A joint torque prediction method fusing biomechanical information, medium and device

Abstract

The present application relates to the technical field of healthcare informatics, and specifically provides a joint torque prediction method, medium and equipment fused with biomechanical information; the method is to predict joint torque by using a joint torque prediction neural network model embedded with biomechanical information; a personalized biomechanical model is embedded in the joint torque prediction neural network model; when training the joint torque prediction neural network model, the joint torque prediction neural network model and the personalized biomechanical model are cooperatively driven and parallelly optimized based on data errors and physical residual errors, and bidirectional back transmission of gradients between the joint torque prediction neural network model and the personalized biomechanical model is realized. The intelligent estimation framework fused with biomechanical prior knowledge realizes efficient and physiologically reasonable torque prediction.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and more specifically, to a method, medium, and device for predicting joint torque by integrating biomechanical information. Background Technology

[0002] Against the backdrop of rapid development in biomedical engineering and rehabilitation technology, precise biomechanical analysis has become a core requirement in fields such as rehabilitation treatment assessment, athlete performance optimization, and the development of assistive devices for people with disabilities. Human movement involves complex interactions of the neuromuscular system, where key variables such as muscle strength and joint kinematic parameters are difficult to measure directly in vivo. Musculoskeletal models, as core simulation tools, can achieve non-invasive estimation of these variables, providing important technical support for related applications. In the fields of rehabilitation treatment and sports biomechanics, the accuracy of musculoskeletal modeling directly affects the personalized development of rehabilitation programs, the prediction of sports injury risks, and the adaptation effect of assistive devices. For example, gait rehabilitation training for stroke patients requires adjusting the assistive strategies of rehabilitation robots based on dynamic predictions of muscle strength and joint movement; in athlete training, musculoskeletal models are needed to analyze the impact of movement posture on muscle load in order to optimize training programs and reduce injury risks. These scenarios all place stringent requirements on the real-time response capability, prediction accuracy, and physical rationality of musculoskeletal models.

[0003] As a crucial link in the movement and force transmission of the lower limbs, the accurate prediction of joint torque is of significant clinical and engineering value for gait rehabilitation training in stroke patients, prediction of sports injury risks, and control of lower limb exoskeleton robots. Currently, mainstream joint torque prediction methods are mainly divided into two categories: physical-based modeling and data-driven modeling.

[0004] On the one hand, while traditional physics-based modeling methods (such as inverse dynamics methods) have clear biomechanical interpretability, they rely on complex dynamic derivations and static optimization solutions, and need to address the redundancy issues of the neuromuscular skeletal system. For example, the Chinese patent "Method and Apparatus for Torque Estimation" (Publication No.: CN111347421A) has expensive data acquisition equipment, cumbersome operation, slow model calculation speed, and high running latency, making it difficult to meet the application requirements of real-time control of rehabilitation robots.

[0005] On the other hand, emerging data-driven modeling methods attempt to achieve rapid torque inference through deep learning. Examples include the Chinese patent "A Neural Network-Based Exoskeleton Robot Assistive Control System and Method" (Publication No.: CN110653817A) and the Chinese patent "Method, Device, Equipment and Storage Medium for Joint Torque Estimation Based on Electromyographic Signals of Lower Limb Exoskeleton" (Publication No.: CN119848487A). However, these methods are essentially "black box" models, lacking integration with the underlying physical mechanisms of the musculoskeletal system. They suffer from significant individual generalization defects and accuracy bottlenecks, their prediction results may violate biomechanical principles, and they are highly dependent on training data. Furthermore, they exhibit weak adaptability across individuals, across movement modes (such as walking, climbing stairs, turning, etc.), and in scenarios with dynamic changes in movement speed.

