A Clinical Teaching and Training Method for Postpartum Rehabilitation Based on Artificial Intelligence

By using multi-rigid-body dynamics models and real-time motion electromyography data to drive simulation, combined with augmented reality technology and force feedback gloves, the problem of not being able to obtain musculoskeletal status in real time in existing postpartum rehabilitation training is solved. This enables personalized visual and tactile feedback, protects vulnerable postpartum tissues, and improves training effectiveness.

CN123120198APending Publication Date: 2026-11-17DAZHOU VOCATIONAL & TECH COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611630216.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing postpartum rehabilitation training methods cannot obtain and utilize the internal mechanical state of the musculoskeletal system in real time, and lack personalized, dynamic, adaptive visual guidance and tactile feedback, resulting in inappropriate training intensity or inaccurate movement selection, and failing to accurately protect vulnerable postpartum tissues.

Method used

Simulations are driven by multi-rigid-body dynamics models and real-time electromyography data, combined with augmented reality technology and force feedback gloves to obtain the musculoskeletal status of postpartum women in real time and provide personalized visual and tactile feedback, dynamically adjusting the training program.

Benefits of technology

It enables precise and personalized postpartum rehabilitation training, protects vulnerable postpartum tissues, avoids excessive stress, and improves the safety and effectiveness of training.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120198A_ABST
    Figure CN123120198A_ABST
Patent Text Reader

Abstract

This invention discloses an artificial intelligence-based clinical teaching and training method for postpartum rehabilitation, belonging to the field of postpartum rehabilitation training technology. The method acquires a set of rehabilitation baseline parameters for postpartum women, including resting tension of the pelvic floor muscles, interbrevior distance, and spinal curvature index, and extracts initial training sequences from a rehabilitation training prescription library accordingly. During training, spatial motion trajectory data of the trunk and limbs are collected using an inertial measurement unit array, and electromyographic activity data of the target muscle groups are collected using a surface electromyography electrode array. A multi-rigid-body dynamics model of the postpartum musculoskeletal system is constructed, and forward dynamics simulation is driven by the spatial motion trajectory data and electromyographic activity data to obtain the reaction torques of each joint and the muscle activation sequence. An augmented reality head-mounted display overlays three-dimensional images of virtual coach demonstration movements generated based on the initial training sequence, and an auxiliary force or resistance is applied to the postpartum woman's hands using a force feedback glove according to the reaction torques and muscle activation sequence.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of postpartum rehabilitation training technology, specifically to a clinical teaching and training method for postpartum rehabilitation based on artificial intelligence. Background Technology

[0002] Postpartum rehabilitation training aims to restore pelvic floor function, diastasis recti, and overall core stability in postpartum women. Current clinical teaching methods largely rely on one-on-one guidance and verbal correction from physical therapists. Therapists judge the effectiveness of a woman's movements by observing their appearance and provide manual assistance or resistance. This model has inherent limitations: therapists cannot perceive the actual exertion of the deep muscle groups in real time, nor can they quantify the load on each joint. Assessment based solely on appearance easily leads to missed compensatory movements, and the strength and timing of manual techniques depend on personal experience, resulting in poor repeatability. Furthermore, training programs often use a uniform pace, failing to individualize the design based on baseline parameters such as resting pelvic floor muscle tension, rectus abdominis distance, and spinal curvature, leading to inappropriate training intensity or incorrect movement selection for some women. Digital rehabilitation training methods commonly use inertial sensors or visual capture to record movements and display standard movements as virtual characters for postpartum women to imitate, or provide simple resistance through pressure feedback devices. However, these solutions lack real-time solutions for the internal biomechanical state of the musculoskeletal system, and cannot use the activation of deep muscles and joint torque as feedback quantities to apply to the training process. The guidance images of the virtual coach will not change dynamically with the actual biomechanical state of the mother. The correlation between tactile feedback and biomechanical state is weak, making it difficult to achieve precise movement guidance while protecting the fragile postpartum tissues.

[0003] This invention aims to address how to acquire and utilize the internal biomechanical state of the musculoskeletal system in real time to achieve precise tactile-assisted training guidance during postpartum rehabilitation training, and how to achieve dynamic adaptive personalized visual guidance based on the individual baseline characteristics of postpartum women. Summary of the Invention

[0004] This invention provides an artificial intelligence-based clinical teaching and training method for postpartum rehabilitation. The purpose is to use multi-rigid-body dynamics models and real-time motion electromyography data to drive simulation, solve joint reaction torques and muscle activation timing, and use them for force feedback control. At the same time, based on individual rehabilitation baselines, augmented reality virtual coaching guidance is generated and dynamically adjusted to achieve precise and personalized postpartum rehabilitation teaching and training.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention proposes a clinical teaching and training method for postpartum rehabilitation based on artificial intelligence, which integrates rehabilitation assessment, personalized prescription generation, motion capture and biomechanical simulation, augmented reality teaching and force feedback interaction to form a closed loop of intelligent training for postpartum women throughout the entire process.

[0006] In one technical solution of the present invention, the method acquires a set of postpartum rehabilitation baseline parameters for the mother, including pelvic floor muscle resting tension, rectus abdominis intermuscular distance, and spinal curvature index. Preferably, the shear wave velocity of the mother's pelvic floor muscles is measured using an ultrasonic shear wave elastography device and converted into a pelvic floor muscle resting tension value; the distance between the medial borders of the rectus abdominis muscles is measured using ultrasound B-mode imaging as the rectus abdominis intermuscular distance; and point cloud data of the mother's trunk surface is acquired using a three-dimensional optical scanner, and the curvature of the spine in the sagittal and coronal planes is calculated to obtain the spinal curvature index. By acquiring the above core baseline parameters through multimodal assessment, the initial state of the mother's postpartum rehabilitation can be objectively quantified, providing a data basis for subsequent personalized training.

[0007] Based on the baseline parameter set, the method extracts an initial training sequence from a rehabilitation training prescription library. This initial training sequence includes the temporal arrangement and intensity coefficients of multiple training phases. As a preferred embodiment of the invention, the resting tension of the pelvic floor muscles, the interabdominal muscle distance, and the spinal curvature index are discretized and encoded to generate a three-dimensional combined code. Using this three-dimensional combined code as an index, the corresponding training prescription record is queried in the index table of the rehabilitation training prescription library. From this, the temporal sequence table of the training phases and the intensity coefficient of each training phase are parsed, thereby generating the initial training sequence. This method utilizes a multi-dimensional feature combination matching mechanism to quickly generate a structured training plan adapted to the postpartum woman's current physical function, avoiding rehabilitation delays or sports injuries caused by inappropriate training plans.

[0008] During training, this method collects spatial motion trajectory data of the mother's trunk and limbs using an inertial measurement unit (IMU) array, and simultaneously collects electromyographic (EMG) activity data of the target muscle groups using a surface electromyography (SEMG) electrode array. Preferably, IMUs are fixed to the mother's pelvis, thoracic spine, left and right upper arms, left and right forearms, left and right thighs, and left and right lower legs. Each IMU contains a three-axis accelerometer and a three-axis gyroscope, and outputs raw data at a high sampling rate. Multiple IMUs are time-stamped using a wireless synchronization protocol, and data fusion is performed using an error state Kalman filter algorithm based on rigid body posture to obtain the absolute posture and spatial coordinates of each body segment. Simultaneously, SEMG electrodes are attached to the surfaces of the mother's rectus abdominis, external oblique, erector spinae, gluteus maximus, and pelvic floor muscles. Weak EMG signals are collected using a bipolar differential measurement mode and a high sampling rate, and the EMG activity data amplitude sequence is obtained after bandpass filtering and sliding root mean square (RMS) calculation. This multi-source sensor fusion scheme can accurately capture the kinematic and electromyographic information of the mother during training, providing high-fidelity driving data for dynamic analysis.

