A human multi-joint torque prediction method and system

Abstract

This invention relates to a method and system for predicting torque at multiple joints in the human body. The method includes the following steps: Step 1: Acquire electromyographic signals and joint kinematic information, calculate the data of the target joint muscle group, and form a time-aligned group training dataset; Step 2: Sequentially construct a convolutional feature extraction module, an attention pooling module, and an MLP parameter output head module to synthesize a convolutional-attention neural network; Step 3: Construct a composite loss function to obtain the initial parameter vector at the group level; Step 4: For the individual to be deployed, collect its short-time continuous data segments as calibration segments, obtain the final parameter vector x of the individual, and save it; Step 5: Construct a multi-dimensional B-spline function mapping from the kinematic information of the target joint to the muscle and tendon unit length, torque arm, and feather angle of the target joint muscle group; Step 6: Output the real-time torque prediction value of the target joint. This method can obtain high-precision torque prediction results.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for predicting torques at multiple joints in the human body. Background Technology

[0002] Joint torque is a core dynamic quantity characterizing the load level of the musculoskeletal system and a key indicator for measuring motor ability and rehabilitation effectiveness. In clinical rehabilitation and sports medicine, joint torque is used to quantify functional impairment and changes in treatment efficacy, assess motor symmetry and risky postures, and assist in developing individualized training and regression standards. In human factors engineering, it is used to estimate joint forces and cumulative loads, identify high-risk work postures, and guide workstation design and workflow optimization. In sports science and training, it is used to analyze force patterns, differences in technical movements, and the effects of fatigue, assisting in developing training prescriptions and evaluating athletic performance. In prostheses, exoskeletons, and other human-machine collaborative devices, joint torque is both a direct control target for the amplitude and timing of assistance and a benchmark indicator for evaluating safety constraints, energy efficiency, and interactive comfort. In biomechanical modeling and digital human simulation, it is also a key truth value and consistency constraint for model calibration, parameter identification, and multi-source sensor inference. Human joint torque is a universal indicator and key reference across clinical, engineering, and sports applications.

[0003] Acquiring high-precision joint torque typically relies on force measurement platforms, optical motion capture, and multi-sensor synchronous acquisition systems under laboratory conditions, and then calculates the true torque value through inverse dynamics. While this process offers high accuracy, the equipment is expensive, deployment is complex, and it has strict requirements for the scene and space, making it difficult to meet the application needs of daily rehabilitation, outdoor training, and real-time control of portable devices. To improve the usability and real-time performance of torque acquisition, two alternative solutions have emerged in research and engineering practice: one is a data-driven end-to-end regression method, which uses electromyography (EMG) and kinematic signals as inputs and employs neural networks to directly regress and output joint torque. This method is simple to implement and has a high personalized baseline after adaptation with a large amount of individual data, but it has poor generalization performance and limited physical consistency and interpretability. The other is an EMG-driven neuromuscular skeletal model and its parameter identification method, which uses EMG and joint angles or angular velocities as inputs, identifies physiological parameters based on simple neural networks or optimization algorithms, and finally inputs them into a physical model to calculate the joint torque. This approach has strong physical interpretability and can achieve relatively acceptable performance with limited data, but it is difficult to improve the personalized baseline. Neither of the above two schemes can achieve stable and high-precision torque prediction under conditions of a small number of samples.

[0004] To address the above issues, a method and system for predicting human multi-joint torques are proposed. This method solves the problems of poor generalization performance and limited physical interpretability caused by direct regression of joint torques by deep learning networks. The method uses a neural network with an attention mechanism to learn physiological parameter vectors based on muscle geometry / state sequences and calculates joint torques by combining a forward physics model. It can achieve stable and high-precision prediction of human multi-joint torques under small batch sample data conditions. Summary of the Invention

[0005] To solve the above-mentioned technical problems, this invention proposes a method and system for predicting torque at multiple joints in the human body.