[0006] Furthermore, existing models either suffer from complex structures leading to difficulties in maintenance and optimization, or lack interpretability due to a lack of physical constraints. The core model structure of some wearable lower limb exoskeleton rehabilitation robots is fixed, lacking an adaptive adjustment mechanism for dynamic changes in different movement patterns and speeds, and thus unable to flexibly match the biomechanical characteristics of the human body in different movement scenarios. To address this issue, the Chinese patent "Method and Device for Joint Torque Prediction Based on Electromyography and Musculoskeletal Physiological Information" (Publication No.: CN119517298A) first constructs and optimizes a neuromuscular skeletal model, fixes the parameters of the model, and then uses these parameters to generate "soft labels," which, together with the actual torque, construct a fusion label for training the neural network. However, this approach uses a serial architecture, using the musculoskeletal physiological model as a preprocessor to correct sensor noise; while easy to implement, identification errors can easily lead to label contamination, and the two-stage structure struggles to balance individual generalization and the limitations of physical constraint expansion.

[0007] In summary, current joint torque prediction methods still have significant limitations in terms of real-time performance, physical plausibility, generalization ability, and adaptive adjustment mechanisms. There is an urgent need to develop a new joint torque prediction method that combines physical interpretability and fast reasoning ability, and can adapt to multi-mode and variable-speed motion scenarios. Summary of the Invention

[0008] To overcome the shortcomings and deficiencies of the existing technology, the present invention aims to provide a method, medium and device for predicting joint torque by integrating biomechanical information. The method integrates an intelligent estimation framework based on prior biomechanical knowledge, which realizes efficient and physiologically reasonable torque prediction and can meet the needs of practical application scenarios such as rehabilitation treatment assessment, athlete performance improvement and on-demand assistive robot design optimization.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: a joint torque prediction method integrating biomechanical information, which uses a joint torque prediction neural network model embedded with biomechanical information to predict joint torque; the joint torque prediction neural network model is embedded with a personalized biomechanical model; when training the joint torque prediction neural network model, the joint torque prediction neural network model and the personalized biomechanical model are optimized in parallel based on the collaborative driving of data error and physical residual, so as to realize the bidirectional backpropagation of gradient between the joint torque prediction neural network model and the personalized biomechanical model.

[0010] Preferably, the training of the joint torque prediction neural network model includes the following steps:

[0011] Step S1: Obtain electromyographic signals and kinematic parameters of lower limb joint movements from the sample population to construct a sample set; and obtain synchronous real joint torque. ;

[0012] Step S2: Construct a personalized biomechanical model based on the constraint relationship between joint torque, electromyographic signals, and kinematic parameters, and separate the personalized calibration parameters adapted to the individual in the personalized biomechanical model.

[0013] Step S3: The joint torque prediction neural network model extracts temporal and spatial dimension information from the input sample information of the sample set, and performs feature fusion on the temporal and spatial dimension information and the initial input sample information, and then outputs the predicted joint torque. ;

[0014] Step S4: Input the sample information into the personalized biomechanical model for calculation to obtain the joint torque. ; to joint torque With joint torque Perform error calculation; calculate the joint torque. With actual joint torque Error calculation is performed; the two error calculation results are then self-attentionally fused into a comprehensive loss; the comprehensive loss of batch data is used to continuously update the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model, so as to achieve convergence of the comprehensive error.

[0015] Preferably, the personalized biomechanical model includes: a neural activation model, a muscle activation model, a muscle mechanics model, and a joint torque synthesis model;

[0016] The neural activation model calculates neural activation u using electromyographic signals and kinematic parameters.

[0017] The muscle activation model is based on the degree of muscle activation. The degree of muscle activation is calculated by considering the exponential relationship between the neural activation u and the activation level. ;

[0018] The muscle biomechanics model is based on the degree of muscle activation. Calculate active contraction force and passive elastic force ;

[0019] The joint torque synthesis model first combines the active contractile forces of each muscle. and passive elastic force The total muscle force is obtained by adding the total muscle force of each muscle, and then the joint torque is obtained by multiplying the total muscle force of each muscle by its lever arm and summing the results. .