[0009] To establish a personalized biomechanical model for postpartum women, this method constructs a multi-rigid-body dynamic model of their musculoskeletal system. This model includes freedom constraints for each joint and muscle lever arm functions. In a preferred embodiment of this invention, standard bone dimensions and joint positions are extracted from a pre-defined human anatomy database as initial templates. These templates are then scaled and deformed according to the postpartum woman's height, weight, and body fat percentage to generate a personalized skeletal chain. The range of motion constraints for each joint are set as boundary conditions based on the measured joint range of motion of the postpartum woman. The origin and insertion coordinates of the major muscle groups are mapped onto the personalized skeletal chain, and a function table showing the muscle lever arm variation with joint angles is established. The multi-rigid-body dynamic model constructed in this way highly fits the individual anatomical characteristics and physiological range of motion of the postpartum woman, making subsequent simulation results closer to the actual stress state.

[0010] After obtaining motion data and a personalized model, spatial motion trajectory data and electromyographic (EMG) activity data are used as driving inputs to perform forward dynamics simulation on a multi-rigid-body dynamics model, obtaining the reaction torques of each joint and the muscle activation sequence. As a preferred approach, the posture quaternion sequence of each body segment is calculated using inertial measurement unit (IMU) data, and the joint angular positions and angular velocities between adjacent segments are derived. The processed EMG signals are converted into muscle activation envelopes, and muscle force vectors are calculated using a muscle lever arm function table and maximum isometric contraction force. The joint angular acceleration sequence and muscle force vectors are substituted into the Newton-Euler equations of the multi-rigid-body dynamics model to obtain the reaction torques of each joint, while the normalized activation sequence is used as the muscle activation sequence. This forward dynamics simulation process maps measured motion and muscle activity to the spatiotemporal distribution of internal joint loads, providing a basis for real-time assessment of training intensity and force feedback control.

[0011] In terms of visual instruction guidance, this method utilizes an augmented reality head-mounted display to overlay a 3D image of a virtual coach demonstrating movements onto the mother's field of vision. This 3D image is generated based on an initial training sequence. Preferably, based on the movement identifier of the current training segment in the initial training sequence, corresponding standard motion capture data is retrieved from the movement library and mapped onto the virtual coach's skeletal drive template to generate a 3D animation. The spatial positioning module of the augmented reality head-mounted display is used to obtain the mother's head position and orientation, and the 3D animation of the virtual coach is overlaid onto a predetermined position in the field of vision. Furthermore, the transparency of the virtual coach is dynamically adjusted based on the reaction torque and muscle activation sequence, reducing the transparency of the virtual coach when the reaction torque exceeds a preset threshold. This augmented reality teaching method couples standard movement demonstrations with actual movement states, allowing the mother to intuitively experience the key points of the movements from a first-person perspective during training, while simultaneously perceiving whether her exertion is excessive or incorrect through changes in transparency.

[0012] This method applies auxiliary force or resistance to the mother's hand using a force feedback glove based on the reaction torque and muscle activation sequence. In a preferred embodiment, the component of the reaction torque related to hand movement is extracted as the desired hand torque and coupled with the activation level of the hand muscles to generate a force feedback instruction set. Each finger of the force feedback glove is embedded with a miniature linear motor and a tension sensor. The output force of the miniature linear motor is controlled according to the instruction set, and the actual contact force between the finger and the finger is measured in real time by the tension sensor. A proportional-integral controller is used to adjust the actual contact force to the target force. This force feedback mechanism transforms the simulated internal load information into tactile cues, guiding the mother to train in the correct muscle coordination manner and promoting the reshaping of neuromuscular control patterns.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By constructing a multi-rigid-body dynamics model that includes constraints on the degrees of freedom of each joint and muscle lever arm functions, the spatial motion trajectory data of the trunk and limbs collected by the inertial measurement unit array and the electromyographic activity data collected by the surface electromyography electrode array are used as driving inputs to perform forward dynamics simulation, obtaining the reaction torque of each joint and the muscle activation sequence. The force feedback glove generates force feedback commands based on the hand-related components of the reaction torque and the activation degree of the hand muscles. Through a miniature linear motor, it applies auxiliary force or resistance to the mother's hand, corresponding in real time to her own musculoskeletal biomechanical state. This scheme makes tactile assistance no longer a preset fixed force, but a closed-loop adjustment derived from the desired torque calculated from the mother's own joint torque and muscle activation, and the real-time contact force. If compensatory forces or abnormal torques occur during training, the force applied by the glove will change accordingly, directly mapping the biomechanical state to tactile perception. This allows the mother to correct her movements in real time, avoiding excessive load on the pelvic floor and abdominal wall due to incorrect force patterns, protecting the weak tissues in postpartum recovery from a biomechanical perspective, and deepening the intrinsic experience of motor learning. Based on a set of rehabilitation baseline parameters consisting of resting pelvic floor muscle tension, rectus abdominis distance, and spinal curvature index, individualized initial training sequences are extracted from a rehabilitation training prescription library. These sequences are then used to generate three-dimensional images of virtual coach demonstration movements, which are overlaid on the mother's field of vision via an augmented reality head-mounted display. Simultaneously, based on the reaction torque and muscle activation sequence obtained from forward dynamics simulation, the transparency of the virtual coach image is dynamically adjusted. When the reaction torque exceeds a preset threshold, the transparency is reduced, forcing the mother to focus more on her own proprioception rather than visual imitation. This virtual guidance method, which integrates individual baseline adaptation with real-time biomechanical states, breaks away from the conventional model of statically playing animations from a motion library. The level of guidance automatically changes according to the mother's actual exertion—weakening external visual dependence and strengthening internal corrective awareness when there are movement deviations or abnormal biomechanical signals; providing clear demonstrations during normal training phases, effectively preventing the risk of secondary injury caused by blindly imitating standard images and ignoring abnormal internal loads, thus improving the individualized adaptability of clinical teaching and training and active rehabilitation participation. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a clinical teaching and training method for postpartum rehabilitation based on artificial intelligence; Figure 2 This is a flowchart of the construction process for a multi-rigid-body dynamics model of the maternal musculoskeletal system. Figure 3It is a curve showing the change in the lever arm of the main muscles corresponding to the hip joint flexion and extension angles; Figure 4 It is the curve showing the combined value of the reaction torque and the numerical change of the alpha channel. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 This invention provides an artificial intelligence-based clinical teaching and training method for postpartum rehabilitation. The overall implementation scheme of this method is as follows: A set of postpartum rehabilitation baseline parameters for the mother is obtained, including resting tension of the pelvic floor muscles, rectus abdominis distance, and spinal curvature index. An initial training sequence is extracted from a rehabilitation training prescription library based on the baseline parameter set. The initial training sequence includes the timing arrangement and intensity coefficients of multiple training stages. During training, spatial motion trajectory data of the mother's trunk and limbs are collected using an inertial measurement unit array, while electromyographic activity data of the target muscle groups are collected using a surface electromyography electrode array. A multi-rigid-body dynamics model of the mother's musculoskeletal system is constructed, including the degrees of freedom constraints of each joint and the muscle lever arm function. Using the spatial motion trajectory data and electromyographic activity data as driving inputs, a forward dynamics simulation is performed on the multi-rigid-body dynamics model to obtain the reaction torques of each joint and the muscle activation sequence. A three-dimensional image of a virtual coach demonstrating movements is superimposed on the mother's field of vision using an augmented reality head-mounted display. The three-dimensional image of the virtual coach's demonstration movements is generated based on the initial training sequence. The force feedback glove applies auxiliary force or resistance to the mother's hand based on the reaction torque and muscle activation sequence.

[0018] Example 1: In its specific implementation, this embodiment provides a detailed description of the process of obtaining the postpartum rehabilitation baseline parameter set for mothers and extracting initial training sequences from the rehabilitation training prescription library based on the baseline parameter set.

