Regional muscle load grade prediction method and device based on deep learning

Abstract

The invention relates to a regional muscle load grade prediction method and device based on deep learning. The method comprises the following steps: acquiring original motion data captured by inertial motion; carrying out standardized pretreatment on the sample; driving an individualized biomechanical simulation model by using the preprocessed data, and generating a high-fidelity muscle load grade label data set through muscle force calculation, regional aggregation and grade discretization; training a deep learning model according to the data set; and finally, using the trained model to quickly predict new motion data, and outputting the load level of each muscle area. According to the method, through an integrated process, the contradiction between label scarcity and real-time requirements in non-intrusive muscle load assessment is solved, and efficient and interpretable muscle load assessment is realized.

Description

Technical Field

[0001] This application relates to the intersection of biomechanics and artificial intelligence technologies, and in particular to a method and device for predicting regional muscle load levels based on deep learning. Background Technology

[0002] In fields such as rehabilitation medicine, ergonomics, virtual reality / augmented reality (VR / AR), and sports science, real-time, accurate, and non-invasive assessment of the load level of human muscles during exercise is crucial for optimizing training programs, preventing injuries, improving work posture, and achieving highly realistic virtual interactions.

[0003] Traditional methods for obtaining muscle load levels mainly fall into two categories. One category is invasive or contact-based measurements, such as surface electromyography (sEMG). This method collects muscle electrical signals through electrodes attached to the skin surface, thereby estimating muscle activation and strength. However, sEMG signals are susceptible to factors such as skin impedance, electrode placement, and crosstalk, and require complex calibration and models to be converted into muscle force, resulting in high equipment costs, cumbersome setup, and a poor user experience. The other category is simulation methods based on physical models, such as using biomechanical simulation software like AnyBody Modeling System (AMS) and OpenSim. These methods are based on multibody dynamics, inverse dynamics, and musculoskeletal models, using optimized algorithms to decompose the measured kinematic data into the force of each muscle. While offering high accuracy, the simulation process is extremely complex and time-consuming, heavily reliant on individualized and precise modeling and parameter calibration, making it difficult to meet the demands of real-time, online applications.

[0004] In recent years, with the development of wearable sensors and artificial intelligence technologies, data-driven muscle load prediction methods have emerged. These methods attempt to bypass complex physical simulations and directly establish a mapping relationship from easily accessible external motion data to internal muscle load. However, the core challenge facing these methods is the scarcity of high-quality supervisory labels. Real muscle force is difficult to measure directly and non-invasively, while relying on simulation software to generate labels faces a contradiction between computational cost and efficiency. In addition, existing studies often focus on specific limbs or single tasks, lacking universality that covers multiple regions and scenarios throughout the body. The output of the models is mostly continuous force values, which are disconnected from the hierarchical assessment methods commonly used in clinical practice, resulting in insufficient interpretability. Therefore, how to construct an efficient, universal, interpretable, and non-invasive end-to-end muscle load assessment system while ensuring prediction accuracy and physiological rationality remains an urgent technical problem to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and device for predicting regional muscle load levels based on deep learning to address the above-mentioned technical problems.

[0006] A deep learning-based method for predicting regional muscle load levels, the method comprising:

[0007] Acquire raw human motion data collected by an inertial motion capture system; the raw human motion data includes angle information of multiple joints; The raw human motion data is standardized and preprocessed to generate driving data compatible with the target biomechanical simulation software; The driving data is input into the biomechanical simulation process of the individualized musculoskeletal model. Through muscle force calculation and region aggregation, muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data are generated. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. Using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, a deep learning model is trained to obtain a deep learning model that can map joint movement to regional muscle load levels. The motion data to be evaluated is input into the trained deep learning model after the standardization preprocessing, and the load level prediction results of each muscle region are output.

[0008] In one embodiment, the method further includes: using a spherical linear interpolation algorithm to correct continuous data frames in the original human motion data that exhibit sensor drift, wherein the interpolation formula is:

[0009] Among them Interpolation factor; The quaternion represented by the data frame , The directional angle.

[0010] In one embodiment, the method further includes: performing inverse dynamics and static optimization calculations based on the Hill-type muscle model and the minimum-maximum fatigue recruitment criterion, with the objective function being:

[0011]

[0012] in, To use muscle strength Let M be the objective function of the variable M, and let M represent the number of muscles. The goal of optimization is to increase muscle strength. minimize; The constant coefficient matrix of the unknown forces; This represents the combined muscle force and joint torque; Given the known external forces and inertial forces; For the first Muscle strength of a muscle group; For the first The ultimate strength of a muscle mass; Indicates the first The activation level of the mass muscle; The calculated muscle forces are aggregated according to the preset anatomical regions to obtain the muscle load values ​​for each region. After normalizing the muscle load values ​​of each region, they are discretized into integer levels within a preset range, which serve as the muscle load level labels.