[0020] Preferably, the muscle activation model, the degree of muscle activation The calculation formula is:

[0021] ;

[0022] in, Here, R is the shape parameter, and R is the maximum neural activation during the maximum random contraction. R is related to the maximum value of neural activation u. By activating the scaling factor Establish a linear relationship;

[0023] The muscle mechanics model, active contractile force and passive elastic force The calculation formulas are as follows:

[0024] ;

[0025] ;

[0026] in, This represents the standardized muscle fiber length. This represents the standardized muscle fiber contraction velocity. Maximum isolength muscle strength; , , These are the muscle fiber force-length relationship function, force-velocity relationship function, and passive elastic force-length relationship function, respectively.

[0027] , , ;

[0028] , ;

[0029] in, The optimal length coefficient; , These are the personalized identification parameters; This represents the actual muscle fiber length. To achieve the optimal muscle fiber length; Data obtained by fitting data using the biomechanical simulation software OpenSIM; This represents the actual contraction speed of the muscle fiber. This represents the maximum contraction velocity of the muscle fiber.

[0030] The personalized calibration parameters for individual adaptation include: shape parameters. Activation ratio factor Personalized identification parameters Personalized identification parameters Optimal length coefficient .

[0031] Preferably, step S4 includes:

[0032] Personalized biomechanical models calculate joint torque for each batch of samples. , joint torque Joint torque predicted by the joint torque prediction neural network model The error between them is used as the physical residual to calculate the loss function. ;

[0033] Calculate the joint torque predicted by the neural network model for each batch of samples. With actual joint torque The error between them is used as data error to calculate the loss function. ;

[0034] loss function and loss function A comprehensive loss is constructed by fusing samples using SoftMax. Based on the comprehensive loss obtained for each batch of samples, the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model are updated and adjusted.

[0035] Preferably, the formula for calculating the comprehensive loss is:

[0036] ;

[0037] in, , These are the attention scaling factors for the p-th batch.

[0038] Preferably, the joint torque prediction neural network model includes an input layer, a TCN model, and an output layer connected in sequence;

[0039] The method for constructing the input sample information of the joint torque prediction neural network model is as follows: the synchronized electromyographic signals and kinematic parameters are concatenated, and the time-series slicing is performed using the sliding window algorithm to obtain the input sample information;

[0040] The TCN model includes Each convolutional block consists of a convolutional layer, a ReLU activation layer, a batch normalization layer, and a Dropout layer. The sample features of the sample set are input through the input layer and then processed by each convolutional block. The convolutional layers of the convolutional blocks perform causal convolution and dilated convolution on the input to capture multi-dimensional features. The output features of the convolutional layers are processed by the ReLU activation layer to extract temporal features and by the batch normalization layer, before being input to the Dropout layer for Dropout processing. The time features after Dropout processing are projected back to the original feature space, transformed in dimension, and then residually connected to the initial sample features. Finally, the predicted joint torque is output through the output layer. .

[0041] Preferably, the electromyographic signals obtained in step S1 are preprocessed before being applied to the sample set; preprocessing refers to:

[0042] The electromyographic signal obtained in step S1 is filtered using a low-pass Butterworth filter to obtain the corresponding electromyographic signal envelope; the electromyographic signal envelope is then normalized to obtain the preprocessed electromyographic signal value.

[0043] A readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the joint torque prediction method incorporating biomechanical information.

[0044] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the joint torque prediction method that integrates biomechanical information.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. Balancing speed and biomechanical rationality: This invention uses a joint torque prediction neural network model for inference, which has a fast computing speed and no high running latency, meeting the needs of real-time applications; at the same time, it incorporates the physical constraints of a personalized biomechanical model, avoiding the problem of physical law violations caused by the "black box" characteristics of traditional models, and the prediction results are more biomechanically rational.

[0047] 2. Enhanced generalization ability and robustness: This invention reduces the reliance on a large amount of training data through the regularization effect of biomechanical knowledge. Even with limited training data, varying data distribution (such as different walking speeds), or cross-subject scenarios, it can still maintain good predictive performance, and its robustness and generalization ability are significantly improved.

