Lifting capacity prediction method and system based on muscle force-joint load dynamic parameters
By fusing multi-source heterogeneous data and biomechanical simulation technology, a personalized skeletal muscle model is constructed. The XGBoost-LSTM hybrid model is used to predict weightlifting ability, which solves the safety risks and low prediction accuracy of traditional 1RM testing and achieves high-precision 1RM prediction and compensatory action recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196347A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of sports biomechanics and intelligent training technology, specifically involving a method and system for predicting the ability of weightlifters based on the fusion of multi-source heterogeneous data. Background Technology
[0002] The current assessment of 1RM (One Repetition Maximum) in weightlifting faces three major technical bottlenecks, which severely restrict the implementation of scientific training: High safety risks. Traditional 1RM testing requires athletes to complete a single lift close to their physiological limits under unprotected or limited protection, which greatly increases the risk of acute sports injuries. According to the "Guidelines for the Prevention of Weightlifting Injuries" (2023 edition), up to 23.7% of participants in routine 1RM testing suffer injuries such as rotator cuff tears, glenohumeral subluxations, or deltoid strains, with the risk being particularly high among adolescent and female athletes. Such injuries not only interrupt training cycles but can also cause permanent joint dysfunction, seriously affecting an athlete's career.
[0003] The prediction accuracy is low and lacks feedback on movement quality. The widely used Epley empirical formula (1RM=w×(1+0.0333×r)) only performs linear extrapolation based on training weight (w) and repetitions (r), completely ignoring individual differences in force exertion patterns. Empirical studies show that its overall prediction error rate is as high as 28.7%; more seriously, when athletes have compensatory force exertion (such as excessive shoulder involvement to compensate for insufficient lower limb drive), the error can surge to 42.3%, causing training plans to deviate significantly from actual abilities. Existing improvement methods, such as the Brzycki formula and the Lander formula, although showing slight optimization in specific populations, still do not solve the problem of the missing core variable of "movement quality".
[0004] The limited data dimensions fail to reflect the underlying mechanical mechanisms. While recent studies have attempted to incorporate machine learning models, the input features remain limited to raw kinematic parameters (such as joint angles and angular velocities) or electromyographic activation amplitude, failing to reveal the dynamic coupling between the internal forces generated by muscle contraction and the external loads borne by the skeletal system. Furthermore, these models neglect the "muscle-joint" biomechanical balance, resulting in an average prediction bias as high as 22.3%, making it difficult to support refined training interventions. For example, when two athletes perform the same snatch, one relies on efficient lower limb drive (hip-knee coordination), while the other relies on shoulder compensation. Existing models cannot distinguish their true potential, leading to inaccurate training prescriptions.
[0005] A deeper technical deficiency lies in the fact that although biomechanical simulation has been successfully used in sports such as diving, gymnastics, and throwing to optimize aerial posture or release dynamics (as described in patent CN114528901A), those skilled in the art generally believe that weightlifting, as a typical ground-supported, high-load (often exceeding 100% of one's own body weight), low-speed resistance exercise, has fundamentally different mechanical characteristics from aerial movements. Therefore, the "secondary mechanical indicators" (such as joint reaction force and muscle force) obtained from simulation have no direct value in 1RM prediction. Summary of the Invention
[0006] This invention proposes a three-in-one technical approach of "data acquisition - biomechanical simulation calculation - intelligent prediction", and the method flow is as follows: Figure 1 As shown, the secondary mechanical indices calculated by biomechanical simulation are used as key input features of the machine learning model. To achieve the above objective, this invention provides the following solution: 1. A method for predicting weightlifting ability based on dynamic parameters of muscle force and joint load, characterized by comprising the following steps: Collect multi-source heterogeneous data when athletes perform snatch movements, including inertial motion capture data, surface electromyography signal data, and video data; The multi-source heterogeneous data is preprocessed, including kinematic data correction, electromyographic signal filtering and noise reduction, time synchronization, and sampling rate matching. Based on the preprocessed data, a personalized musculoskeletal model was constructed using the OpenSim simulation platform, and inverse dynamics and static optimization algorithms were used to calculate muscle force and joint reaction force. A multimodal fusion prediction model is constructed. Data is divided into time series based on the snatching action. LSTM is used to process dynamic dependencies, including electromyographic time series signals. XGBoost is used to integrate and decompose non-time series features. An XGBoost-LSTM hybrid model is constructed. Muscle force and joint reaction force are input into the XGBoost-LSTM hybrid model after feature engineering. The model outputs 1RM prediction values and compensatory action prompts.