[0006] This invention relates to a method for predicting torque at multiple joints in the human body, comprising the following steps:

[0007] Step 1: Collect electromyographic signals through an electromyography sensor and collect joint kinematic information through an inertial sensor. Input the original biokinematic and mechanical signals obtained from the test human uniform walking gait experiment into the musculoskeletal model, calculate and obtain the time-series data of muscle and tendon unit length, torque arm and pendentive angle of the target joint muscle group, and calculate the true value of joint torque through the inverse dynamics model to form a time-aligned group training dataset.

[0008] Step 2: Construct the convolutional feature extraction module, the attention pooling module, and the MLP parameter output head module in sequence, and combine the above three modules into a convolutional-attention neural network;

[0009] Step 3: Construct a composite loss function to train the convolutional-attention neural network built in Step 2, and obtain the initial parameter vector at the group level;

[0010] Step 4: For the individual to be deployed, collect its short-term continuous data segment as a calibration segment; during calibration, use the initial parameter vector at the population level as the optimization variable, and obtain the final parameter vector x of the individual by minimizing the loss function on the calibration segment and save it.

[0011] Step 5: Construct a multidimensional B-spline function mapping from the kinematic information of the target joint to the muscle and tendon unit lengths, torque arms, and pendent angles of the muscle group of the target joint;

[0012] Step 6: Collect kinematic information of the tested individual using an inertial sensor, and obtain the muscle and tendon unit length, torque arm, and feather angle of the target joint muscle group through multidimensional B-spline function mapping; input the mapped data information, parameter vector x, and real-time acquired electromyographic signals into the physical forward calculation module, and output the real-time torque prediction value of the target joint.

[0013] Furthermore, in step 2, the convolutional feature extraction module is constructed by performing one-dimensional convolutional feature extraction on the electromyographic signals, joint kinematic information, muscle and tendon unit lengths of the target joint muscle group, and torque arm temporal data collected in step 1, in order to extract dynamic features at different time scales.

[0014] Furthermore, in step 2, the attention pooling module is constructed by adaptively weighting and pooling the one-dimensional convolutional features extracted by the convolutional feature extraction module in the time dimension to obtain attention weights and pooling features.

[0015] Furthermore, in step 2, the parameter output head module maps the pooling features to the parameter logits vector z, and then maps it to a pre-defined upper and lower bound value range through a Sigmoid-type boundary mapping function, thus converting the parameter logits vector z into an initial parameter vector representing the physiological properties of the target joint muscle group.

[0016] Furthermore, in step 3, the composite loss function is constructed by weighting and combining the above six terms: torque error term, correlation and slope consistency term, physiological regularization term, dynamic smoothing and passive term penalty, attention stability regularization, and global gain regularization.

[0017] Further, in step 3, the torque error term is constructed using the weighted root mean square error (RMSE). Time weights w(t) are constructed based on the torque magnitude and gradient. The weight of the resting / low torque segment before joint exertion is set to 0.3, and the weight of the high torque / drastic torque change segment after the joint torque reaches the gradient threshold is set to 1.5. The correlation and slope consistency terms are constructed by introducing correlation loss and slope loss to enhance waveform and scale consistency. The physiological regularization term is constructed by penalizing deviations from the physiologically reasonable range for the five physiological parameters of each muscle. The dynamic smoothing and passive term penalty is constructed by applying smoothing constraints to the change of muscle force over time and penalizing the passive muscle force component to avoid non-physiological bursts. The attention stability regularization term introduces entropy regularization and time smoothing regularization into the attention weights to avoid attention collapse or violent jitter under conditions of few samples. The global gain regularization term penalizes the degree to which the global gain g deviates from 1.

[0018] Furthermore, in step 4, the parameter vector x includes five physiological parameters for each muscle and a global gain g.

[0019] Furthermore, in step 4, the five physiological parameters for each muscle are the maximum isometric force Fmax, tendon relaxation length LtSlack, optimal muscle fiber length LmOpt, scaling factor s, and the nonlinear shape parameter A for EMG conversion into muscle activation.

