement errors and residuals.

[0148]

[0149]

[0150] In the formula, w J ,w K ,w L ,w M They represent the root mean square error between the experimental and simulated pelvis, pelvis and lower limb rotation, waist and upper limb rotation, and residual force and moment, respectively. i ,R j represent the kinematic degrees of freedom and residual forces / torques, respectively.

[0151] In order to make the dynamic musculoskeletal model track the desired action under a given external force, computational muscle control (CMC) is a method of estimating actuator control based on feedback control, so that the actuator produces motion that tracks the desired trajectory (i.e., the joint angle calculated based on experimental data). Specifically, it is achieved by combining proportional-differential control and static optimization to calculate a set of muscle excitation points, driving the generalized coordinates (e.g., joint angles) of the dynamic musculoskeletal model toward the desired motion trajectory.

[0152] Calculate a set of desired accelerations to drive the model coordinates toward the experimentally derived coordinates. Use the following PD control law to calculate the desired accelerations

[0153]

[0154] In the formula, Represent the feedback gains of velocity and position error respectively, Denote the experimental coordinates and model coordinates, respectively. Since the force exerted by the muscle on the body cannot change instantaneously, the required acceleration is calculated over a short period of time T in the future. For musculoskeletal models, the time interval is required to be short enough and fully controllable when the muscle force changes. T is usually selected to be about 0.01 seconds. If these expected accelerations are achieved, the errors between the model coordinates and the experimentally derived coordinates are driven to zero. In order to reduce these errors to zero, the velocity gain can be selected using the following relationship:

[0155]

[0156] Usually, the controller generalized force f is calculated by static optimization:

[0157]

[0158]

[0159] Where x represents the controller. When the actuator force is torque, f = τ, the controller is now an excitation, and the torque generation has muscle dynamics:

[0160]

[0161] Where τ represents the joint torque, A(q) represents the muscle moment arm matrix, and f m represents the function related to muscle activity a(x), r represents the torque actuator, and l represents the muscle length. The muscle force limit is obtained by the following formula:

[0162]

[0163]

[0164]

[0165] In the formula, f m represents muscle force, and R represents residual. To obtain muscle activity a(x), we can get the following value based on muscle force:

[0166]

[0167] According to the above calculations, we can analyze and get the desired muscle parameters. All curves are compared pairwise, and the second clustering is performed with similarity as the goal. A set of initial training programs R is selected for all subjects, and a set of variable coefficients is configured for the training programs. An autonomous learning optimization network (NET1) is designed to optimize the variable coefficients until the muscle force change curves detected by subjects of the same type under the same training program configuration R' are almost the same, that is, the training effect of this new training program configuration on subjects of the same type is the most similar, and the obtained change curve is identified as the best expected curve F under this training program configuration. E .

[0168] At the same time, an autonomous learning optimization network (NTE2) affected by errors is added on the basis of network NET1. For new trainees who have never been trained and tested, data processing and analysis are carried out in the same way as for experimental subjects, and the trainees are classified into the existing experimental subject category. According to the training program designed for this category, the trainees are trained and the actual muscle force change curve F is detected. R .

[0169] The actual muscle force change curve F of the trainee collected R The expected muscle force change curve F obtained for this category E Do weighted least squares processing, that is:

[0170]

[0171] Consider e as the error between actual and expected, and modify the variable coefficient of the training scheme through feedback until the error e→0.

[0172] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention. The techniques, shapes, and structural parts not described in detail in the present invention are all known technologies.

Claims

1. Personalized and precise training method based on human lower limb muscle model, It is characterized in that The following steps are involved: S10. Select subjects for muscle training and perform the first clustering based on gender, age, height, and weight; S20. Use multi-axis inertial sensors and plantar pressure sensors to collect and process relevant data during training; S30. Establishing a human lower limb musculoskeletal model that conforms to human kinematics and dynamics, and performing muscle characteristic analysis on the human lower limb muscle model; S40. Perform a second clustering based on muscle characteristics analysis among subjects of the same category; S50. Based on the second clustering result and in combination with the training scheme, the autonomous learning optimization network is designed, and the coefficients are continuously modified until a training scheme configuration with the highest similarity in training effects is produced among the experimenters of the same category, and the training effect is identified as the expected curve under the training scheme configuration; S60. After the new trainees are classified, they are trained using the training program of the corresponding category after classification, and the actual muscle force change curve during the whole process is optimized by weighted least squares with the expected muscle force change curve under the category. The result is regarded as the error and the variable coefficient is corrected by the autonomous learning optimization network through feedback control until the error converges; The step S30 includes the following steps: S31. Establish a human lower limb musculoskeletal model with virtual sensors that conforms to human kinematics and dynamics; S32. The multi-axis inertial sensor is used to collect and process the quaternion Q = (q 0 ,q 1 ,q 2 ,q 3 ), converting the sensor coordinate system in quaternion form and represented by direction cosine matrix into the rotation matrix of the reference coordinate system; S33. Perform residual reduction processing on the human lower limb musculoskeletal model, and apply a set of weighting factors w to all degrees of freedom of the human lower limb musculoskeletal model. q ; S34. A set of muscle excitation points are calculated by combining proportional differential control and static optimization to drive the generalized coordinates of the dynamic human lower limb musculoskeletal model toward the desired motion trajectory.