[0048] 3. Higher prediction accuracy: Through dual optimization of root mean square loss and biomechanical loss, the trained joint torque prediction neural network model has a smaller root mean square loss in muscle strength and joint angle prediction, a higher Pearson correlation coefficient, and a better fit between the prediction results and the true values.

[0049] 4. Solving the model generalization problem under complex physical constraints: This invention achieves bidirectional backpropagation of gradients between the joint torque prediction neural network model and the personalized biomechanical model by constructing a parallel optimization framework driven by data error and physical residual. If the biomechanical model has deviations, it will be compensated in real time through gradient descent. If there is noise in the sample set, the physical consistency constraint can effectively suppress the overfitting of the joint torque prediction neural network model to random disturbances. This endogenous coupling optimization mechanism not only significantly improves the robustness of the system, but also solves the model generalization problem under complex physical constraints.

[0050] 5. High degree of personalization: The personalized biomechanical model is equipped with personalized calibration parameters specifically adapted to the individual, which can accurately match the individual differences between different subjects, thereby ensuring that the prediction results of the joint torque prediction neural network model embedded with this personalized information have a high degree of personalization. Attached Figure Description

[0051] Figure 1 This is a flowchart of the training process for the joint torque prediction neural network model of this invention;

[0052] Figure 2 This is a schematic diagram of the joint torque synthesis model of the present invention;

[0053] Figure 3 The results of joint torque prediction by the invented joint torque prediction neural network model and fully connected neural network in representative subjects;

[0054] Figure 4 This is a comparison chart of the evaluation metrics of the invented joint torque prediction neural network model and the fully connected neural network. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0056] Example 1

[0057] This embodiment presents a method for predicting joint torque by integrating biomechanical information. The method uses a joint torque prediction neural network model embedded with biomechanical information to predict joint torque. The joint torque prediction neural network model is embedded with a personalized biomechanical model. When training the joint torque prediction neural network model, the joint torque prediction neural network model and the personalized biomechanical model are optimized in parallel based on the collaborative driving of data error and physical residual, so as to realize the bidirectional backpropagation of gradient between the joint torque prediction neural network model and the personalized biomechanical model.

[0058] Training of a neural network model for predicting joint torque, such as Figure 1 As shown, it includes the following steps:

[0059] Step S1: Obtain electromyographic signals and kinematic parameters of lower limb joint movements from the sample population to construct a sample set; and obtain synchronous real joint torque. .

[0060] The method for obtaining electromyographic signals of lower limb joint movement is as follows:

[0061] Electromyographic signals from the gastrocnemius (GM), tibialis anterior (TA), and soleus (SL) muscles on the medial side of the lower leg were acquired using three wireless electromyography (EMG) sensors at a sampling frequency of 2000 Hz.

[0062] The method for obtaining kinematic parameters is as follows: an optical motion capture system consisting of sixteen reflective markers and six cameras is used to collect kinematic parameters at a sampling frequency of 100Hz;

[0063] The method for obtaining joint torque signals is as follows: a force-measuring treadmill is used to collect ground reaction force (GRF) data at a sampling frequency of 1000Hz, and then the joint torque signals are calculated.

[0064] The electromyographic signals obtained in step S1 are preprocessed before being applied to the sample set; preprocessing refers to:

[0065] The electromyographic signal obtained in step S1 is filtered using a fourth-order low-pass Butterworth filter with a frequency range of 10-400Hz to obtain the corresponding electromyographic signal envelope; the electromyographic signal envelope is then normalized to obtain the preprocessed electromyographic signal value.

[0066] The envelope amplitude of the electromyographic signals collected in each exercise mode was normalized to the maximum voluntary contraction value across all exercise modes, calculated using the following formula:

[0067] ;

[0068] in, This represents the raw value of the electromyographic signal envelope. This represents the maximum voluntary contraction value across all movement patterns. This represents the minimum value of the electromyographic signal envelope. The normalized electromyographic signal value is used to eliminate the influence of individual differences and signal fluctuations.

[0069] Step S2: Construct a personalized biomechanical model based on the constraint relationship between joint torque, electromyographic signals, and kinematic parameters, and separate the personalized calibration parameters adapted to the individual in the personalized biomechanical model.