[0013] In one embodiment, it further includes: for the first Each muscle region has a continuous muscle force denoted as . The minimum and maximum values ​​of muscle strength in this area during the training session are denoted as follows: , ,right Perform linear normalization to obtain normalized variables. :

[0014] in ,Will Mapping to discrete levels from 1 to 10 yields level labels. :

[0015] in, For the first Force level labels for individual muscle regions This represents the floor operator.

[0016] In one embodiment, the deep learning model is further comprising: a multilayer perceptron model, including: an input layer with the number of neurons matching the dimension of the joint angle features; at least one hidden layer with Randomized ReLU activation function and Dropout layers inserted between layers; and an output layer with the number of neurons matching the number of muscle regions, used to output the predicted load level values ​​for each region.

[0017] In one embodiment, the number of anatomical regions is 12, including the chest, back, abdomen, waist, left upper arm, right upper arm, left forearm, right forearm, left thigh, right thigh, left calf, and right calf.

[0018] In one embodiment, the method further includes: selecting at least one typical human movement posture as a verification condition; By using the forward computation of the trained deep learning model and the offline computation of the biomechanical simulation process, the region prediction level and reference level of each muscle region under typical human movement postures are obtained. Calculate the absolute error between the predicted level and the reference level for all regions; If the absolute error of all regions does not exceed the preset threshold, then the accuracy of the deep learning model is determined to meet the requirements.

[0019] A deep learning-based regional muscle load level prediction device, the device comprising: The data acquisition and standardization module is used to acquire raw human motion data collected by the inertial motion capture system; the raw human motion data includes angle information of multiple joints; the raw human motion data is preprocessed and standardized to generate driving data compatible with the target biomechanical simulation software. The simulation label generation module is used to input the driving data into the biomechanical simulation process of the individualized musculoskeletal model, and generate muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data through muscle force calculation and region aggregation. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. The model training module is used to train a deep learning model using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, to obtain a deep learning model that can map joint movement to regional muscle load levels. The prediction module is used to input the motion data to be evaluated into the trained deep learning model after the standardization preprocessing, and output the load level prediction results for each muscle region.

[0020] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire raw human motion data collected by an inertial motion capture system; the raw human motion data includes angle information of multiple joints; The raw human motion data is standardized and preprocessed to generate driving data compatible with the target biomechanical simulation software; The driving data is input into the biomechanical simulation process of the individualized musculoskeletal model. Through muscle force calculation and region aggregation, muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data are generated. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. Using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, a deep learning model is trained to obtain a deep learning model that can map joint movement to regional muscle load levels. The motion data to be evaluated is input into the trained deep learning model after the standardization preprocessing, and the load level prediction results of each muscle region are output.

[0021] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire raw human motion data collected by an inertial motion capture system; the raw human motion data includes angle information of multiple joints; The raw human motion data is standardized and preprocessed to generate driving data compatible with the target biomechanical simulation software; The driving data is input into the biomechanical simulation process of the individualized musculoskeletal model. Through muscle force calculation and region aggregation, muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data are generated. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. Using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, a deep learning model is trained to obtain a deep learning model that can map joint movement to regional muscle load levels. The motion data to be evaluated is input into the trained deep learning model after the standardization preprocessing, and the load level prediction results of each muscle region are output.

[0022] The aforementioned deep learning-based method and apparatus for predicting regional muscle load levels solves the core problem of scarce training data labels by pre-processing the high-cost biomechanical simulation process into a one-time, offline labeling process. Furthermore, using the generated high-fidelity, graded label dataset, a lightweight deep learning model is trained. In practical applications, simply running this model is sufficient to quickly obtain interpretable muscle load levels. This method creatively constructs an integrated end-to-end process encompassing data acquisition, simulation modeling, model training, and result verification. Under the premise of non-invasiveness and low cost, it balances the physiological rationality of the prediction results, computational efficiency, and output interpretability, effectively overcoming the limitations of traditional methods in terms of real-time performance, universality, and ease of use. Attached Figure Description