[0019] The baseline parameter set for postpartum recovery in postpartum women includes three parts: resting tension of the pelvic floor muscles, distance between the rectus abdominis muscles, and spinal curvature index. Shear wave velocity of the pelvic floor muscles is measured using an ultrasonic shear wave elastography device, and this velocity is converted into a resting tension value. The probe of the ultrasonic shear wave elastography device is placed in the perineum, aligned with the pelvic floor muscle group, and emits acoustic radiation force pulses to generate shear waves in the tissue. The propagation process of the shear waves is captured, and the shear wave velocity is calculated and recorded as... Shear wave velocity Converted to resting tension of pelvic floor muscles The conversion relationship is given by the following formula:

[0020] in, This indicates the resting tension of the pelvic floor muscles, expressed in centimeters of water column. This indicates the shear wave velocity, measured in meters per second. The first conversion factor is a constant obtained through in vitro tissue phantom calibration experiments. The value is 0.95; The second conversion factor is an exponent determined based on the power-law relationship between the elastic modulus of the pelvic floor muscles and the shear wave velocity. The value is 2.0. First conversion factor. The value is determined based on the following: In the isolated tissue phantom experiment, the phantom material is subjected to known different tension levels, and the shear wave velocity is measured simultaneously. The optimal value is obtained by fitting the data using the least squares method. It is 0.95. Second conversion factor. The value of is based on the following: According to research on the mechanical properties of biological soft tissue, the elastic modulus is proportional to the square of the shear wave velocity; therefore, the second conversion coefficient is determined as follows: The setting is 2.0. The distance between the medial borders of the rectus abdominis muscles in the parturient is measured using ultrasound B-mode imaging, and this distance is taken as the intermuscular distance. In practice, a linear array probe of the ultrasound B-mode imaging device is placed horizontally three centimeters above the umbilicus of the parturient, and the probe angle is adjusted until the structures of the rectus abdominis muscles and linea alba on both sides are clearly displayed. On the frozen sonogram, the pixel positions of the medial borders of the left and right rectus abdominis muscles are marked respectively. The Euclidean distance between the two pixels is calculated using the device's built-in ranging function to obtain the intermuscular distance, denoted as . The unit is millimeters. Measurements were repeated three times, and the average of the three measurements was taken as the final rectus abdominis distance. Point cloud data of the mother's torso surface was acquired using a 3D optical scanner. The curvature of the spine in the sagittal and coronal planes was calculated to obtain the spinal curvature index. The 3D optical scanner consisted of four depth sensors distributed around the mother, simultaneously acquiring torso surface data and generating a dense point cloud. After denoising and downsampling the point cloud data, principal component analysis was used to extract the long axis direction of the point cloud as the preliminary spinal orientation. Control points were selected at equal intervals along the preliminary spinal orientation. Ellipses were fitted to the torso cross-sectional point cloud at each control point, and the center points of the ellipses were connected to form the spinal centerline. The spinal centerline was projected onto the sagittal plane to obtain the sagittal spinal centerline projection. The sagittal spinal curvature was calculated using circular arc fitting, i.e., the reciprocal of the radius of the fitted arc, denoted as . Projecting the spinal centerline onto the coronal plane yields the coronal projection of the spinal centerline. Circular arc fitting is then used to calculate the coronal spinal curvature, denoted as . The spinal curvature index is calculated by measuring the sagittal spinal curvature. The square of the coronal spinal curvature The arithmetic square root of the sum of squares is obtained, denoted as . .

[0021] The process of extracting initial training sequences from the rehabilitation training prescription library based on the baseline parameter set is as follows. The resting tension of the pelvic floor muscles... Inter-rectus abdominis and spinal curvature index Discretize and encode separately. (Pelvic floor muscle resting tension) The discretization coding rule is: when the resting tension of the pelvic floor muscles is... When the water column is less than 15 centimeters, it is coded as the number 1; when the resting tension of the pelvic floor muscles is... When the water column is between 15 cm and 25 cm, it is coded as number 2; when the resting tension of the pelvic floor muscles... A water column greater than 25 centimeters is coded as the number 3. (Intermuscular distance between the rectus abdominis muscles) The discretization coding rule is: when the inter-abdomen distance... When the distance between the rectus abdominis muscles is less than 20 millimeters, it is coded as the letter A; when the distance between the rectus abdominis muscles is less than 20 millimeters, it is coded as the letter A. When the distance between the rectus abdominis muscles is between 20 and 30 millimeters, it is coded as the letter B; when the distance between the rectus abdominis muscles is between 20 and 30 millimeters... For values ​​greater than 30 millimeters, the code is the letter C. Spinal curvature index. The discretization coding rule is as follows: the 25th and 75th percentiles of the statistical distribution of spinal curvature index of a large-scale postpartum population, which are pre-stored in a rehabilitation training prescription database, are used as the dividing thresholds. When the spinal curvature index... When the value is less than the 25th percentile, it is coded with the letter X; when the spinal curvature index When the 25th percentile is between the 75th percentile, it is coded with the letter Y; when the spinal curvature index... For values ​​greater than the 75th percentile, the code is the letter Z. The three discrete codes are concatenated in a fixed order: pelvic floor muscle resting tension code, rectus abdominis distance code, and spinal curvature index code, to generate a three-dimensional combined code. This three-dimensional combined code is used as a query index for precise matching in the index table of the rehabilitation training prescription database. The index table uses the three-dimensional combined code as the primary key, with each three-dimensional combined code corresponding to a unique identifier for a training prescription record. When a record with the exact same three-dimensional combined code exists in the index table, the corresponding training prescription record is directly read; when no completely matching record exists, the Hamming distance between the generated three-dimensional combined code and all three-dimensional combined codes in the index table is calculated, and the training prescription record corresponding to the three-dimensional combined code with the smallest Hamming distance is selected. The training session sequence table and the intensity coefficient of each training session are parsed from the training prescription record. The training prescription record is stored in a structured data format, containing fields: number of training sessions, session identifier sequence, session duration sequence, and session intensity coefficient sequence. Each identifier in the stage identifier sequence corresponds to a preset training action type. Each value in the stage duration sequence represents the duration of the corresponding training stage in seconds. Each value in the stage intensity coefficient sequence is a floating-point number between zero and one, representing the ratio of the training intensity of the corresponding training stage to the maximum intensity of that stage. After parsing the training prescription record, the stage identifier sequence, stage duration sequence, and stage intensity coefficient sequence are combined to generate the initial training sequence. The initial training sequence is an ordered list, and each element in the list contains three data items: stage identifier, duration, and intensity coefficient.

[0022] Example 2: In specific implementation, please refer to Figure 2 The process of constructing a multi-rigid-body dynamics model of the musculoskeletal system of a parturient woman begins with the extraction of an initial template. Standard skeletal dimensions and joint positions are extracted from a pre-defined human anatomy database as the initial template. This database stores skeletal geometric parameters and joint position parameters for multiple standard human bodies. Skeletal geometric parameters include the length, inertia tensor, and center of mass position of each bone, while joint position parameters include the three-dimensional coordinates of the joint center in the local skeletal coordinate system. When extracting the initial template, standard human data of the same sex as the parturient woman from the pre-defined human anatomy database are selected as the baseline, and the skeletal geometric parameters and joint position parameters from this baseline are copied as the initial template.

[0023] In practice, the initial template is scaled and deformed based on the mother's height, weight, and body fat percentage to generate a personalized skeletal chain. The scaling method employs a segmented proportional scaling approach. The bones in the initial template are divided into multiple bone groups according to their respective body segments. These bone groups include the head and neck bone group, trunk bone group, left upper arm bone group, left forearm bone group, right upper arm bone group, right forearm bone group, left thigh bone group, left lower leg bone group, right thigh bone group, and right lower leg bone group. A scaling factor is calculated for each bone group. Taking the trunk bone group as an example, its scaling factor is... Determined by the following formula:

[0024] in, This represents the scaling factor for the trunk skeleton group; This represents the height-to-body ratio weighting factor, with a value of 0.5. This represents the body weight percentage weighting factor, with a value of 0.3. This represents the body fat percentage weighting factor, with a value of 0.2. This indicates the mother's height, in centimeters. This represents the height of the standard human body corresponding to the initial template, in centimeters. This indicates the mother's weight, expressed in kilograms. This represents the weight of the standard human body corresponding to the initial template, in kilograms; This indicates the mother's body fat percentage, expressed as a percentage. This represents the body fat percentage of the standard human body corresponding to the initial template, expressed as a percentage. , and The values ​​are determined as follows: height is the most important factor determining bone length, assigned the highest weight of 0.5; weight reflects overall circumference, assigned a weight of 0.3; body fat percentage affects subcutaneous fat distribution, thus altering limb shape, assigned a weight of 0.2. For limb skeleton groups, a limb length ratio correction factor is added when calculating the scaling factor. The correction factor is the ratio of the measured length of the corresponding limb of the mother to the length of the initial template limb. The length dimension in the skeletal geometry parameters of each skeleton group is multiplied by the scaling factor of the corresponding skeleton group, the inertia tensor in the skeletal geometry parameters is multiplied by the fifth power of the scaling factor, and the centroid position in the skeletal geometry parameters is multiplied by the scaling factor to complete the scaling deformation of the initial template. The deformed bones and their connections constitute a personalized skeletal chain.