[0007] 2. The weightlifting ability prediction method based on muscle force-joint load dynamic parameters according to claim 1, characterized in that the acquisition of multi-source heterogeneous data specifically includes: Kinematic data of key skeletal points throughout the body are collected using an inertial motion capture system; Electromyographic signals of key muscles were acquired using a wireless surface electromyography system; Time synchronization is achieved through software triggering and event marking, and an optical and acoustic event marker time point is generated when the action begins.
[0008] 3. The method for predicting weightlifting ability based on dynamic parameters of muscle strength and joint load according to claim 1, characterized in that the preprocessing includes: Correct sensor pose offset using kinematic data; The electromyographic signals were subjected to 60Hz notch filtering and normalization. Align the time axis of kinematic and electromyographic data with the moment the barbell leaves the ground as the key point; Resample the electromyography data to match the frame rate of the kinematic data.
[0009] 4. The weightlifting ability prediction method based on muscle force-joint load dynamic parameters according to claim 1, characterized in that the construction of a personalized musculoskeletal model includes: Import the general model into the OpenSim platform; Based on the athlete's height, weight, and body segment dimensions, the model's geometric parameters and mass inertia parameters are scaled up. By matching motion capture markers with virtual markers in the model through inverse kinematics calculations, the joint angle sequence is solved.
[0010] 6. The weightlifting ability prediction method based on muscle force-joint load dynamic parameters according to claim 1, characterized in that the method for calculating muscle force and joint reaction force using inverse dynamics and static optimization algorithms includes: In the OpenSim platform, the multi-rigid-body dynamics equations of the human body are established based on the Newton-Euler formula; The joint forces and moments of the four main joints—hip, knee, ankle, and spine—are solved using dynamic equations. Using the Hill-type muscle-tendon complex model, and under the premise of satisfying joint torque constraints, the objective function is set as minimizing the sum of squares of activation of all muscles. A static optimization algorithm is used to calculate the muscle force of each group of key muscles.
[0011] 6. The weightlifting ability prediction method based on muscle force-joint load dynamic parameters according to claim 1, characterized in that the input features of the XGBoost-LSTM hybrid model include: joint angle, electromyographic timing features, and peak muscle force and joint load growth rate.
[0012] 7. A weightlifting capacity prediction system based on dynamic parameters of muscle force and joint load, applied to the method described in any one of claims 1-8, characterized in that the system comprises: The data acquisition module is used to simultaneously acquire inertial motion capture, electromyography signals, and video data; A preprocessing module, connected to the data acquisition module, is used for data correction, filtering, and time alignment. The modeling and calculation module, connected to the preprocessing module, is used to construct personalized models and solve for muscle forces and joint reaction forces. The predictive analysis module, connected to the modeling and calculation module, is used to generate 1RM prediction values and compensation action prompts.