[0070] Personalized biomechanical models include: neural activation models, muscle activation models, muscle mechanics models, and joint torque synthesis models.

[0071] The neural activation model calculates neural activation u using electromyographic signals and kinematic parameters.

[0072] Neural activation u originates from muscle signals, and its discrete analytical form can be represented as a critically damped linear second-order differential system:

[0073] ;

[0074] in, For neural activation; This is a muscle signal; For index number; This refers to the sampling frequency of the electromyography signal; , , These are second-order dynamic parameters; specifically, It is a neural activation level gain parameter; , These are parameters that regulate the increase and decrease of neural activation; The delay time is 40ms; the parameters satisfy the following constraints:

[0075] ,and ,and .

[0076] The muscle activation model is based on the degree of muscle activation. The degree of muscle activation is calculated by considering the exponential relationship between the neural activation u and the activation level. :

[0077] ;

[0078] in, Here, R is the shape parameter, and R is the maximum neural activation during the maximum random contraction. R is related to the maximum value of neural activation u. By activating the scaling factor Establish a linear relationship. .

[0079] The muscle biomechanics model is based on the degree of muscle activation. Calculate active contraction force and passive elastic force :

[0080] ;

[0081] ;

[0082] in, This represents the standardized muscle fiber length. This represents the standardized muscle fiber contraction velocity. Maximum isolength muscle strength; , , These are the muscle fiber force-length relationship function, force-velocity relationship function, and passive elastic force-length relationship function, respectively.

[0083] , , ;

[0084] , ;

[0085] in, The optimal length coefficient; , These are the personalized identification parameters; This represents the actual muscle fiber length. To achieve the optimal muscle fiber length; Data obtained by fitting data using the biomechanical simulation software OpenSIM; This represents the actual contraction speed of the muscle fiber. This represents the maximum contraction velocity of the muscle fiber.

[0086] ;

[0087] ;

[0088] ;

[0089] Where ζ=0.45 is the muscle force-length shape factor; =1.5, which is the maximum normalized muscle elongation; =0.25, which is the muscle force-velocity curve factor.

[0090] The joint torque synthesis model, such as Figure 2 As shown, firstly, the active contraction force of each muscle... and passive elastic force The total muscle force is obtained by adding the components together; considering the influence of the muscle penum angle, the joint torque is obtained by multiplying the total muscle force of each muscle by its lever arm and then summing the results. :

[0091] ;

[0092] in, , For the i-th muscle, there are the active contractile force and passive elastic force. Let be the lever arm of the i-th muscle; n is the number of muscles involved in the joint movement. The feathered angle at rest;

[0093] ;

[0094] lever arm From the sum of muscle and tendon length For joint angle Partial derivatives:

[0095] ;

[0096] ;

[0097] in, The length of the tendon changes very little during muscle activity, so it is set to a constant.

[0098] Personalized biomechanical models employ a large number of individualized calibration parameters, including shape parameters. Activation ratio factor Personalized identification parameters , Optimal length coefficient Therefore, during the embedding of the biomechanical model, individual parameter identification for each participant allows the personalized biomechanical model to adjust its parameters and functions accordingly based on the physiological characteristics of different individuals. These calibration routines incorporate the solved parameters as optimization variables into the joint torque prediction neural network model.

[0099] Step S3: The joint torque prediction neural network model extracts temporal and spatial dimension information from the input sample information of the sample set, and performs feature fusion on the temporal dimension information, spatial dimension information, and initial input sample information, and then outputs the predicted joint torque. .

[0100] The joint torque prediction neural network model consists of an input layer, a TCN model, and an output layer connected in sequence.