[0020] Further, in step 4, the parameter vector x is used by substituting the five physiological parameters of each muscle in the parameter vector x into the physical forward calculation module, and combining it with the electromyographic signals collected in step 1. The physical forward calculation module is then used to calculate the temporal muscle force of the muscle and tendon unit. The joint torque prediction value is obtained by superimposing the temporal muscle force with the torque arm temporal data obtained in step 1. The physical forward calculation module consists of a series of differentiable continuous equations, based on the Hill-type muscle mechanics relationship, and calculates the force exerted by the corresponding muscle and tendon unit by inputting electromyographic signals.

[0021] The present invention also relates to a system for predicting human multi-joint torques, the system comprising a computer module for running the human multi-joint torque prediction method.

[0022] Beneficial effects

[0023] This invention improves the stability of physiological parameter identification and the physical consistency of torque prediction under small sample conditions by employing a composite loss constraint and parameter-level calibration strategy. The optimized parameter vector calibration strategy enables rapid individualized adaptation under real, continuous, short-time data conditions, resulting in high-precision torque prediction results. It is applicable to multi-joint torque prediction and can be extended to various scenarios such as rehabilitation assessment, exoskeleton / prosthetic control, and sports training. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall technical solution of a method for predicting torque at multiple joints in the human body according to the present invention.

[0025] Figure 2 This is a block diagram illustrating the working principle of the offline training phase of the convolutional-attention neural network of this invention.

[0026] Figure 3 This is a bar chart comparing the root mean square error test results of the predicted torque and the actual torque of the present invention.

[0027] Figure 4 A bar chart comparing the Pearson correlation coefficient test results of the predicted torque and the actual torque of this invention;

[0028] Figure 5 This is a comparison chart of the average predicted torque curve and the true average torque curve of the present invention;

[0029] Figure 6 This is a comparison chart of the torque prediction results of the L5-S1 intervertebral disc joint according to the present invention. Detailed Implementation

[0030] The following combination Figures 1 to 6 This implementation method will be described in detail.

[0031] The purpose of this invention is to provide a method for predicting torque at multiple joints in the human body. This method uses muscle geometry / muscle-bone state parameters obtained by processing raw biological kinematic and mechanical signals through a musculoskeletal model as the core input. A convolutional-attention neural network is constructed, and the output is a population parameter vector constrained by physiological boundaries. This is combined with physical forward computation and a composite loss function constraint. In the new individual experimental calibration stage, a calibration strategy of "optimizing the parameter vector without changing the network weights" is adopted, ultimately obtaining individualized physiological parameter results. Combined with physical model forward computation, this method achieves high-precision torque prediction results for each individual in only 4-5 gait cycles. It is applicable to scenarios such as musculoskeletal rehabilitation assessment, exoskeleton / prosthetic control, and sports training.

[0032] This invention comprises three stages in its implementation: an offline training stage, an individual calibration stage, and an online inference stage. Steps 1-3 belong to the offline training stage, steps 4-5 belong to the individual calibration stage, and step 6 belongs to the online inference stage. The overall technical solution flowchart of this invention is as follows: Figure 1 As shown, the specific implementation method is as follows:

[0033] Step 1: Collect electromyographic signals using an electromyography sensor and joint kinematic information using an inertial sensor. Input the raw bio-kinematic and mechanical signals obtained from the uniform walking gait experiment of the tested human body into the musculoskeletal model. Calculate the temporal data of the muscle and tendon unit length, torque arm, and pendentive angle of the target joint muscle group. Calculate the true value of the joint torque using an inverse dynamics model to form a time-aligned group training dataset.

[0034] Step 2: Construct the convolutional feature extraction module, the attention pooling module, and the MLP parameter output head module in sequence, and combine these three modules into a convolutional-attention neural network, as follows: Figure 2 As shown.

[0035] The convolutional feature extraction module is constructed by performing one-dimensional convolutional feature extraction on the electromyographic signals, joint kinematic information, muscle and tendon unit lengths of the target joint muscle group, and torque arm temporal data collected in step 1, in order to extract dynamic features at different time scales.

[0036] The attention pooling module is constructed by adaptively weighting and pooling the one-dimensional convolutional features extracted by the convolutional feature extraction module in the time dimension to obtain attention weights and pooling features.