[0013] 8. The system according to claim 7, wherein the data acquisition module comprises: Inertial motion capture unit; Electromyography (EMG) acquisition unit; Video recording unit; A synchronization control unit, through photoacoustic event marking, achieves time synchronization of multiple devices. According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: it collects multi-source heterogeneous data to calculate secondary mechanical indicators such as muscle strength and joint load, rather than raw physiological data, revealing the dynamic coupling relationship between the internal force generated by muscle contraction and the external load borne by the skeletal system, and uses secondary mechanical indicators such as muscle strength and joint load for weightlifting ability prediction, which significantly improves the prediction accuracy. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a flowchart of the weightlifting ability prediction method based on muscle force-joint load dynamic parameters of the present invention; Figure 2 This is a schematic diagram of the data acquisition scheme; Figure 3 This is a schematic diagram of the module structure of the weightlifting ability prediction system based on muscle force-joint load dynamic parameters of the present invention. Detailed Implementation
[0016] 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, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0017] The purpose of this invention is to provide a method and system for predicting weightlifting ability based on dynamic parameters of muscle force and joint load. Through multi-source heterogeneous data fusion and biomechanical simulation technology, it quantifies the deep mechanical parameters in the snatch movement and achieves 1RM (one-repetition maximum load) prediction and compensatory movement recognition based on a machine learning model. This method effectively solves the problems of high safety risks, low prediction accuracy, and limited data dimensions in traditional 1RM testing. This invention proposes a three-in-one technical approach of "data acquisition—biomechanical simulation calculation—intelligent prediction," and the method flow is as follows: Figure 1 As shown, the secondary mechanical indices calculated by biomechanical simulation are used as key input features of the machine learning model.
[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] The weightlifting ability prediction method based on muscle force-joint load dynamic parameters of this invention includes: S1: Synchronous acquisition of multi-source heterogeneous data.
[0020] Multi-source data of the athlete's snatch movements are simultaneously acquired using an inertial motion capture system, a surface electromyography system, and a high-definition camera. Specifically, this includes: The Noitom Perception Neuron 3.0 inertial motion capture system (sampling rate 200Hz) was used to collect kinematic data of 17 key skeletal points throughout the body, covering the rotation angles and angular velocities of major joints in the trunk, upper limbs, and lower limbs.
[0021] Simultaneously, the electromyographic signals of eight core muscle groups, including the trapezius, deltoid, and gluteus maximus, were recorded using the Tianjin Tech 8-channel ErgoLAB wireless surface electromyography system (sampling rate 1024Hz). The electrode attachment positions strictly followed the SENIAM guidelines.
[0022] Side-view video was recorded at 120fps using a Sony DSC-RX10M4 camera for motion quality assessment and data backup.
[0023] All devices achieve millisecond-level time synchronization through software triggering and photoacoustic event markers, and use the moment the barbell leaves the ground as the unified zero point of time to ensure consistent motion phase.
[0024] S2: Multi-source data preprocessing.
[0025] The collected raw data is preprocessed to ensure data quality and synchronization.
[0026] Manually correct pose shifts in kinematic data caused by sensor interference.
[0027] The electromyography signal was subjected to a 60Hz notch filter to eliminate power frequency interference, high-pass / low-pass filtering, and normalization to suppress physiological noise such as electrocardiogram and skin conductance.
[0028] Using the moment the barbell leaves the ground as the key point, the time axis of kinematic and electromyographic data is aligned with video data. The electromyography data was downsampled to 200Hz using a 15 / 16 ratio using Python to match the frame rate of the kinematic data.
[0029] S3: Personalized skeletal muscle model construction.
[0030] The subject's height, weight, and body segment dimensions were imported into the OpenSim simulation platform and then personalized scaling was performed based on the general model gait2392.
[0031] Adjust the model's geometric dimensions and mass inertia parameters based on measured human body parameters.
[0032] By matching motion capture markers with virtual markers in the model through inverse kinematics calculations, the joint angle sequence is solved.
[0033] An equivalent ground reaction force, represented by a muscle ball, is set between the model and the ground to provide boundary conditions for dynamic calculations.
[0034] S4: Simulation calculation of muscle strength and joint load.