[0025] In practice, the range of motion constraints for each joint are set as boundary conditions based on the measured joint range of motion values ​​of the parturient. A joint range of motion goniometer is used to measure the parturient's joint range of motion, including shoulder flexion / extension, shoulder abduction / adduction, elbow flexion / extension, wrist flexion / extension, hip flexion / extension, hip abduction / adduction, knee flexion / extension, and ankle dorsiflexion / plantarflexion. Each joint is measured three times in each direction of motion. The average of the three measurements is taken as the maximum permissible range of motion for that joint in that direction, and zero is taken as the minimum permissible range of motion. The maximum and minimum permissible ranges of motion for each joint in each direction are used as boundary conditions and written into the definition of the degree of freedom constraints for the corresponding joints in the multi-rigid-body dynamics model.

[0026] In practice, the origin and insertion coordinates of major muscle groups are mapped onto a personalized skeletal chain, establishing a function table for the muscle lever arm as a function of joint angle. Major muscle groups include the rectus abdominis, external oblique, erector spinae, gluteus maximus, iliopsoas, quadriceps femoris, hamstrings, tibialis anterior, gastrocnemius, deltoid, biceps brachii, and triceps brachii. For each muscle, the coordinates of its origin and insertion in the corresponding local skeletal coordinate system are extracted from a pre-defined human anatomy database. These coordinates are then scaled proportionally according to the scaling factor of the corresponding bone in the personalized skeletal chain; the scaled coordinates are the muscle origin and insertion coordinates mapped onto the personalized skeletal chain. For each muscle spanning a single joint, the joint angle is traversed from the minimum to the maximum permissible range of motion, step by step, according to a pre-defined angle. At each traversed angle, the muscle lever arm is calculated based on the muscle origin and insertion coordinates and the joint rotation center. The muscle lever arm is defined as the perpendicular distance from the joint rotation center to the muscle's line of action. The muscle lever arm values ​​at each traversed angle are recorded, forming a discrete data table with joint angles as input and muscle lever arm as output. For muscles spanning multiple joints, a multi-dimensional data table is established to represent the change in muscle lever arm with each joint angle. The discrete data tables are stored according to the direction of joint movement, forming a function table of muscle lever arm variation with joint angle. During forward dynamics simulation, the muscle lever arm values ​​at any joint angle are obtained from the function table using interpolation methods.

[0027] See Figure 3The graph shows the horizontal axis representing the hip flexion-extension angle, ranging from 0 to 120 degrees, and the vertical axis representing the muscle lever arm, in meters. It illustrates the lever arm trends of four major muscles—the gluteus maximus, quadriceps femoris, iliopsoas, and hamstrings—at different hip flexion-extension angles. The gluteus maximus lever arm (solid black line) initially rises rapidly with increasing hip flexion-extension angle, reaching a maximum of 0.09 meters at approximately 90 degrees, before slightly decreasing. The quadriceps femoris lever arm (dashed blue line) shows an overall upward trend, starting at approximately 0.07 meters at 0 degrees, gradually increasing to nearly 0.09 meters, and then slightly decreasing at 100 degrees. The iliopsoas lever arm (dashed red line) shows the opposite trend, peaking at approximately 0.086 meters at 0 degrees, and continuously decreasing with increasing flexion-extension angle, reaching near 0 at 120 degrees. The hamstring lever arm (green dotted line) first decreases to a minimum of approximately 0.029 meters within the 0-30 degree range, then gradually increases, reaching approximately 0.07 meters at 120 degrees. Overall, the lever arms of the four muscles exhibit a clear non-linear variation characteristic with changes in the hip flexion and extension angles, reflecting the dynamic change law of muscle lever arms with joint movement, and providing specific data support for the construction of muscle lever arm function tables in multi-rigid-body dynamics models.

[0028] Example 3: In practice, the process of collecting spatial motion trajectory data of the mother's torso and limbs through an inertial measurement unit array is as follows: An inertial measurement unit is fixed in each of the ten body segments of the mother, namely the pelvis, thoracic vertebrae, left upper arm, right upper arm, left forearm, right forearm, left thigh, right thigh, left calf, and right calf. The inertial measurement unit at the pelvis was fixed to the body surface at the midpoint of the line connecting the posterior superior iliac spines of the mother. The inertial measurement unit at the thoracic spine was fixed to the body surface at the midpoint of the line connecting the spinous processes of the seventh cervical vertebra and the first thoracic vertebra. The inertial measurement units for the left and right upper arms were fixed to the body surface at the deltoid tuberosity on the lateral side of the mother's left and right upper arms, respectively. The inertial measurement units for the left and right forearms were fixed to the body surface 3 cm proximal to the radial styloid process on the dorsal side of the mother's left and right forearms, respectively. The inertial measurement units for the left and right thighs were fixed to the body surface 10 cm proximal to the upper edge of the patella on the anterior side of the mother's left and right thighs, respectively. The inertial measurement units for the left and right lower legs were fixed to the body surface 5 cm distal to the tibial tuberosity on the anterior side of the mother's left and right lower legs, respectively. Each inertial measurement unit was tightly fitted to the mother's body surface with an elastic band, and a 5 mm thick silicone pad was placed between the elastic band and the body surface to suppress measurement noise introduced by soft tissue vibration.

[0029] Each inertial measurement unit (IMU) contains a three-axis accelerometer and a three-axis gyroscope. The accelerometer's range is ±16 times the standard gravitational acceleration, and the gyroscope's range is ±2000 degrees per second. Each IMU outputs raw acceleration and angular velocity data at a sampling rate of 100 Hz. Each sample outputs six floating-point values: the first three are the acceleration components along the three orthogonal axes of the IMU's local coordinate system, and the last three are the angular velocity components around the three orthogonal axes. The ten IMUs are synchronized via a wireless synchronization protocol to eliminate data packet transmission delay. The wireless synchronization protocol uses a master-slave mode, with the IMU fixed at the thoracic spine designated as the master node and the other nine as slave nodes. At the beginning of each sampling period, the master node broadcasts a synchronization beacon frame via radio frequency. The synchronization beacon frame contains the master node's current sampling sequence number and its local clock count. Upon receiving a synchronization beacon frame, each slave node records its local clock count at the moment of reception. Based on the difference between the master node's local clock count and the slave node's local clock count in the synchronization beacon frame, it calculates the slave node's clock offset relative to the master node. It then uses linear interpolation to re-mark the timestamps of the raw acceleration and angular velocity data that have been buffered but not yet processed by the slave node, aligning them to the master node's time base. If any slave node does not receive a synchronization beacon frame within a predetermined waiting time, it uses the clock drift rate calculated during the most recent successful synchronization to timestamp subsequent sampled data until a new synchronization beacon frame is received.