[0023] Figure 1 This is an application scenario diagram of a deep learning-based regional muscle load level prediction method in one embodiment; Figure 2This is a flowchart illustrating a deep learning-based method for predicting regional muscle load levels in one embodiment. Figure 3 This is a structural block diagram of a deep learning-based regional muscle load level prediction device in one embodiment. Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] The deep learning-based regional muscle load level prediction method provided in this application can be applied to, for example... Figure 1 The application environment is illustrated. The user wears an inertial motion capture suit 101 (such as a VD suit), which integrates multiple inertial measurement units (IMUs) for real-time acquisition of motion angle data from various joints of the human body. A data acquisition terminal 102 (e.g., a portable computer) receives the raw motion data wirelessly or via wired connection and performs preliminary processing and forwarding. A server 104 (which can be a local server or a cloud server cluster) communicates with the terminal 102 via a network, responsible for running biomechanical simulation software to generate training labels or deploying trained deep learning models for real-time prediction. The roles of the terminal 102 and server 104 can be adjusted according to actual deployment needs; for example, in resource-rich edge computing scenarios, all processes can be integrated into a single high-performance terminal.

[0026] In one embodiment, such as Figure 2 As shown, a deep learning-based method for predicting regional muscle load levels is provided, which can be applied to... Figure 1 Taking the server in the example, the following steps are included: Step 202: Obtain raw human motion data collected by the inertial motion capture system.

[0027] This step involves acquiring data using an inertial motion capture system. For example, a commercially available inertial motion capture suit such as the VD suit is used to collect full-body motion data of the subject in various preset tasks or daily activities at a fixed sampling rate. These activities include walking, squatting, carrying, and raising arms. The system directly outputs a BVH (BioVision Hierarchy) file containing complete skeletal hierarchical information. The BVH file is a common format for describing skeletal animation, containing joint offsets and Euler angle rotation data for each frame, thus defining the posture of the human skeleton at each moment. The raw data contains angular information from multiple joints, from the root joints to the end effectors.

[0028] Step 204: Standardize and preprocess the raw human motion data to generate driving data compatible with the target biomechanical simulation software.

[0029] Raw human motion data typically cannot be directly used for biomechanical simulations due to two main issues: first, the names of skeletal joints defined by the motion capture system are inconsistent with the bone naming conventions of the target simulation software (such as AMS); second, sensor noise and drift can cause clipping of skeletal models or abnormal jitter in the data. The purpose of this step is to solve these two problems and generate clean, standardized motion data sequences that are fully compatible with the simulation software, serving as the driving input for subsequent simulations.

[0030] Specifically, the preprocessing step may include: Establish a mapping table between the skeleton names in the motion capture system and the skeleton names in the target simulation software, and perform batch replacement of joint names in the BVH file. For example, map "Pelvis" in the motion capture data to "Hips" in the AMS model.

[0031] For abnormal frames exhibiting obvious skeletal clipping or data jitter, manual or rule-based automatic correction can be performed. More importantly, for smoothness errors caused by sensor drift that do not meet kinematic constraints, a quaternion-based spherical linear interpolation (Slerp) algorithm is used for repair. This algorithm can smoothly interpolate rotations on a three-dimensional sphere, maintaining the physical correctness of rotations and velocity uniformity better than linear interpolation Euler angles.

[0032] Through the above processing, standardized and smoothed motion data are obtained, and its data format, frame rate, and rotation order all meet the input requirements of simulation software such as AMS.

[0033] Step 206: Input the driving data into the biomechanical simulation process of the individualized musculoskeletal model, and generate muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data through muscle force calculation and region aggregation. Construct a high-fidelity label dataset based on the driving data and muscle load level labels.

[0034] This step is one of the core innovations of this invention, aiming to utilize biomechanical simulation to generate high-quality, physiologically meaningful supervisory labels required for machine learning. This step includes: In the AMS software, an individualized model is established by adjusting the default human musculoskeletal model based on the subject's actual height, weight, and other parameters. This model includes bones, joints, and multiple muscles using a Hill-type model. The standardized motion data (BVH format) obtained in step 204 is used as the kinematic driver and input into the AMS model. The software performs inverse dynamics analysis to calculate joint torques, and then distributes the joint torques to each muscle based on static optimization and specific muscle recruitment criteria, calculating the activation and muscle force of each muscle at each time step. This optimization process aims to find the combination of muscle forces that minimizes the sum of squares of the total muscle activation; its objective function reflects the efficiency principle of muscle synergistic exertion in physiology.