[0037] The parameter output head module is constructed by mapping pooling features into a parameter logits vector z, and then mapping it to a pre-defined upper and lower bound value range through a Sigmoid-type boundary mapping function, thus converting the parameter logits vector z into an initial parameter vector representing the physiological properties of the target joint muscle group.

[0038] Step 3: Construct a composite loss function to train the convolutional-attention neural network built in Step 2, and obtain the initial parameter vector at the group level.

[0039] The composite loss function is a composite loss function constructed by weighting and combining the above six terms: torque error term, correlation and slope consistency term, physiological regularization term, dynamic smoothing and passive term penalty, attention stability regularization, and global gain regularization.

[0040] The torque error term is constructed using the weighted root mean square error (RMSE). Time weights w(t) are constructed based on torque magnitude and gradient, with the weight of the resting / low torque segment before joint exertion set to 0.3, and the weight of the high torque / drastic torque change segment after the joint torque reaches the gradient threshold set to 1.5. The correlation and slope consistency terms are constructed by introducing correlation loss and slope loss to enhance waveform and scale consistency. The physiological regularization term is constructed by penalizing deviations from the physiologically reasonable range for five physiological parameters of each muscle. The dynamic smoothing and passive term penalty is constructed by applying smoothing constraints to the time-varying muscle force and penalizing the passive muscle force component to avoid non-physiological bursts. The attention stability regularization term introduces entropy regularization and time smoothing regularization into the attention weights to avoid attention collapse or violent jitter under conditions of few samples. The global gain regularization term penalizes the degree to which the global gain g deviates from 1.

[0041] Step 4: For the individual to be deployed, collect its short-term continuous data segment as a calibration segment; during calibration, use the initial parameter vector at the group level as the optimization variable, obtain the final parameter vector x of the individual by minimizing the loss function on the calibration segment and save it for reuse in the subsequent online inference stage.

[0042] The parameter vector x includes five physiological parameters for each muscle and a global gain g. The five physiological parameters for each muscle are: maximum isometric force Fmax, tendon relaxation length LtSlack, optimal muscle fiber length LmOpt, scaling factor s, and nonlinear shape parameter A for EMG conversion to muscle activation.

[0043] The parameter vector x is used by substituting the five physiological parameters of each muscle in the parameter vector x into the physical forward calculation module. Combined with the electromyographic signals collected in step 1, the physical forward calculation module is used to calculate the temporal muscle force of the muscle-tendon unit. The predicted joint torque value is obtained by superimposing the temporal muscle force with the torque arm temporal data obtained in step 1. The physical forward calculation module consists of a series of differentiable continuous equations, based on Hill-type muscle mechanics, and calculates the force exerted by the corresponding muscle-tendon unit by inputting electromyographic signals.

[0044] Step 5: Construct a multidimensional B-spline function mapping from the kinematic information of the target joint to the muscle and tendon unit lengths, torque arms, and feather angles of the muscle group of the target joint.

[0045] Step 6: Collect kinematic information of the tested individual using an inertial sensor, and obtain the muscle and tendon unit length, torque arm, and feather angle of the target joint muscle group through multidimensional B-spline function mapping; input the mapped data information, parameter vector x, and real-time acquired electromyographic signals into the physical forward calculation module, and output the real-time torque prediction value of the target joint. Figure 3 , 4 A bar chart was created to compare the root mean square error and Pearson correlation coefficient of the predicted torque and the actual torque of the hip, knee and ankle joints of the lower limbs obtained from testing on a public dataset. Compared with the comparison method, the root mean square error of this method is significantly reduced, and the Pearson correlation coefficient is significantly improved. Figure 5 The figure shows a comparison between the average predicted torque curve and the true average curve. As can be seen from the figure, the predicted torque curve of this method is closer to the true curve. Figure 6 The image shows a comparison of the torque prediction results of the L5-S1 intervertebral disc joint tested using self-collected data. It can be seen that this method also has high prediction accuracy in the L5-S1 intervertebral disc joint.

[0046] The present invention also relates to a system for predicting human multi-joint torques, the system comprising a computer module for running the human multi-joint torque prediction method.