[0030] An error-state Kalman filter (ERF) algorithm based on rigid body attitude is employed to fuse acceleration and angular velocity data collected by each inertial measurement unit (IMU), outputting the absolute attitude and position coordinates of the fixed body segments of each IMU in the world coordinate system. The state vector of the ERF algorithm comprises nine components: three components representing the position coordinates of the body segments in the world coordinate system, three components representing the velocities of the body segments relative to the world coordinate system, and three components representing the attitude quaternion error of the body segments relative to the world coordinate system. The observation model of the ERF algorithm uses acceleration data from the three-axis accelerometer as linear acceleration observations and angular velocity data from the three-axis gyroscope as angular velocity observations. The prediction step of the ERF algorithm is executed at each sampling time. Using the posterior state estimate from the previous sampling time and the angular velocity observation from the current sampling time, the prior attitude quaternion estimate for the current sampling time is calculated based on the quaternion kinematic equations. Simultaneously, the velocity is predicted using the gravitational acceleration vector in the world coordinate system. After prediction, the update step begins. The difference between the observed acceleration at the current sampling moment and the gravitational acceleration vector is used as a correction for the position error. The Kalman gain matrix is ​​calculated and used to correct the prior estimate of the state vector, resulting in the posterior state estimate. The output stage uses the three components of the position coordinates from the posterior state estimate as the position coordinates of the body segments in the world coordinate system, and the attitude quaternion corresponding to the posterior state estimate as the absolute attitude. This processing is performed independently for each inertial measurement unit (IMU), ultimately outputting the position coordinates and absolute attitude of each of the ten body segments in the world coordinate system.

[0031] In practice, the process of collecting electromyographic activity data of target muscle groups using a surface electromyographic electrode array is as follows: Surface electromyographic electrodes are attached to the surface of the rectus abdominis, external oblique, erector spinae, gluteus maximus, and pelvic floor muscles of the mother. Two electrodes are attached to the rectus abdominis: one at a horizontal line 3 cm above the umbilicus and two cm to the left and right of the linea alba. One electrode is attached to the external oblique: one at the midpoint of the line connecting the anterior superior iliac spine and the lower edge of the ribs. One electrode is attached to the erector spinae: one at a point 3 cm lateral to the spinous process of the second lumbar vertebra. One electrode is attached to the gluteus maximus: one at the midpoint of the line connecting the posterior superior iliac spine and the greater trochanter of the femur. The pelvic floor muscles use vaginal probe electrodes with four electrode contacts arranged in a ring on the probe surface, spaced 90 degrees apart. Before attaching all surface electrodes, the skin surface is cleaned with alcohol swabs to remove dead skin cells and oil.

[0032] Each surface electromyography (EMG) electrode employs a bipolar differential measurement mode, with a distance of 20 mm between the two differential electrodes. The sampling rate for each EMG electrode is 1 kHz, and the common-mode rejection ratio (CMRR) of the front-end amplifier for each electrode is greater than 100 dB. The acquired raw EMG signals undergo bandpass filtering. The passband frequency of the bandpass filter is set between 20 Hz and 450 Hz. The 20 Hz passband frequency is used to suppress motion artifacts and low-frequency drift, while the 450 Hz passband frequency is used to avoid high-frequency electromagnetic interference and signal aliasing. A fourth-order Butterworth filter is used in the bandpass filter process, which is a zero-phase filtering process. This involves first performing forward filtering on the signal, then inverting the filtered signal and performing a second filtering to eliminate phase distortion.

[0033] The moving mean square (RMS) value of the bandpass-filtered electromyography (EMG) signal was calculated to obtain the amplitude sequence of the EMG activity data. The window length for the moving mean square calculation was 200 milliseconds, corresponding to 200 sampling points at a sampling rate of 1 kHz. The step size for the moving mean square calculation was 10 sampling points, corresponding to a time step of 10 milliseconds. At each position of the moving window, the RMS value was calculated as shown in the following formula:

[0034] in, This represents the amplitude of electromyographic activity data at the position of the m-th sliding window, in microvolts. This represents the length of the sliding window, with a value of 200. represents the sliding window step length, with a value of 10; p represents the index of the sampling point of the filtered electromyography signal. This represents the amplitude of the electromyographic signal at the p-th sampling point after bandpass filtering, in microvolts. The moving average root mean square (RMS) calculation is performed independently on the signals acquired by each surface electromyographic electrode, ultimately outputting an amplitude sequence of electromyographic activity data for each electrode channel.

[0035] Example 4: In practice, spatial motion trajectory data and electromyographic activity data are used as driving inputs to perform forward dynamic simulation on the multi-rigid-body dynamics model, thereby obtaining the reaction torque of each joint and the muscle activation sequence.

[0036] In practice, the acceleration and angular velocity data collected by the inertial measurement unit array (IMU) are fused using a complementary filtering algorithm to obtain the attitude quaternion sequence for each body segment. The complementary filtering algorithm is executed independently for the data collected by each IMU. Taking a body segment as an example, the original acceleration and angular velocity data vectors at the current sampling time are obtained from the IMU. The original acceleration data vector is normalized to obtain a normalized acceleration vector. The normalized acceleration vector is cross-producted with the pre-stored gravity acceleration reference vector in the local coordinate system of the body segment described by the attitude quaternion output at the previous sampling time. The result of the cross-product is an error vector, which describes the deviation between the gravity direction represented by the normalized acceleration vector and the gravity direction calculated from the attitude quaternion. The error vector is multiplied by a proportional gain coefficient and then added to the original angular velocity data vector at the current sampling time to obtain the corrected angular velocity vector. The proportional gain coefficient is set to 0.1. This value is determined by adjusting the proportional gain coefficient to ensure the transition band slope of the filter's amplitude-frequency response curve meets dynamic response requirements, with the complementary filter's cutoff frequency set to 0.05 Hz. Experimental calibration confirmed 0.1 as the optimal proportional gain coefficient. The corrected angular velocity vector is integrated with the attitude quaternion output from the previous sampling time using a quaternion differential equation to obtain the attitude quaternion at the current sampling time. The integration time step is the reciprocal of the sampling period, i.e., 0.01 seconds. This process is repeated for all inertial measurement units at each sampling time, resulting in attitude quaternion sequences for each of the ten body segments. Each attitude quaternion sequence is a set of attitude quaternions arranged chronologically.

[0037] In practice, the joint angle positions and angular velocities between adjacent segments are calculated based on the posture quaternion sequence. The definition of adjacent segments is based on the connections within the human skeletal chain: the pelvic segment and the thoracic vertebrae form the lumbar joint; the pelvic segment and the left thigh segment form the left hip joint; the pelvic segment and the right thigh segment form the right hip joint; the left thigh segment and the left calf segment form the left knee joint; the right thigh segment and the right calf segment form the right knee joint; the thoracic vertebrae segment and the left upper arm segment form the left shoulder joint; the thoracic vertebrae segment and the right upper arm segment form the right shoulder joint; the left upper arm segment and the left forearm segment form the left elbow joint; and the right upper arm segment and the right forearm segment form the right elbow joint. Taking the lumbar spine joint as an example, for the same sampling time, the posture quaternions of the pelvic segment and the thoracic segment are known. The conjugate quaternion of the pelvic segment posture quaternion is calculated, and then a quaternion multiplication operation is performed between the conjugate quaternion of the pelvic segment posture quaternion and the posture quaternion of the thoracic segment to obtain the quaternion representing the relative rotation of the lumbar spine joint. The relative rotation quaternion of the lumbar spine joint is converted into Euler angles, where the three angular components correspond to the joint angle positions in the sagittal, coronal, and horizontal planes, respectively. The angular velocity is calculated by numerically differentiating the joint angle positions of two adjacent sampling times. The numerical differentiation uses the central difference method. Except for the first and last sampling times, the joint angular velocity at the nth sampling time is equal to the difference between the joint angle position at the (n+1)th sampling time and the joint angle position at the (n-minus)th sampling time, divided by twice the sampling period, which is 0.01 seconds. Perform the above calculations on all adjacent segment combinations at each sampling time to obtain the joint angular positions and angular velocities of all joints.