[0101] The method for constructing the input sample information of the joint torque prediction neural network model is as follows: the synchronized electromyographic signals and kinematic parameters are concatenated, and the time-series slicing is performed using the sliding window algorithm to obtain the input sample information;

[0102] The TCN model (Temporal Convolutional Network model) includes =4 convolutional blocks; each convolutional block consists of a convolutional layer, a ReLU activation layer, a batch normalization layer, and a Dropout layer; the sample features of the sample set are input through the input layer and then processed by each convolutional block; the convolutional layers of the convolutional blocks perform causal convolution and dilated convolution on the input to achieve multi-dimensional feature capture; the output features of the convolutional layers are processed by the ReLU activation layer to extract temporal features and the batch normalization layer to normalize them before being input into the Dropout layer for Dropout processing; the time features after Dropout processing are projected back to the original feature space, the dimensions are transformed, and a residual connection is formed with the initial sample features; finally, the predicted joint torque is output through the output layer. .

[0103] Step S4: Input the sample information into the personalized biomechanical model for calculation to obtain the joint torque. ; to joint torque With joint torque Perform error calculation; calculate the joint torque. With actual joint torque Error calculation is performed; the results of the two error calculations are then self-attentionally fused into a comprehensive loss; the comprehensive loss from batch data is used to continuously update the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model, in order to achieve convergence of the comprehensive error. Specifically, the steps include:

[0104] The electromyographic signals and kinematic parameters of the sample information are input into a personalized biomechanical model to calculate the joint torque. ;

[0105] Calculate the joint torques of the personalized biomechanical models for each batch of samples (sample size B). Joint torque predicted by the joint torque prediction neural network model The root mean square loss between them is used as the loss function. :

[0106] ;

[0107] Calculate the joint torque predicted by the neural network model for each batch of samples. With actual joint torque The error between them is used as the loss function :

[0108] ;

[0109] loss function and loss function A comprehensive loss is constructed by fusing data using SoftMax; the formula for calculating the comprehensive loss is as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] in, , These are the attention scaling factors for the p-th batch; , Loss functions for batches p and q, respectively. ; , Loss functions for batches p and q, respectively. n is the batch quantity;

[0114] The backpropagation algorithm is used based on the comprehensive loss obtained from each batch of samples, and the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model are updated and adjusted according to the gradient descent algorithm.

[0115] This invention has been fully validated for its feasibility and effectiveness through real-human experiments. The experimental design employed a subject-object dataset architecture: the subject dataset was used to evaluate the inversion and identification capabilities of the joint torque prediction neural network model, while the object dataset was used to verify the general applicability of this invention in different scenarios. The experiment included six healthy adult subjects (4 males and 2 females, age 24.6 ± 0.4 years, height 173.9 ± 1.5 cm, weight 68.7 ± 2.6 kg). Signals from two exercise modes—weighted walking and brisk walking—were simultaneously collected using an optical motion capture system, a force-measuring treadmill, and a wireless surface electromyography (EMG) sensor. The sampling frequencies were 100 Hz, 1000 Hz, and 2000 Hz, respectively. The predicted joint torque was the ankle joint torque.

[0116] In the inverse problem modeling and verification phase of the experiment, the weighted walking data was used as the main training set, and the parameters to be optimized and the adaptive scheduling mechanism were independently set for each muscle of each subject (S1-S6). During the training process, the overall loss showed a stable convergence trend.

[0117] Table 1. Parameter identification results of the three muscles in the subjects.

[0118] Table 1 presents the parameter identification results for three muscles in representative subjects. As shown in Table 1, there are significant physiological differences in individualized parameters among different muscles. Specifically, the shape parameter A of the gastrocnemius muscle on the medial side of the calf is -2.83, and the activation ratio factor k of the soleus muscle is... r The score reached 2.57, and the optimal length coefficient λ for the tibialis anterior muscle was 0.72. These identification results are consistent with prior anatomical knowledge. In the forward prediction validation stage, the individualized parameters obtained from the inverse problem were fixed to construct a forward prediction neural network model, which was then cross-compared with a fully connected neural network FCNN to obtain... Figure 3 and Figure 4 .