2. The personalized precision training method based on the human lower limb muscle model according to claim 1, It is characterized in that The first clustering comprises the following steps: S11. Generate a certain number N of quasi-random numbers in the area where the cluster samples are located; S12. Find k neighboring values ​​of all quasi-random number points; S13. Calculate the weight coefficient of each quasi-random point based on the k-neighbor distance, as follows: Among them, r i represents the average k-nearest neighbor distance of the i-th quasi-random point; S14. Perform weighted k-means clustering on the quasi-random number points to obtain an initial cluster center u; S15. According to the initial cluster center mean vector u={u 1 ,u 2 ,...,u k }, calculate the sample x j With each mean vector u i The distance is as follows: d ji =||x j -u i ||,(1≤i≤k); S16. Determine x based on the nearest mean vector j Cluster markers: <h2 style=";text-align:left;direction:ltr">γ<h2 style=";text-align:left;direction:ltr"> j <h2 style=";text-align:left;direction:ltr"> =argmin<h2 style=";text-align:left;direction:ltr"> i∈{1,2,...,k} <h2 style=";text-align:left;direction:ltr"> d<h2 style=";text-align:left;direction:ltr"> ji <h2 style=";text-align:left;direction:ltr"> ; The sample x j Assign to the corresponding clusters: Calculate the new mean vector for the cluster samples divided for the first time: If u i '≠u i , then update the current mean vector to u i ', otherwise, keep the current mean vector unchanged, repeat steps S15 and S16 until the current mean vector is not updated, and divide the clusters: C={C 1 ,C 2 ,...,C k }。 3. The personalized precision training method based on the human lower limb muscle model according to claim 2, It is characterized in that The step S20 includes the following steps: S21. The clustered subjects wear inertial sensors and stand on the three-dimensional plantar pressure sensor to collect relevant data; S22. Collect plantar pressure data using a three-dimensional plantar pressure sensor and process the data using an SG filtering algorithm; S23. Use multi-axis inertial sensors to collect kinematic data, convert the kinematic data into quaternions and perform Kalman filtering. Update the system's time and state information through cyclic iteration of the equation, and then estimate the optimal estimate value at each moment.

4. The personalized precision training method based on the human lower limb muscle model according to claim 3, It is characterized in that In step S22, the plantar pressure is mainly divided by region, and the plantar pressure data of different regions are subjected to the same data processing. The 2n+1 data before and after a certain moment of all the data in the region are filtered and fitted with a k-1 order polynomial.

5. The personalized precision training method based on the human lower limb muscle model according to claim 4, It is characterized in that The kinematic data in step S23 includes acceleration, angular velocity and magnetometer data.

6. The personalized precision training method based on the human lower limb muscle model according to claim 1, It is characterized in that The residual reduction applies a set of weighting factors w to all degrees of freedom of the human lower limb musculoskeletal model. q The steps are as follows: Limiting the amount of residual forces and moments applied to the human lower extremity musculoskeletal model is achieved by the following formula: My:J=K err +R err In the formula, J represents the modified objective function, K err ,R err represents the kinematic measurement errors and residuals, In the formula, w J ,w K ,w L ,w M They represent the root mean square error between the experimental and simulated pelvis, pelvis and lower limb rotation, waist and upper limb rotation, and residual force and moment, respectively. i ,R j represent the kinematic degrees of freedom and residual forces / torques, respectively.

7. The personalized precision training method based on the human lower limb muscle model according to claim 6, It is characterized in that The step S40 is to perform a second clustering based on the muscle characteristic parameters of the experimenter during the entire training process and taking the similarity of all curves compared pairwise as the target.

8. The personalized precision training method based on the human lower limb muscle model according to claim 7, It is characterized in that The S50 steps are as follows: Select a set of initial training plans for all subjects and configure a set of variable coefficients for the training plans. Design an autonomous learning optimization network to optimize the variable coefficients until the muscle force change curves detected by subjects of the same type under the same training plan configuration are almost the same, that is, the training effect of this training plan configuration on subjects of the same type is the most similar, and the obtained change curve is identified as the optimal expected curve F under this training plan configuration. E .

9. The personalized precision training method based on the human lower limb muscle model according to claim 8, It is characterized in that The S60 steps are as follows: For new trainees who have never been trained and tested, the data processing and analysis are carried out in the same way as the experimental subjects, and the trainees are classified into the existing experimental subject categories. According to the training program designed for this category, the trainees are trained and the actual muscle force change curve F is detected. R ; The actual muscle force change curve F of the trainee collected R The expected muscle force change curve F obtained for this category E Do weighted least squares processing, that is: Consider e as the error between actual and expected, and modify the variable coefficient of the training scheme through feedback until the error e→0.

Images