[0038] In practice, the electromyography (EMG) signals acquired by the surface EMG electrode array are rectified and low-pass filtered to obtain the muscle activation envelope. The band-pass filtered EMG signals acquired by each surface EMG electrode channel are processed independently. Rectification involves taking the absolute value of each sampling point of the band-pass filtered EMG signal to obtain a full-wave rectified EMG signal. The full-wave rectified EMG signal is then low-pass filtered using a second-order Butterworth filter with a cutoff frequency of 3 Hz. The cutoff frequency of 3 Hz is chosen because the upper limit of the mechanical response frequency of muscle contraction is approximately 2.5 Hz. Setting the low-pass filter cutoff frequency to 3 Hz effectively filters out noise components higher than the contraction frequency while completely preserving the envelope characteristics of muscle activation information. The signal after low-pass filtering is the muscle activation envelope, which is a sequence of real numbers with the same number of sampling points as the original EMG signal. The value at each sampling point represents the instantaneous value of the muscle activation level at the corresponding moment.

[0039] In practice, joint angular position, angular velocity, and muscle activation envelope are used as inputs to the multi-rigid-body dynamics model. The recursive Newton-Euler algorithm is used to calculate the reaction torques at each joint. The recursive Newton-Euler algorithm consists of two stages: outward recursion and inward recursion. In the outward recursion stage, starting from the base of the skeletal chain (the pelvic segment), the linear and angular accelerations of each segment are calculated progressively towards the distal segments. The known input is the linear acceleration of the base segment in the world coordinate system, which is the opposite direction to gravitational acceleration and has a magnitude of 9.81 m / s². Based on the joint angular position and angular velocity, the rotation matrix and rotational angular velocity of the current segment relative to the parent segment are calculated. Combining this with the linear and angular accelerations of the parent segment, the linear and angular accelerations of the current segment are recursively obtained. In the inward recursion stage, starting from the distal segments, the reaction forces and reaction torques at each joint are calculated progressively towards the base. At each segment, the segmental inertial force is calculated based on the linear acceleration and mass of the current segment, and the segmental inertial torque is calculated based on the angular acceleration and segmental inertial tensor. The segmental mass and segmental inertial tensor are provided by the geometric parameters of the corresponding bones in the personalized skeletal chain. The muscle activation envelope is used to correct the instantaneous lever arm of the corresponding muscle in the muscle lever arm function table. During the inward recursion phase of the recursive Newton-Euler algorithm to solve for the joint reaction torque, a muscle contribution torque term is added. The muscle contribution torque term is equal to the current activation value of the corresponding muscle in the muscle activation envelope multiplied by the lever arm value of that muscle at the current joint angle in the muscle lever arm function table multiplied by the maximum isometric contractile force of that muscle. The maximum isometric contractile force is obtained by querying the preset human anatomy database by muscle name index. The linear equations of the joint reaction torque are solved, and the joint reaction torque components of each joint in the three motion directions of the sagittal, coronal, and horizontal planes are output.

[0040] In practice, the spatial motion trajectory data is simultaneously converted into angular displacement, angular velocity, and angular acceleration sequences for each joint. The spatial motion trajectory data consists of the position coordinates and absolute attitude of ten body segments in the world coordinate system. The conversion process follows the joint kinematics calculation method based on attitude quaternions and numerical differentiation. The angular position sequence of each joint is used as the angular displacement sequence, the first-order numerical derivative of the joint angular position is used as the angular velocity sequence, and the first-order numerical derivative of the joint angular velocity is used as the angular acceleration sequence.

[0041] In practice, electromyography (EMG) activity data is converted into a normalized activation sequence for each muscle. EMG activity data is the sequence of EMG activity amplitude values ​​output by each surface electromyography (SEM) electrode channel. For each monitored muscle, there is at least one corresponding SEM channel. When a muscle corresponds to multiple SEM channels, the maximum value among the EMG activity amplitude values ​​of multiple SEM channels at the same sampling time is taken as the original activation amplitude value of that muscle at that sampling time. For each muscle, a maximum voluntary contraction calibration is performed before formal training. The mother applies maximum force to the muscle according to instructions, and the EMG activity amplitude sequence of the corresponding SEM channel at the time of maximum voluntary contraction is recorded. The maximum value in this sequence is taken as the maximum EMG amplitude reference value for that muscle. The normalized activation sequence is calculated by dividing the original activation amplitude value of the muscle at each sampling time by the maximum EMG amplitude reference value for that muscle, obtaining a normalized activation value ranging from zero to one. These values ​​are then arranged in chronological order to form a normalized activation sequence.

[0042] In practice, based on the muscle lever arm function table, the normalized activation degree is multiplied by the maximum isometric contraction force of the corresponding muscle to obtain the muscle force vector. The muscle lever arm function table is established when constructing the multi-rigid-body dynamics model, storing the corresponding relationship between the lever arm of each muscle and the joint angle. At each sampling moment during the simulation, the current angular displacement values ​​of all joints are obtained, and the lever arm value of each muscle at the current joint angle is found in the muscle lever arm function table through linear interpolation. The maximum isometric contraction force is retrieved from a pre-set human anatomy database. The maximum isometric contraction force of each muscle is a value based on standard human body data and linearly scaled according to the mother's weight, with the scaling ratio being the ratio of the mother's weight to the standard human body weight. The magnitude of the muscle force vector is the normalized activation degree multiplied by the maximum isometric contraction force, and the direction of the muscle force vector is along the line connecting the muscle's origin and insertion points.

[0043] In practice, the angular acceleration sequence and muscle force vectors are substituted into the Newton-Euler equations of the multi-rigid-body dynamics model, and the reaction torques of each joint are obtained by solving the equations. The Newton-Euler equations of the multi-rigid-body dynamics model are expressed in the following form:

[0044] in, This represents the vector formed by the angular displacements of each joint; This represents the vector formed by the angular velocities of each joint. This represents the vector formed by the angular accelerations of each joint. The mass matrix is ​​represented by the mass of each segment and the inertia tensor of each segment as a function of the joint angle. Represents the matrix of Coriolis force and centripetal force; Represents the gravity term vector; This represents the reaction torque vector of each joint, which is the unknown quantity to be solved. This represents the muscle lever arm matrix, where the element in the a-th row and b-th column is the lever arm value of the b-th muscle with respect to the a-th joint. Let represent the muscle force vector, where the b-th element is the magnitude of the muscle force vector of the b-th muscle at the current sampling time. Substituting the angular acceleration sequence and the muscle force vector into the Newton-Euler equations, a linear equation system solution method based on LU decomposition is used to solve the problem. This yields the reaction torques of each joint.

[0045] In practice, the normalized activation sequence is arranged chronologically to form the muscle activation time sequence. The normalized activation sequence of each muscle is the muscle activation time sequence of that muscle. The muscle activation time sequences of all muscles constitute a multidimensional time series. The first dimension of the multidimensional time series is the muscle number, and the second dimension is the time sampling point number.