[0035] Directly predicting the force of each individual muscle results in a model with extremely high output dimensionality and is difficult to train. Based on muscle synergy theory, muscles with similar functions will be co-activated in specific tasks. Therefore, this application proposes a region-level load prediction strategy. Based on anatomical location and function, the major muscles of the body are divided into multiple regions. For each region, the forces of all muscles within that region at the same time are aggregated to obtain a continuous value representing the overall load level of that region.

[0036] To make the model output more interpretable and correspond to clinically used muscle strength grading (such as MRC grading) or ergonomic load levels, continuous load values ​​are discretized into finite integer levels. Specifically, for each region, the maximum and minimum load values ​​are found across all training samples. Then, the load value for that region in each sample is linearly normalized. Finally, the normalized value is mapped to an integer level of 1-10 (level 10 represents the maximum load observed in the training set for that region, and level 1 represents the minimum load). Thus, each frame of motion data corresponds to a 12-dimensional integer level vector label.

[0037] All standardized motion data were paired with corresponding muscle load level labels to construct a structured high-fidelity dataset for subsequent model training.

[0038] Step 208: Using a high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, train a deep learning model to obtain a deep learning model that can map joint movement to regional muscle load levels.

[0039] This step utilizes the dataset constructed in S206 to train a model capable of learning the complex mapping relationship from joint angles to muscle load levels. Specifically: Considering the highly nonlinear spatiotemporal mapping relationship between input joint angles and output levels in muscle load level prediction tasks, the spatial and temporal dependencies are simplified to some extent after region partitioning and grading. A preferred embodiment of this application employs a Multilayer Perceptron (MLP) as the prediction model. MLPs have a simple structure, high parameter efficiency, and are adept at learning complex static nonlinear mappings. The number of nodes in its input layer equals the dimension of the joint angle features, and the number of nodes in its output layer equals the number of muscle regions. Multiple hidden layers can be set in between, using activation functions such as Randomized ReLU to enhance nonlinear expressive power, and a Dropout layer is introduced to prevent overfitting.

[0040] The constructed dataset is divided into training, validation, and test sets. The Adam optimizer is used to train the MLP model with the cross-entropy loss and mean squared error (MSE) between the predicted and true rank as the loss function. Performance on the validation set is monitored during training until the model converges.

[0041] Step 210: After standardization and preprocessing, the motion data to be evaluated is input into the trained deep learning model, which outputs the load level prediction results for each muscle region.

[0042] Once the model is trained, it can be used to predict muscle load levels in real-time or offline on new human motion data. The new motion data to be evaluated first undergoes the same standardized preprocessing procedure as S204 to transform it into driving data. This data is then input into the trained deep learning model, which can directly output the load level (level 1-10) of each muscle region in the current posture end-to-end.

[0043] The aforementioned deep learning-based method for predicting regional muscle load levels addresses the core challenge of scarce training data labels by pre-processing the costly biomechanical simulation process into a one-time, offline labeling process. Furthermore, a lightweight deep learning model is trained using the generated high-fidelity, pre-graded labeled dataset. In practical applications, simply running this model quickly yields interpretable muscle load levels. This method creatively constructs an integrated end-to-end process encompassing data acquisition, simulation modeling, model training, and result verification. Under the premise of non-intrusiveness and low cost, it balances the physiological rationality of the prediction results, computational efficiency, and output interpretability, effectively overcoming the limitations of traditional methods in terms of real-time performance, universality, and ease of use.

[0044] In one embodiment, a spherical linear interpolation algorithm is used to correct continuous data frames in the original human motion data that exhibit sensor drift. The interpolation formula is as follows:

[0045] Among them Interpolation factor; The quaternion represented by the data frame , The directional angle.

[0046] This embodiment can effectively correct the jitter caused by data drift and generate a smooth, continuous and physiologically reasonable motion sequence, providing high-quality input for subsequent simulation. This is a key preprocessing step to ensure the quality of the labeling throughout the process.

[0047] In one embodiment, inverse dynamics and static optimization calculations are performed based on the Hill-type muscle model and the minimum-maximum fatigue recruitment criterion. The objective function is:

[0048]

[0049] in, To use muscle strength Let M be the objective function of the variable M, and let M represent the number of muscles. The goal of optimization is to increase muscle strength. minimize; The constant coefficient matrix of the unknown forces; This represents the combined muscle force and joint torque; Given the known external forces and inertial forces; For the first Muscle strength of a muscle group; For the first The ultimate strength of a muscle mass; Indicates the first The activation level of the muscle mass is calculated; the calculated muscle force is aggregated according to the preset anatomical region division to obtain the muscle load value of each region; after normalizing the muscle load value of each region, it is discretized into integer levels within a preset range as muscle load level labels.