[0119] Figure 3 The invention visually demonstrates the predictive performance of the PINN joint torque prediction neural network model and the fully connected FCNN neural network within a gait cycle: during the abrupt change in ground reaction force (early support phase), the PINN joint torque prediction neural network model can maintain a more stable neuromuscular response, and the peak alignment error is significantly reduced; during the swing phase, the torque curve of the PINN joint torque prediction neural network model has a higher degree of agreement with the true value, and the residual attention mechanism effectively suppresses the jitter caused by high-frequency noise in the electromyographic signal.

[0120] Figure 4 A bar chart comparing the average performance of the proposed method across subjects, including error bars, systematically compares the performance of the proposed method on five core evaluation metrics, covering maximum error (MER), shape similarity (SS), maximum single-step error (PEA), coefficient of determination (R²), and root mean square error (RMSE). Shape similarity and coefficient of determination are positive indicators (higher values ​​indicate better performance), while the other three are negative indicators (lower values ​​indicate smaller errors and better performance). Error bars reflect the degree of variability of each indicator among subjects. The results show that the proposed joint torque prediction neural network model PINN comprehensively outperforms the fully connected neural network FCNN in four metrics across all subjects: root mean square error, shape similarity, peak alignment error, and maximum error range. Specifically, the mean shape similarity is improved by 12.3%, and the peak alignment error is reduced by 35.7%, demonstrating the convergence ability and cross-individual stability of the proposed method on sparse datasets.

[0121] In particular, cross-motion mode transfer experiments show that when the personalized biomechanical model obtained from weight-bearing walking training is directly applied to fast walking prediction, the PINN joint torque prediction neural network model still maintains high accuracy, while the performance of the fully connected neural network FCNN significantly decreases. This verifies the crucial role of physical constraints in improving model robustness. The experimental results fully demonstrate that this invention, by integrating biomechanical prior knowledge and attention mechanisms, achieves high-precision, low-data-dependency, and strong generalization ability for joint torque prediction in real human motion scenarios, meeting the technical requirements for real-time control of exoskeleton robots.

[0122] Example 2

[0123] This embodiment provides a readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the joint torque prediction method incorporating biomechanical information as described in Embodiment 1.

[0124] Example 3

[0125] This embodiment discloses a computer device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the joint torque prediction method that integrates biomechanical information as described in Embodiment 1.

[0126] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for predicting joint torque by integrating biomechanical information, characterized in that: A joint torque prediction neural network model embedded with biomechanical information is used to predict joint torque; a personalized biomechanical model is embedded in the joint torque prediction neural network model; when training the joint torque prediction neural network model, the joint torque prediction neural network model and the personalized biomechanical model are optimized in parallel based on data error and physical residual, so as to realize the bidirectional backpropagation of gradient between the joint torque prediction neural network model and the personalized biomechanical model. The training of the joint torque prediction neural network model includes the following steps: Step S1: Obtain electromyographic signals and kinematic parameters of lower limb joint movements from the sample population to construct a sample set; and obtain synchronous real joint torque. ; Step S2: Construct a personalized biomechanical model based on the constraint relationship between joint torque, electromyographic signals, and kinematic parameters, and separate the personalized calibration parameters adapted to the individual in the personalized biomechanical model. Step S3: The joint torque prediction neural network model extracts temporal and spatial dimension information from the input sample information of the sample set, and performs feature fusion on the temporal and spatial dimension information and the initial input sample information, and then outputs the predicted joint torque. ; Step S4: Input the sample information into the personalized biomechanical model for calculation to obtain the joint torque. ; to joint torque With joint torque Perform error calculation; calculate the joint torque. With actual joint torque Error calculation is performed; the two error calculation results are self-attentionally fused into a comprehensive loss; the comprehensive loss of batch data is used to continuously update the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model, so as to achieve comprehensive error convergence; The personalized biomechanical models include: neural activation model, muscle activation model, muscle mechanics model, and joint torque synthesis model; The neural activation model calculates neural activation u using electromyographic signals and kinematic parameters. The muscle activation model is based on the degree of muscle activation. The degree of muscle activation is calculated by considering the exponential relationship between the neural activation u and the activation level. ; The muscle biomechanics model is based on the degree of muscle activation. Calculate active contraction force and passive elastic force ; The joint torque synthesis model first combines the active contractile forces of each muscle. and passive elastic force The total muscle force is obtained by adding the total muscle force of each muscle, and then the joint torque is obtained by multiplying the total muscle force of each muscle by its lever arm and summing the results. .