[0035] This step, based on preprocessed multi-source data, uses the biomechanical simulation software OpenSim to calculate the dynamic parameters of muscle force and joint load during the snatch movement. The process includes inverse dynamics analysis, joint reaction force calculation, and static optimization algorithm to solve for muscle force. The specific steps are as follows: First, based on the Newton-Euler equations of rigid body dynamics Establish the biomechanical equilibrium equations for various parts of the human body. Among them, rigid body i angular momentum relative to the center of mass and Rigid bodies i The mass and moment of inertia about the simplified center, and Rigid bodies i linear acceleration and angular acceleration, and Acting on rigid bodies i The force system consists of the principal vectors and principal moments about the simplified center. This force system includes both external forces and moments, as well as constraint forces and moments between and within the rigid body. Let a rigid body i with its center of mass Pi be subjected to constraint forces... and constraint torque , external force and external torque (or constraint reaction force) and constraint reaction moment And gravity Gi. If a coordinate system is established at point O as shown in the figure and used as an inertial basis, according to the Newton-Euler formula, we have: .in, Let X be the linear acceleration of the rigid body. , , These are the angular accelerations of the rigid body rotating about the x-axis, y-axis, and z-axis, respectively. , , These are the moments of inertia of the rigid body about its center of curvature about the x-axis, y-axis, and z-axis, respectively. Forces acting on the rigid body , And the additional torque generated by inertial forces about the moment center. Therefore, for any rigid body i, we have the matrix form of the dynamic equations: .
[0036] Once the external forces, external torques, and motion parameters are determined, the equation can be used to solve for the constraint reactions and torques. In biomechanics, the constraint reactions and torques within human segments are known as joint forces and joint torques.
[0037] The Hill-type muscle-tendon complex model was adopted, taking into account muscle activation, muscle fiber length, contraction speed, and tendon elastic nonlinearity.
[0038] The Hill model is a basic model. The following three papers elaborate on the Hill model. This embodiment uses the improved Hill-type muscle-tendon complex model proposed in the 2000 paper.
[0039] [1]HARDYK ATT. Force and power-velocity relationships in a multi-joint movement[D]. Pennsylvania State: The Pennsylvania State University, 2000:102-123. [2] FENN WO, MARSH BS. Muscular force at different speeds of shortening[J]. The Journal of Physiology, 1935, 85(3):277-297. [3]POLISSAR MJ. Physical chemistry of contractile process in muscle. I. A physicochemical model of contractile mechanism[J]. The American Journal of Physiology, 1952, 168(3):766-811. By employing static optimization algorithms or their extensions, and under the premise of satisfying joint torque constraints, physiologically reasonable optimization objectives are set, such as minimizing the sum of squares of muscle activation or metabolic energy consumption, to estimate the activation level and force output of each muscle.
[0040] S5: Construction of multimodal fusion prediction model.
[0041] Secondary biomechanical indicators such as muscle strength and joint load are fused with kinematic and electromyographic features and input into a machine learning model. Input features include joint angles (such as hip extension angle and trunk forward tilt angle), electromyographic timing features (such as trapezius activation peak time and integrated electromyographic value), and simulation parameters such as peak muscle force and joint load growth rate.
[0042] Model structure: After the data is divided into temporal sequences based on the snatch action, feature engineering is performed using tsfresh, and then a hybrid model of XGBoost and LSTM is constructed. LSTM processes the dynamic dependence of temporal signals such as electromyography (e.g., the peak of deltoid activation is delayed by 12ms from the time the barbell leaves the ground), and XGBoost integrates the decomposed non-temporal features (e.g., the peak of deltoid activation is delayed by 12ms from the time the barbell leaves the ground).
[0043] Output results: The model outputs 1RM prediction values and risk warnings for compensatory movements (such as shoulder compensation and lumbar overload).
[0044] Training optimization: Layered 5-fold cross-validation is adopted, Huber Loss is selected as the loss function, and an early stopping mechanism (patience=10) is introduced to prevent overfitting.
[0045] like Figure 2 As shown, a hardware integration scheme for multi-source heterogeneous data acquisition is presented.
[0046] Hardware components: Inertial motion capture unit: Employs 17 sensor nodes from the Noitom Perception Neuron 3.0 system, distributed as shown in the attached diagram. Figure 2 As shown in (a), key skeletal points such as the trunk and limbs are covered to ensure that kinematic data are collected without any blind spots. Electromyography (EMG) acquisition unit: 8-channel wireless surface EMG system. Electrode placement strictly follows the SENIAM guidelines, as shown in the attached image. Figure 2 (b) Covers 8 core muscle groups including the trapezius, deltoid, and gluteus maximus. Video recording unit: Sony camera shoots from the sagittal plane at 120fps for motion quality verification and data backup.