[0050] The physical meaning of the objective function in this embodiment is to find a set of muscle forces that minimizes the sum of squares of activation of all muscles, under the constraints of satisfying the equation of motion (i.e., the generated joint torque is equal to the torque required for inverse dynamics calculation) and the constraints that muscle forces are non-negative and less than the maximum isometric contraction force. This simulates the physiological assumption that the human nervous system tends to minimize overall muscle fatigue when performing a specific action.

[0051] In one embodiment, for the first Each muscle region has a continuous muscle force denoted as . The minimum and maximum values ​​of muscle strength in this area during the training session are denoted as follows: , ,right Perform linear normalization to obtain normalized variables. :

[0052] in ,Will Mapping to discrete levels from 1 to 10 yields level labels. :

[0053] in, For the first Force level labels for individual muscle regions This represents the floor operator.

[0054] In this embodiment, after obtaining the strength of each muscle, each frame of data generates a fixed-dimensional level label vector with clear ordinal meaning, which greatly facilitates the supervised training of the deep learning model and the intuitive understanding of the results.

[0055] In one embodiment, the deep learning model is a multilayer perceptron model, comprising: an input layer with the same number of neurons as the joint angle feature dimension; at least one hidden layer with a Randomized ReLU activation function and a Dropout layer inserted between layers; and an output layer with the same number of neurons as the number of muscle regions, used to output the predicted load level values ​​for each region.

[0056] In this embodiment, the key advantage is that, through the aforementioned process of regional division and hierarchical classification, the originally extremely complex biomechanical mapping problem is transformed into a relatively standard multivariate regression / classification problem, which can then be understood by deep learning models.

[0057] In one embodiment, the number of anatomical regions is 12, including the chest, back, abdomen, waist, left upper arm, right upper arm, left forearm, right forearm, left thigh, right thigh, left calf, and right calf.

[0058] Specifically, the muscle region allocation is shown in Table 1.

[0059] Table 1 Muscle Region Allocation Table

[0060] In another embodiment, the verification process includes: selecting at least one typical human motion posture as the verification condition; obtaining the region prediction level and reference level of each muscle region under the typical human motion posture through forward computation of the trained deep learning model and offline computation of the biomechanical simulation process; calculating the absolute error between the prediction level and the reference level of all regions; if the absolute error of all regions does not exceed the preset threshold, the accuracy of the deep learning model is determined to meet the requirements.

[0061] The validation method proposed in this embodiment possesses internal consistency and engineering applicability. Typical movement postures, sourced from the same training data but not involved in the training, are used as test items. The model's performance is quantified by comparing the model's predicted answers with the standard answers calculated through a complete biomechanical simulation process. Calculating the absolute error is simple and intuitive, reflecting the absolute distance between the predicted level and the true level. In ergonomic assessments or rehabilitation monitoring, the approximate range of load levels is more important than the precise force value; a level 1 error typically does not lead to misjudgments of work condition risk or muscle status. This validation method not only assesses the model's accuracy but also verifies the feasibility and reliability of the entire process.

[0062] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0063] In one embodiment, such as Figure 3 As shown, a deep learning-based regional muscle load level prediction device is provided, comprising: a data acquisition and normalization module 302, a simulation label generation module 304, a model training module 306, and a prediction module 308, wherein: The data acquisition and standardization module 302 is used to acquire raw human motion data collected by the inertial motion capture system; the raw human motion data includes angle information of multiple joints; the raw human motion data is preprocessed and standardized to generate driving data compatible with the target biomechanical simulation software; The simulation label generation module 304 is used to input the driving data into the biomechanical simulation process of the individualized musculoskeletal model, and generate muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data through muscle force calculation and region aggregation, and construct a high-fidelity label dataset based on the driving data and the muscle load level labels. The model training module 306 is used to train a deep learning model using the high-fidelity labeled dataset, with the joint angle in the driving data as the input feature and the muscle load level as the supervision label, to obtain a deep learning model that can map joint movement to the muscle load level of the region. The prediction module 308 is used to input the motion data to be evaluated into the trained deep learning model after the standardization preprocessing, and output the load level prediction results of each muscle region.

[0064] Specific limitations regarding the deep learning-based regional muscle load level prediction device can be found in the limitations of the deep learning-based regional muscle load level prediction method described above, and will not be repeated here. Each module in the aforementioned deep learning-based regional muscle load level prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0065] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores motion capture data, simulation tag datasets, model parameters, etc. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting regional muscle load levels.