[0046] Example 5: In practice, the process of overlaying a 3D image of a virtual coach's demonstration movements onto the mother's field of vision using an augmented reality head-mounted display is as follows: Based on the motion identifier of the current training segment in the initial training sequence, the corresponding standard motion capture data is retrieved from the motion library. Each training segment in the initial training sequence contains a segment identifier, which serves as the motion identifier. The motion library is a pre-built structured database. Each record in the motion library uses a motion identifier as its primary key and stores the corresponding standard motion capture data. The standard motion capture data is collected by an optical motion capture system on a standard demonstrator. The standard demonstrator has a preset standard human skeleton proportion, and the standard motion capture data contains a time series of the 3D position coordinates of 52 skeletal nodes throughout the standard demonstrator's body, with a capture frame rate of 120 frames per second. The retrieval process uses the motion identifier as the query key, performing hash matching in the motion library's index table. If a match is successful, the corresponding standard motion capture data file is read and loaded into memory. The standard motion capture data is then mapped onto the virtual coach's skeletal drive template to generate the virtual coach's 3D animation. The virtual coach's skeletal drive template contains virtual bone nodes that correspond one-to-one with the fifty-two bone nodes in the standard motion capture data. Each virtual bone node has an initial position in the template's local coordinate system. The mapping process converts the 3D position coordinates of each bone node in each frame of the standard motion capture data into the target position of the corresponding virtual bone node in the virtual coach's skeletal drive template. The conversion method involves multiplying the coordinates of the bone nodes in the standard motion capture data by a scaling matrix, where the diagonal elements of the scaling matrix are the ratio of the virtual coach's bone length to the standard presenter's bone length. The mapped bone node target position sequence is used as a driving signal and input into the inverse kinematics solver of the virtual coach's skeletal drive template. The inverse kinematics solver outputs the joint angle sequence of each joint of the virtual coach. The joint angle sequence drives the deformation of the virtual coach's 3D skinned mesh model, generating a 3D animation of the virtual coach frame by frame. The spatial positioning module of the augmented reality head-mounted display is used to obtain the position and orientation of the mother's head, and the 3D animation of the virtual coach is superimposed onto a predetermined position in the mother's field of vision. The spatial positioning module of the augmented reality head-mounted display (ARHDD) includes two front-facing wide-angle grayscale cameras and an inertial measurement unit (IMU). The two cameras acquire images of the environment in front of the mother at a frame rate of 60 frames per second. A sparse point cloud map of the environment is constructed in real time using a simultaneous localization and mapping (SMR) algorithm, and the six-DOF pose of the ARHDD is calculated within this map. The mother's head position is determined by the translation component of the ARHDD's six-DOF pose, and her head orientation is determined by the rotation component. The predetermined position is set as a fixed world coordinate system point 1.5 meters directly in front of the mother's field of vision, at a height of 0.8 meters above the ground.The root skeleton node of the virtual coach's 3D animation is placed in a predetermined position. The observation and projection matrices are calculated based on the mother's head position and orientation. Perspective projection rendering of the virtual coach's 3D animation is then performed and overlaid on the transparent optical display lens of the augmented reality head-mounted display. The transparency of the virtual coach is dynamically adjusted based on the reaction torque and muscle activation sequence. When rendering the virtual coach's 3D animation on the augmented reality head-mounted display, the virtual coach's 3D animation material uses a texture format with an alpha channel. The alpha channel value controls the transparency of the virtual coach. The initial value of the alpha channel value is set to 1.0, corresponding to a completely transparent effect. In each rendering frame, the reaction torque of all joints at the current moment is obtained, and the root mean square value of the reaction torque of all joints is calculated as the comprehensive reaction torque value. The comprehensive reaction torque value is compared with a preset threshold of 15 Nm. The preset threshold of 15 Nm is based on the upper limit of comfort torque determined in a repetitive load tolerance study of adult female wrist and elbow joints. When the comprehensive reaction torque value is greater than the preset threshold, the alpha channel value is updated according to the following formula:

[0047] in, This indicates the updated Alpha channel value; This represents the alpha channel value of the current rendered frame; This represents the transparency attenuation coefficient, with a value of 0.1 per Newton-meter. The basis for this value is that, through experimental calibration, when the combined value of the reaction torque exceeds a preset threshold of 5 Newton-meters, the transparency of the virtual coach produces a perceptible change within the semi-transparent range. This represents the combined value of the reaction torque, expressed in Newton-meters (Nm). This represents a preset threshold, expressed in Newton-meters (Nm). When the combined reaction torque is less than or equal to the preset threshold, the alpha channel value gradually recovers to its initial value of 1.0, with a recovery rate of 0.05 per frame. The updated alpha channel value is assigned to the alpha channel of the virtual coach's 3D animation material, and then blended and rendered by the rendering pipeline of the augmented reality head-mounted display, thereby reducing the virtual coach's transparency when the reaction torque exceeds the preset threshold. In specific implementation, the process of applying auxiliary force or resistance to the mother's hand through the force feedback glove based on the reaction torque and muscle activation sequence is as follows: The component related to hand movement in the reaction torque is extracted as the desired hand torque. The reaction torque vectors of each joint are obtained from the forward dynamic simulation of the multi-rigid-body dynamics model. In the process, torque components corresponding to joints related to hand movements are selected. These joints include the flexion-extension components of the left wrist joint, the abduction-adduction components of the left wrist joint, the flexion-extension components of the right wrist joint, the abduction-adduction components of the right wrist joint, the flexion-extension components of the left thumb metacarpophalangeal joint, the flexion-extension components of the left index finger metacarpophalangeal joint, the flexion-extension components of the left middle finger metacarpophalangeal joint, the flexion-extension components of the left ring finger metacarpophalangeal joint, the flexion-extension components of the left little finger metacarpophalangeal joint, and the corresponding components of the right hand. These twenty torque components are combined into a twenty-dimensional vector as the desired hand torque. The activation degree of hand muscle groups in the muscle activation sequence is coupled with the desired hand torque to generate a force feedback instruction set. The muscle activation sequence contains the normalized activation degree sequence of all monitored muscles. The hand muscle groups include the left flexor digitorum superficialis, the left extensor digitorum commonis, the right flexor digitorum superficialis, and the right extensor digitorum commonis. The activation degree of the hand muscle groups is extracted from the muscle activation sequence by indexing the hand muscle group name at each sampling time. The coupling method is as follows: for each element in the desired torque vector of the hand, the element value is multiplied by the normalized activation degree of the corresponding finger flexor or extensor muscle to obtain the target force value of the corresponding micro linear motor in the force feedback command set. The force feedback command set consists of twenty target force values, each corresponding to one of the twenty micro linear motors in the stress feedback glove. Each finger sleeve of the force feedback glove embeds a micro linear motor and a tension sensor. The force feedback glove contains ten finger sleeves, which are worn on the mother's left thumb, left index finger, left middle finger, left ring finger, left little finger, and the corresponding five fingers on the right hand. Each finger sleeve has a micro linear motor embedded on both the palmar and dorsal sides, and each finger sleeve also has a tension sensor embedded in it. The micro linear motors are connected to fixed anchor points at the fingertips of the finger sleeves via steel wire ropes. The pushing and pulling motion of the micro linear motors is converted into pulling or pushing forces applied to the fingers through the steel wire ropes, thereby realizing the output of auxiliary force or resistance. A tension sensor is connected in series between the steel wire rope and the fingertip anchor point of the finger sleeve. The tension sensor measures the actual contact force between the steel wire rope and the finger in real time. The output force of the micro linear motor is controlled according to the force feedback command set. Each target force value in the force feedback command set is wirelessly transmitted to the control board of the force feedback glove via Bluetooth communication protocol. The microcontroller on the control board generates a pulse-width modulation drive signal for the micro linear motor corresponding to the target force value. By measuring the actual contact force between the finger sleeve and the finger in real time through the tension sensor, a proportional-integral (PI) controller is used to adjust the actual contact force to the target force in the force feedback command set. The control law of the PI controller is expressed as:

[0048] in, This represents the output voltage of the miniature linear motor at time t, in volts. This represents the force error at time t. The force error is equal to the target force value in the force feedback command set minus the actual contact force value measured by the tension sensor, and the unit is Newtons. This represents the proportional gain coefficient, with a value of 0.8 volts per newton. This represents the integral gain coefficient, with a value of 0.1 volts per second. Proportional gain coefficient. and integral gain coefficient The values ​​were determined based on the following: Step response tests were conducted on the fingertips of the force feedback glove. The proportional gain and integral gain coefficients were adjusted to ensure the system overshoot was less than 5% and the settling time was less than 0.2 seconds. The optimal combination was then calibrated. , The microcontroller reads the force sensor value in each control cycle, calculates the force error, executes the proportional-integral control law, and updates the drive voltage of the micro linear motor, so that the actual contact force tracks the target force value.

[0049] See Figure 4 In the graph, the horizontal axis represents time, ranging from 0 to 10 seconds; the left vertical axis represents the comprehensive reaction torque value, in Newton-meters (Nm), ranging from approximately 0 to 26 Nm; the right vertical axis represents the alpha channel value, ranging from approximately 0.4 to 1.0; the dashed line represents the preset threshold, fixed at 15 Nm. The solid black line in the graph is the curve of the comprehensive reaction torque value, which exhibits a periodic fluctuation trend. The comprehensive reaction torque value peaks at multiple time points, reaching a maximum of approximately 25 Nm and a minimum close to zero. The dashed red line represents the alpha channel value, which shows an overall downward trend. Initially, the alpha channel value is approximately 1.0, then decreases in a stepwise manner over time, with several obvious abrupt drops, finally reaching approximately 0.45 at the end of the 10-second period.