[0047] Synchronization mechanism: Millisecond-level time alignment of multiple devices is achieved through photoacoustic event tagging, ensuring the consistency of the time axis of kinematic, electromyographic, and video data at the moment the barbell leaves the ground.
[0048] This data acquisition scheme provides high-precision, multi-dimensional input data for subsequent personalized modeling and is the foundation for calculating the "muscle strength-joint load dynamic parameters".
[0049] like Figure 3 As shown, the weightlifting ability prediction system based on muscle force-joint load dynamic parameters of the present invention includes: a data acquisition module 1, a preprocessing module 2, a personalized model construction module 3, a simulation calculation module 4, a prediction analysis module 5, and related data flow design.
[0050] Specifically, the data acquisition module corresponds to the attached Figure 2 The hardware configuration is responsible for acquiring raw data. It includes an inertial motion capture unit, an electromyography (EMG) acquisition unit, and a video recording unit.
[0051] The preprocessing module performs sensor pose correction, electromyographic signal filtering and noise reduction, and time synchronization.
[0052] The personalized model building module completes the construction of personalized models based on measured human body data in the OpenSim platform.
[0053] The personalized model construction module is subjected to inverse dynamics and static optimization calculations in the OpenSim platform.
[0054] The predictive analysis module integrates XGBoost based on time-series partitioning and LSTM model based on time-series data to output 1RM prediction values and compensation action prompts.
[0055] The modules are designed with unidirectional data flow to ensure the closed-loop nature of the data processing logic.
[0056] The 17-channel 200Hz kinematic data acquired by the inertial motion capture unit in data acquisition module 1 is imported into the preprocessing module for pose correction and format conversion.
[0057] The 8-channel 1024Hz electromyographic data acquired by the electromyographic acquisition unit in data acquisition module 1 is imported into the preprocessing module for noise reduction, high-pass and low-pass filtering, and normalization.
[0058] By combining the video information collected by the video recording unit in data acquisition module 1, the time stamp is aligned with the moment when the athlete exerts force (barbell leaves the ground) as the key point, thereby aligning the time axis of kinematic data and electromyographic data.
[0059] Finally, the electromyography (EMG) data was resampled. The frame rate was synchronized by using Python to downsample the data at a ratio of 15 / 16, so that the EMG data sampling rate matched the kinematic data frame rate, thus ensuring that the two were synchronized in time.
[0060] The output of preprocessing module 2 is directly used as the input of the modeling and calculation module, avoiding data redundancy and error propagation.
[0061] In the personalized model building module 3, model scaling is performed. The general whole-body skeletal muscle model gait2392 is imported into the system, and the model's geometric dimensions and mass inertial parameters are scaled based on the athlete's height, weight, and length of each body segment measured by the inertial motion capture unit in the data acquisition module 1, to generate a personalized biomechanical model.
[0062] Subsequently, inverse kinematics calculations are performed, matching the motion capture marker data processed by preprocessing module 2 with the virtual markers of the model to obtain the joint angle sequence that best matches the measured motion. Based on this, a muscle ball is placed between the model and the ground to represent the ground reaction force.
[0063] Then, inverse dynamics and static optimization calculations are performed in simulation calculation module 4.
[0064] The calculation results are exported from the simulation calculation module 4 and input into the prediction analysis module 5 for prediction calculation.
[0065] Example effect: The study used 23 young weightlifters as test subjects and validated the method under loads ranging from 70% to 90% of their 1RM. The results showed: The mean absolute error (MAE) of 1RM prediction is 4.2 kg, which is lower than the error rate of the traditional Epley formula (28.7%). The accuracy rate of recognizing compensatory movements reached 89.5%, effectively providing early warning of lumbar spine overload risk. Model determination coefficient (R²) 2 The R² value was 0.88, significantly outperforming prediction methods that rely solely on kinematic features (R²). 2 =0.62).
[0066] This invention breaks through the limitations of traditional 1RM testing by integrating multi-source data and biomechanical simulation, providing reliable technical support for the scientific and personalized development of weightlifting training.