[0047] Example

[0048] Step 1, Data Preparation: For each training individual, muscle length, torque arm, and feather angle are calculated based on the musculoskeletal model, and the true value of joint torque is obtained through inverse dynamics. The torque curve is processed by zero-phase low-pass filtering, with the cutoff frequency set to 6 Hz.

[0049] Network structure: The network input is a sequence of muscle geometry / musculoskeletal state features arranged in time, specifically n-dimensional EMG, 1-dimensional joint angle, n-dimensional muscle length, and n-dimensional lever arm.

[0050] Step 2: The convolutional feature extraction module performs one-dimensional convolution on the input to extract short-term and medium-term dynamic features. The module structure consists of two layers of one-dimensional convolution, both operating on the time axis. The first convolution layer has 3n+1 input channels and 32 output channels; the kernel length is 21, and ReLU is applied after convolution. The second convolution layer has 32 input channels and 64 output channels; the kernel length is also 21, and ReLU is applied after convolution.

[0051] After convolution, the input to the attention pooling module generates attention weights, adaptively converging temporal information into a global representation. The attention pooling layer takes a 64-dimensional time series as input, calculates a scalar score for the 64-dimensional features at each time point, and then normalizes all the scores to transform them into a set of weights that sum to 1. This way, moments with significant force exertion and drastic torque changes automatically receive greater weights, while resting periods receive smaller weights, thus focusing the model's representation on the "critical force exertion phase." A weighted average of these weights across the entire time series yields a vector representing the "information of the entire motion segment." Regularization constraints are applied to the attention weights in the loss function.

[0052] The parameter vector is then output through the parameter output header. The parameter prediction header structure consists of two fully connected layers. The first layer takes a 64-dimensional input and outputs a 128-dimensional output; the second layer takes a 128-dimensional input and outputs a 5n+1-dimensional output, resulting in 5n unconstrained parameter vectors and a global gain. Through a sigmoid-type mapping and pre-defined upper and lower bounds, the parameter vectors are transformed into physically meaningful physiological parameters. The global gain is then limited to a small range close to 1, thus ensuring the rationality of the physiological parameters and the stability of subsequent calculations.

[0053] Step 3, Composite Loss: During training, calculate the weighted RMSE constraint torque. Construct time weights w(t) based on the torque magnitude and gradient. The weights decrease to 0.3 for resting / low torque segments and increase to 1.5 for segments with drastic torque changes (high gradient). This makes the network focus more on fitting the force and rise-fall phases, rather than resting segments. Trend Consistency Loss: Correlation loss, slope loss. Parameter and Physiological Regularization: Activate the scaling factor s prior (to keep it fluctuating around 1, avoiding unconstrained amplification or reduction of muscle force); Dynamic Suppression (suppresses unreasonable large dynamic dependence). Attention Regularization: Entropy Regularization (focuses attention on key time points), Temporal Smoothing (maintains smooth changes in attention over time, suppressing high-frequency jitter). Global Gain Regularization (a slight global calibration).

[0054] Training was performed using a first-order Adam optimizer, jointly training the convolutional layers, attention pooling layers, and parameter regression head in the parametric network. The initial learning rate was set to 1×10^(-3), and a total of 300 training epochs were conducted.

[0055] Step 4: Calibration segment selection: Collect a small batch of gait cycle data for the individual to be tested as the calibration segment.

[0056] Calibration variable constraints: The parameter vector of this individual is optimized using the population parameter vector obtained from offline training as the initial value, without changing the network structure and loss function. The first-order Adam optimizer is also used, with a learning rate of 1×10^(-2), and a total of 800 updates are performed.

[0057] To compensate for the problem of asynchronous data acquisition (potentially resulting in millisecond-level residual misalignment), a temporal lag alignment is introduced during the new individual calibration phase. This prevents the model from forcibly fitting the temporal misalignment by modifying physiological parameters and global gain, which would lead to distortion of the final optimized parameters. First, the predicted moment curve for the new individual calibration phase is calculated using the population baseline parameters. Then, the shift of the sampling point that maximizes the correlation between the predicted and actual moment curves is found. After alignment, the physiological muscle parameters are optimized.