[0050] The combined reaction torque value exceeded the preset threshold of 15 Nm multiple times throughout the time period, with the peak values ​​significantly higher than the threshold at 1 second, 2 seconds, 4 seconds, 6 seconds, 8 seconds, and 9 seconds, triggering the alpha channel value's decreasing mechanism. The decrease in the alpha channel value corresponds to the time points when the combined reaction torque value exceeds the threshold, and its rate of decrease is related to the magnitude of the reaction torque exceeding the threshold. When the reaction torque is below or close to the threshold, the alpha channel value tends to stabilize or slightly recover, but overall exhibits a gradual trend towards transparency.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A clinical teaching and training method for postpartum rehabilitation based on artificial intelligence, characterized in that, include: Obtain a set of baseline parameters for postpartum recovery of the mother, including resting tension of the pelvic floor muscles, inter-abdominal distance, and spinal curvature index; An initial training sequence is extracted from the rehabilitation training prescription library based on the baseline parameter set. The initial training sequence contains the timing arrangement and intensity coefficient of multiple training sessions. During the training process, spatial motion trajectory data of the mother's trunk and limbs were collected through an inertial measurement unit array, while electromyographic activity data of the target muscle groups were collected through a surface electromyography electrode array. A multi-rigid-body dynamics model of the musculoskeletal system of a parturient woman is constructed, wherein the multi-rigid-body dynamics model includes the degree-of-freedom constraints of each joint and the muscle lever arm function. Using the spatial motion trajectory data and the electromyographic activity data as driving inputs, a forward dynamic simulation is performed on the multi-rigid-body dynamics model to obtain the reaction torque of each joint and the muscle activation sequence. An augmented reality head-mounted display overlays a 3D image of a virtual coach demonstrating movements onto the mother's field of vision. The 3D image of the virtual coach demonstrating movements is generated based on the initial training sequence. The force feedback glove applies auxiliary force or resistance to the mother's hand based on the reaction torque and muscle activation sequence.

2. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The acquisition of the postpartum recovery baseline parameter set for mothers includes: The shear wave velocity of the pelvic floor muscles in postpartum women is measured using an ultrasonic shear wave elastography device, and the shear wave velocity is converted into a resting tension value of the pelvic floor muscles. The distance between the medial borders of the rectus abdominis muscles in postpartum women was measured by ultrasound B-mode imaging, and the distance between the medial borders of the rectus abdominis muscles was taken as the rectus abdominis distance. Point cloud data of the mother's torso surface was acquired using a 3D optical scanner, and the curvature of the spine in the sagittal and coronal planes was calculated to obtain the spinal curvature index.

3. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The multi-rigid-body dynamics model of the musculoskeletal system of the mother includes: Standard bone dimensions and joint positions are extracted from a pre-defined human anatomy database as initial templates; The initial template is scaled and deformed based on the mother's height, weight, and body fat percentage to generate a personalized skeletal chain; Set the range of motion constraints of each joint to boundary conditions based on the measured values ​​of the joint mobility of the mother. The origin and insertion coordinates of the major muscle groups are mapped onto the personalized skeletal chain, and a function table is established to show how the muscle lever arm changes with the joint angle.

4. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The step of using the spatial motion trajectory data and electromyographic activity data as driving inputs to perform forward dynamic simulation on the multi-rigid-body dynamics model includes: The acceleration and angular velocity data collected by the inertial measurement unit array are fused using a complementary filtering algorithm to obtain the attitude quaternion sequence of each body segment; Calculate the joint angular position and angular velocity between adjacent segments based on the attitude quaternion sequence; The electromyographic signals acquired by the surface electromyographic electrode array are rectified and low-pass filtered to obtain the muscle activation envelope. The joint angular position, angular velocity, and muscle activation envelope are used as inputs to a multi-rigid-body dynamics model, and the reaction torque of each joint is calculated using the recursive Newton-Euler algorithm.

5. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The method of overlaying a 3D image of a virtual coach demonstrating movements onto the mother's field of vision using an augmented reality head-mounted display includes: Based on the action identifier of the current training stage in the initial training sequence, retrieve the corresponding standard motion capture data from the action library; The standard motion capture data is mapped onto the skeleton-driven template of the virtual coach to generate a 3D animation of the virtual coach. The spatial positioning module of the augmented reality head-mounted display is used to obtain the position and orientation of the mother's head, and the 3D animation of the virtual coach is superimposed onto the predetermined position in the mother's field of vision. The transparency of the virtual coach is dynamically adjusted based on the reaction torque and muscle activation sequence, and the transparency of the virtual coach is reduced when the reaction torque exceeds a preset threshold.

6. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The force feedback glove applies auxiliary force or resistance to the mother's hand based on the reaction torque and muscle activation sequence, including: The component of the reaction torque related to hand movement is extracted and used as the desired hand torque; The activation degree of the hand muscle group in the muscle activation sequence is coupled with the desired torque of the hand to generate a force feedback instruction set; Each finger sleeve of the force feedback glove is embedded with a miniature linear motor and a tension sensor, and the output force of the miniature linear motor is controlled according to the force feedback instruction set. The actual contact force between the finger sleeve and the finger is measured in real time by a tension sensor, and the actual contact force is adjusted to the target force in the force feedback command set by a proportional-integral controller.

7. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The step of extracting initial training sequences from the rehabilitation training prescription library based on the baseline parameter set includes: The resting tension of the pelvic floor muscles, the inter-rectus distance, and the spinal curvature index are discretized and encoded to generate a three-dimensional combined code; Using the three-dimensional combination code as an index, the corresponding training prescription record is queried in the index table of the rehabilitation training prescription database; The time series table of the training phases and the intensity coefficient of each training phase are parsed from the training prescription record to generate the initial training sequence.

8. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The acquisition of spatial motion trajectory data of the mother's torso and limbs via an inertial measurement unit array includes: Inertial measurement units were fixed to the pelvis, thoracic vertebrae, left and right upper arms, left and right forearms, left and right thighs, and left and right calves of the mother. Each inertial measurement unit contains a three-axis accelerometer and a three-axis gyroscope, outputting raw acceleration and angular velocity data at a sampling rate of 100 Hz; The timestamps of all inertial measurement units are aligned using a wireless synchronization protocol to eliminate data packet transmission delay; An error state Kalman filter algorithm based on rigid body attitude is used to fuse acceleration and angular velocity data, and output the absolute attitude and position coordinates in the world coordinate system of each body segment where the sensor is located.

9. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The acquisition of electromyographic activity data of the target muscle group via a surface electromyographic electrode array includes: Surface electromyography electrodes were attached to the surface of the rectus abdominis, external oblique, erector spinae, gluteus maximus, and pelvic floor muscles of the mother. Each surface electromyography electrode employs a bipolar differential measurement mode with a sampling rate of 1000 Hz and a common-mode rejection ratio greater than 100 dB; The raw electromyography signals were bandpass filtered, with the passband frequency set to 20 to 450 Hz. The amplitude sequence of electromyographic activity data was obtained by performing sliding root mean square calculation on the bandpass filtered electromyographic signal with a window length of 200 milliseconds.

10. The postpartum rehabilitation clinical teaching and training method based on artificial intelligence according to claim 1, characterized in that, The forward dynamic simulation of the multi-rigid-body dynamics model, to obtain the reaction torques of each joint and the muscle activation sequence, includes: The spatial motion trajectory data is converted into sequences of angular displacement, angular velocity, and angular acceleration for each joint; The electromyographic activity data was converted into normalized activation sequences for each muscle. Based on the muscle lever arm function table, the normalized activation degree is multiplied by the maximum isometric contraction force of the corresponding muscle to obtain the muscle force vector; Substitute the angular acceleration sequence and the muscle force vector into the Newton-Euler equations of the multi-rigid-body dynamics model, and obtain the reaction torque of each joint by solving the equation system. The normalized activation sequence is arranged in chronological order as the muscle activation sequence.