Claims
1. A method for predicting weightlifting ability based on dynamic parameters of muscle force and joint load, characterized in that, Includes the following steps: Collect multi-source heterogeneous data when athletes perform snatch movements, including inertial motion capture data, surface electromyography signal data, and video data; The multi-source heterogeneous data is preprocessed, including kinematic data correction, electromyographic signal filtering and noise reduction, time synchronization, and sampling rate matching. Based on the preprocessed data, a personalized musculoskeletal model was constructed using the OpenSim simulation platform, and inverse dynamics and static optimization algorithms were used to calculate muscle force and joint reaction force. A multimodal fusion prediction model is constructed. Data is divided into time series based on the snatching action. LSTM is used to process dynamic dependencies, including electromyographic time series signals. XGBoost is used to integrate and decompose non-time series features. An XGBoost-LSTM hybrid model is constructed. Muscle force and joint reaction force are input into the XGBoost-LSTM hybrid model after feature engineering. The model outputs 1RM prediction values and compensatory action prompts.
2. The method for predicting weightlifting ability based on dynamic parameters of muscle force and joint load according to claim 1, characterized in that, The collection of multi-source heterogeneous data specifically includes: Kinematic data of key skeletal points throughout the body are collected using an inertial motion capture system; Electromyographic signals of key muscles were acquired using a wireless surface electromyography system; Time synchronization is achieved through software triggering and event marking, and an optical and acoustic event marker time point is generated when the action begins.
3. The method for predicting weightlifting ability based on dynamic parameters of muscle strength and joint load according to claim 1, characterized in that, The preprocessing includes: Correct sensor pose offset using kinematic data; The electromyographic signals were subjected to 60Hz notch filtering and normalization. Align the time axis of kinematic and electromyographic data with the moment the barbell leaves the ground as the key point; Resample the electromyography data to match the frame rate of the kinematic data.
4. The method for predicting weightlifting ability based on dynamic parameters of muscle strength and joint load according to claim 1, characterized in that, The construction of the personalized skeletal muscle model includes: Import the general model into the OpenSim platform; Based on the athlete's height, weight, and body segment dimensions, the model's geometric parameters and mass inertia parameters are scaled up. By matching motion capture markers with virtual markers in the model through inverse kinematics calculations, the joint angle sequence is solved.
5. The method for predicting weightlifting ability based on dynamic parameters of muscle strength and joint load according to claim 1, characterized in that, Methods for calculating muscle force and joint reaction force using inverse dynamics and static optimization algorithms include: In the OpenSim platform, the multi-rigid-body dynamics equations of the human body are established based on the Newton-Euler formula; The joint forces and moments of the four main joints—hip, knee, ankle, and spine—are solved using dynamic equations. Using the Hill-type muscle-tendon complex model, and under the premise of satisfying joint torque constraints, the objective function is set as minimizing the sum of squares of activation of all muscles. A static optimization algorithm is used to calculate the muscle force of each group of key muscles.
6. The method for predicting weightlifting ability based on dynamic parameters of muscle force and joint load according to claim 1, characterized in that, The input features of the XGBoost-LSTM hybrid model include: joint angles, electromyographic timing features, peak muscle force, and joint load growth rate.
7. A weightlifting capacity prediction system based on dynamic parameters of muscle force and joint load, applied to the method described in any one of claims 1-8, characterized in that, The system includes: The data acquisition module is used to simultaneously acquire inertial motion capture, electromyography signals, and video data; A preprocessing module, connected to the data acquisition module, is used for data correction, filtering, and time alignment. The modeling and calculation module, connected to the preprocessing module, is used to construct personalized models and solve for muscle forces and joint reaction forces. The predictive analysis module, connected to the modeling and calculation module, is used to generate 1RM prediction values and compensation action prompts.
8. The system according to claim 7, characterized in that, The data acquisition module includes: Inertial motion capture unit; Electromyography (EMG) acquisition unit; Video recording unit; The synchronization control unit achieves time synchronization of multiple devices through photoacoustic event markers.