[0058] Once calibration is complete, the output will be saved for subsequent inference.

[0059] Physics-based forward computation: The optimized physiological parameter vector representing the physiological attributes of the target joint muscle group is input into the physics-based forward model. First, the EMG signal is bandpass filtered with cutoff frequencies of 20Hz and 500Hz, and a 50Hz notch filter is used to remove power frequency interference. A 6th-order Butterworth filter with zero phase delay is selected. Then, the absolute value of the filtered sEMG signal is taken to highlight the pulse firing characteristics of the sEMG signal. The processed sEMG signal is then filtered by a low-pass filter with a cutoff frequency of 2Hz to extract the envelope.

[0060] The neural activation is extracted from the processed EMG using the following formula. A first-order recursive filter is applied and a fixed time delay is superimposed, which roughly corresponds to a simplified electrophysiological process of "filtering + dynamic characteristics + conduction delay".

[0061]

[0062] In the formula, — The neural activation function at time t; t — the current time step; —Processed EMG; , , —Neural activation coefficients, with values ​​of 0.9486, 0.052, and 0.00627, respectively; —Time delay, with a value of 36ms; —Sampling time interval.

[0063] Then, the muscle activation function is calculated using the neural activation function:

[0064]

[0065] In the formula, —Muscle activation function; — Nonlinear shape parameters.

[0066] Muscle force is calculated from muscle activation using a simplified Hill-type muscle model, which incorporates physiological parameters optimized by network learning.

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] In the formula, —Muscle fiber length; — Optimal muscle fiber length; — Muscle-tendon unit length; — Tendon relaxation length; —Normalized muscle fiber length; — Scaling factor; —Maximum isolongitudinal muscle strength; —The main driving force; —Passive force; —Pinnate horns; —Single muscle strength.

[0073] Finally, the forces of each muscle are multiplied by their torque arms and summed, then multiplied by the global gain to obtain the predicted joint torque. The predicted joint torque is then compared with the true joint torque obtained in step 1 to calculate the composite loss function value.

[0074] Step 5: Construct a multidimensional B-spline function mapping from the kinematic information of the target joint to the muscle and tendon unit lengths, torque arms, and feather angles of the muscle group of the target joint.

[0075] Step 6: In the online phase, kinematic information is obtained through IMU or joint angle data; the length, torque arm, and feather angle of the relevant muscle tendon unit (MTU) are estimated through a pre-established mapping model, and these are input into the physical forward module along with the saved individual parameter vector to calculate the joint torque and output it in real time.

[0076] This invention is applicable to all major joints of the human body where relevant muscle geometry / state information can be obtained, with the output unit being Nm.

[0077] The above description of the present invention is only a preferred embodiment of the present invention and is not intended to limit the implementation of the present invention. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection claimed in the claims.

Claims

1. A method for predicting torque at multiple joints in the human body, characterized in that, Includes the following steps: Step 1: Collect electromyographic signals through an electromyography sensor and collect joint kinematic information through an inertial sensor. Input the original biokinematic and mechanical signals obtained from the test human uniform walking gait experiment into the musculoskeletal model, calculate and obtain the time-series data of muscle and tendon unit length, torque arm and pendentive angle of the target joint muscle group, and calculate the true value of joint torque through the inverse dynamics model to form a time-aligned group training dataset. Step 2: Construct the convolutional feature extraction module, the attention pooling module, and the MLP parameter output head module in sequence, and combine the above three modules into a convolutional-attention neural network; Step 3: Construct a composite loss function to train the convolutional-attention neural network built in Step 2, and obtain the initial parameter vector at the group level; Step 4: For the individual to be deployed, collect its short-term continuous data segment as a calibration segment; during calibration, use the initial parameter vector at the population level as the optimization variable, and obtain the final parameter vector x of the individual by minimizing the loss function on the calibration segment and save it. Step 5: Construct a multidimensional B-spline function mapping from the kinematic information of the target joint to the muscle and tendon unit lengths, torque arms, and pendent angles of the muscle group of the target joint; Step 6: Collect kinematic information of the tested individual using an inertial sensor, and obtain the muscle and tendon unit length, torque arm, and pendent angle of the target joint muscle group through multidimensional B-spline function mapping; The mapped data, parameter vector x, and real-time acquired electromyographic signals are input into the physical forward calculation module, which outputs the real-time torque prediction value of the target joint.