2. The joint torque prediction method integrating biomechanical information according to claim 1, characterized in that: The muscle activation model, the degree of muscle activation The calculation formula is: ； in, Here, R is the shape parameter, and R is the maximum neural activation during the maximum random contraction. R is related to the maximum value of neural activation u. By activating the scaling factor Establish a linear relationship; The muscle mechanics model, active contractile force and passive elastic force The calculation formulas are as follows: ； ； in, This represents the standardized muscle fiber length. This represents the standardized muscle fiber contraction velocity. Maximum isolength muscle strength; , , These are the muscle fiber force-length relationship function, force-velocity relationship function, and passive elastic force-length relationship function, respectively. ， ， ； ， ； in, The optimal length coefficient; , These are the personalized identification parameters; This represents the actual muscle fiber length. To achieve the optimal muscle fiber length; Data obtained by fitting data using the biomechanical simulation software OpenSIM; This represents the actual contraction speed of the muscle fiber. This represents the maximum contraction velocity of the muscle fiber. The personalized calibration parameters for individual adaptation include: shape parameters. Activation ratio factor Personalized identification parameters Personalized identification parameters Optimal length coefficient .

3. The joint torque prediction method integrating biomechanical information according to claim 1, characterized in that: Step S4 includes: Personalized biomechanical models calculate joint torque for each batch of samples. , joint torque Joint torque predicted by the joint torque prediction neural network model The error between them is used as the physical residual to calculate the loss function. ; Calculate the joint torque predicted by the neural network model for each batch of samples. With actual joint torque The error between them is used as data error to calculate the loss function. ; loss function and loss function A comprehensive loss is constructed by fusing samples using SoftMax. Based on the comprehensive loss obtained for each batch of samples, the personalized calibration parameters of the personalized biomechanical model and the parameters of the joint torque prediction neural network model are updated and adjusted.

4. The joint torque prediction method integrating biomechanical information according to claim 3, characterized in that: The formula for calculating the comprehensive loss is as follows: ； in, , These are the attention scaling factors for the p-th batch.

5. The joint torque prediction method integrating biomechanical information according to claim 1, characterized in that: The joint torque prediction neural network model includes an input layer, a TCN model, and an output layer connected in sequence. The method for constructing the input sample information of the joint torque prediction neural network model is as follows: the synchronized electromyographic signals and kinematic parameters are concatenated, and the time-series slicing is performed using the sliding window algorithm to obtain the input sample information; The TCN model includes Each convolutional block consists of a convolutional layer, a ReLU activation layer, a batch normalization layer, and a Dropout layer. The sample features of the sample set are input through the input layer and then processed by each convolutional block. The convolutional layers of the convolutional blocks perform causal convolution and dilated convolution on the input to capture multi-dimensional features. The output features of the convolutional layers are processed by the ReLU activation layer to extract temporal features and by the batch normalization layer, before being input to the Dropout layer for Dropout processing. The time features after Dropout processing are projected back to the original feature space, transformed in dimension, and then residually connected with the initial sample features. Finally, the predicted joint torque is output through the output layer. .

6. The joint torque prediction method integrating biomechanical information according to claim 1, characterized in that: The electromyographic signals obtained in step S1 are preprocessed before being applied to the sample set; preprocessing refers to: The electromyographic signal obtained in step S1 is filtered using a low-pass Butterworth filter to obtain the corresponding electromyographic signal envelope; the electromyographic signal envelope is then normalized to obtain the preprocessed electromyographic signal value.

7. A readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, causes the processor to perform the joint torque prediction method that integrates biomechanical information as described in any one of claims 1-6.

8. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the joint torque prediction method that integrates biomechanical information as described in any one of claims 1-6.

Images