[0066] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0067] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0068] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting regional muscle load levels based on deep learning, characterized in that, The method includes: Acquire raw human motion data collected by an inertial motion capture system; the raw human motion data includes angle information of multiple joints; The raw human motion data is standardized and preprocessed to generate driving data compatible with the target biomechanical simulation software; The driving data is input into the biomechanical simulation process of the individualized musculoskeletal model. Through muscle force calculation and region aggregation, muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data are generated. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. Using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, a deep learning model is trained to obtain a deep learning model that can map joint movement to regional muscle load levels. The motion data to be evaluated is input into the trained deep learning model after the standardization preprocessing, and the load level prediction results of each muscle region are output.

2. The method according to claim 1, characterized in that, The raw human motion data undergoes standardized preprocessing, including: The spherical linear interpolation algorithm is used to correct continuous data frames in the original human motion data that exhibit sensor drift. The interpolation formula is as follows: Among them Interpolation factor; The quaternion represented by the data frame , The directional angle.

3. The method according to claim 1, characterized in that, Generate muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data, including: Inverse dynamics and static optimization calculations are performed based on the Hill-type muscle model and the minimum-maximum fatigue recruitment criterion. The objective function is: in, To use muscle strength Let M be the objective function of the variable M, and let M represent the number of muscles. The goal of optimization is to increase muscle strength. minimize; The constant coefficient matrix of the unknown forces; This represents the combined muscle force and joint torque; Given the known external forces and inertial forces; For the first Muscle strength of a muscle group; For the first The ultimate strength of a muscle mass; Indicates the first The activation level of the mass muscle; The calculated muscle forces are aggregated according to the preset anatomical regions to obtain the muscle load values ​​for each region. After normalizing the muscle load values ​​of each region, they are discretized into integer levels within a preset range, which serve as the muscle load level labels.

4. The method according to claim 3, characterized in that, After normalizing the muscle load values ​​for each region, they are discretized into integer levels within a preset range, including: For the first Each muscle region, with continuous muscle force recorded as... The minimum and maximum values ​​of muscle strength in this area during the training session are denoted as follows: , For continuous muscle force Perform linear normalization to obtain normalized variables. : in , will continuously muscle force Mapping to discrete levels from 1 to 10 yields level labels. : in, For the first Force level labels for individual muscle regions This represents the floor operator.

5. The method according to claim 1, characterized in that, The deep learning model is a multilayer perceptron model, comprising: an input layer with the number of neurons matching the dimension of the joint angle features; at least one hidden layer with Randomized ReLU activation function and Dropout layers inserted between layers; and an output layer with the number of neurons matching the number of muscle regions, used to output the predicted load level values ​​for each region.

6. The method according to claim 3 or 4, characterized in that, The number of anatomical regions is 12, including the chest, back, abdomen, waist, left upper arm, right upper arm, left forearm, right forearm, left thigh, right thigh, left calf, and right calf.

7. The method according to claim 1, characterized in that, The method further includes: Select at least one typical human movement posture as the verification condition. By using the forward computation of the trained deep learning model and the offline computation of the biomechanical simulation process, the region prediction level and reference level of each muscle region under typical human movement postures are obtained. Calculate the absolute error between the predicted level and the reference level for all regions; If the absolute error of all regions does not exceed the preset threshold, then the accuracy of the deep learning model is determined to meet the requirements.

8. A deep learning-based regional muscle load level prediction device, characterized in that, The device includes: The data acquisition and standardization module is used to acquire raw human motion data collected by the inertial motion capture system; the raw human motion data includes angle information of multiple joints; the raw human motion data is standardized and preprocessed to generate driving data compatible with the target biomechanical simulation software; The simulation label generation module is used to input the driving data into the biomechanical simulation process of the individualized musculoskeletal model, and generate muscle load level labels covering multiple muscle regions corresponding to each moment in the driving data through muscle force calculation and region aggregation. A high-fidelity label dataset is constructed based on the driving data and the muscle load level labels. The model training module is used to train a deep learning model using the high-fidelity labeled dataset, with joint angles in the driving data as input features and muscle load levels as supervision labels, to obtain a deep learning model that can map joint movement to regional muscle load levels. The prediction module is used to input the motion data to be evaluated into the trained deep learning model after the standardization preprocessing, and output the load level prediction results for each muscle region.

Images