2. The method for predicting torque at multiple joints of the human body according to claim 1, characterized in that, In step 2, the convolutional feature extraction module is constructed by performing one-dimensional convolutional feature extraction on the electromyographic signals, joint kinematic information, muscle and tendon unit lengths of the target joint muscle group, and torque arm temporal data collected in step 1, in order to extract dynamic features at different time scales.

3. The method for predicting torque at multiple joints of the human body according to claim 1, characterized in that, In step 2, the attention pooling module is constructed by adaptively weighting and pooling the one-dimensional convolutional features extracted by the convolutional feature extraction module in the time dimension to obtain attention weights and pooling features.

4. The method for predicting torque at multiple joints of the human body according to claim 1, characterized in that, In step 2, the parameter output head module is constructed by mapping the pooling features into a parameter logits vector z, and then mapping it to a pre-defined upper and lower bound value range through a Sigmoid-type boundary mapping function, thus converting the parameter logits vector z into an initial parameter vector representing the physiological properties of the target joint muscle group.

5. The method for predicting torque at multiple joints of the human body according to claim 1, characterized in that, In step 3, the composite loss function is constructed by separately constructing a torque error term, a correlation and slope consistency term, a physiological regularization term, a dynamic smoothing and passive term penalty, an attention stability regularization term, and a global gain regularization term, and then weighting and combining the above six terms to form a composite loss function.

6. The method for predicting torque at multiple joints of the human body according to claim 5, characterized in that, In step 3, the torque error term is constructed using the weighted root mean square error (RMSE). Time weights w(t) are constructed based on the torque magnitude and gradient. The weight of the resting / low torque segment before joint exertion is set to 0.3, and the weight of the high torque / drastic torque change segment after the joint torque reaches the gradient threshold is set to 1.

5. The correlation and slope consistency terms are constructed by introducing correlation loss and slope loss to enhance waveform and scale consistency. The physiological regularization term is constructed by penalizing deviations from the physiologically reasonable range for the five physiological parameters of each muscle. The dynamic smoothing and passive term penalty is constructed by applying smoothing constraints to the change of muscle force over time and penalizing the passive muscle force component to avoid non-physiological bursts. The attention stability regularization term introduces entropy regularization and time smoothing regularization into the attention weights to avoid attention collapse or violent jitter under conditions of few samples. The global gain regularization term penalizes the degree to which the global gain g deviates from 1.

7. The method for predicting torque at multiple joints of the human body according to claim 1, characterized in that, In step 4, the parameter vector x includes five physiological parameters for each muscle and a global gain g.

8. The method for predicting torque at multiple joints of the human body according to claim 7, characterized in that, In step 4, the five physiological parameters for each muscle are: maximum isometric force Fmax, tendon relaxation length LtSlack, optimal muscle fiber length LmOpt, scaling factor s, and nonlinear shape parameter A for EMG conversion into muscle activation.

9. The method for predicting torque at multiple joints of the human body according to claim 8, characterized in that, In step 4, the parameter vector x is used by substituting the five physiological parameters of each muscle in the parameter vector x into the physical forward calculation module, and combining them with the electromyographic signals collected in step 1. The time-series muscle force of the muscle tendon unit is calculated by inputting the time-series muscle force into the physical forward calculation module. The joint torque prediction value is obtained by superimposing the time-series muscle force with the torque arm time-series data obtained in step 1. The physical forward calculation module consists of a series of differentiable continuous equations. Based on Hill-type muscle mechanics, it calculates the force exerted by the corresponding muscle and tendon units by inputting electromyographic signals.

10. A system for implementing the human multi-joint torque prediction method according to any one of claims 1 to 9, characterized in that, The system includes a computer module for running a human multi-joint torque prediction method.

Images