Generalized mechanics parameter prediction method and system enhanced by cross-domain generalization

Abstract

The present application belongs to the technical field of joint torque prediction, in particular to a generalized mechanics parameter prediction method and system with cross-domain generalization enhancement, aiming at the problems of insufficient generalization ability, sensitivity to monocular depth noise, single domain adaptation mechanism, lack of credibility evaluation and degradation bottom mechanism of existing monocular video joint torque prediction in cross-person crowd scenarios, the present application proposes the following scheme, which includes the following steps: S1: obtaining the human kinematics sequence of the target to be predicted, the kinematics sequence at least contains joint angle, joint angular velocity, joint angular acceleration time series data, the present application inhibits monocular noise and eliminates bone scale difference through anti-noise difference and height normalization cooperative preprocessing; adopting global and local heterogeneous domain adaptation parallel prediction, orthogonal complement, combining monte carlo credibility quantification and adaptive weighted fusion, automatically degrading bottom when low credibility, realizing high precision and high reliability of cross-person joint torque prediction.

Description

Technical Field

[0001] This invention relates to the field of joint torque prediction technology, and in particular to a method and system for predicting generalized mechanical parameters with cross-domain generalization enhancement. Background Technology

[0002] Joint torque is a core dynamic parameter in fields such as human motion analysis, rehabilitation assessment, sports biomechanics research, and human-computer interaction. Accurate joint torque data is of great value for gait abnormality diagnosis, rehabilitation training effect evaluation, and exoskeleton control strategy design. Traditional methods for accurately acquiring joint torque rely on the combined acquisition of data from a 3D motion capture system (such as Vicon) and a force platform (such as Bertec), and then calculate it through inverse dynamics. While this method offers high accuracy and reliability, it suffers from limitations such as expensive equipment (the entire system costs hundreds of thousands to millions of dollars), highly specialized operation, fixed location requirements, and the need for subjects to wear markers, making it difficult to promote and apply in scenarios such as large-scale population screening, home rehabilitation monitoring, and community sports and health management. In recent years, non-contact joint torque prediction methods based on monocular video have become a research hotspot. These methods utilize ordinary RGB cameras to acquire human motion videos and directly regress joint torques from 2D or 3D human pose estimation results using deep learning models. This significantly reduces the data acquisition threshold and equipment costs, and holds promise for convenient and universally applicable joint torque measurement. However, existing monocular video-based torque prediction technologies still face the following key technical challenges, severely limiting their generalization ability and reliability in practical applications. Challenge 1: Noise amplification in joint angle time-series data in monocular video scenarios. In 3D human pose estimation from monocular video, the depth direction (Z-axis) coordinate error is much larger than the image plane (X / Y-axis) coordinate error, typically more than 10 times the plane error, and there are low-frequency systematic drifts and inter-limb segment correlation drifts. Traditional naive numerical difference methods (such as finite difference) significantly amplify the aforementioned high-frequency noise and low-frequency drifts when performing first and second-order differentiation on joint angle time-series data to calculate angular velocity and angular acceleration. This leads to instability in subsequent training of models such as Physical Information Neural Networks (PINN) and a significant deviation of predicted torques from the true values. Although existing technologies employ ordinary low-pass filtering or moving average smoothing, these often result in over-smoothing of the signal and loss of key dynamic features in the gait, making it difficult to achieve a good balance between noise suppression and signal fidelity. Challenge 2: Inconsistent physical dimensions due to differences in skeletal scale in cross-population scenarios. There are significant differences in skeletal length and limb proportions between different individuals (e.g., children and adults, men and women, different ethnic groups). Traditional global normalization methods (such as simple scaling based on height and weight) cannot eliminate the influence of skeletal scale on kinematic features such as joint angular velocity and angular acceleration from a physical mechanism perspective, resulting in a lack of comparability of model input features across different populations. This causes a sharp decline in the generalization performance of prediction models trained well in the source domain (specific training population) on the zero-sample target population (cross-population scenarios). Experiments show that the pass rate of physical constraints for traditional single PINN models in cross-population scenarios is only about 12.2%, far from meeting the needs of practical applications. Challenge 3: Existing domain adaptation methods are limited and cannot fully cover cross-population distribution drift. To improve cross-population generalization performance, some studies have attempted to introduce domain adaptation methods, such as CORAL (covariance alignment) for global distribution matching, or DANN (adversarial domain adaptation) for learning domain-invariant features. However, existing solutions typically employ only a single domain adaptation mechanism. CORAL only focuses on the global alignment of second-order statistics, making it difficult to eliminate local domain-specific artifacts (such as local feature biases caused by different shooting environments, clothing, and lighting). While DANN can extract domain-invariant features, its adversarial training process is unstable and prone to losing discriminative information from the source domain. A single domain adaptation mechanism cannot simultaneously address both global distribution drift and local feature biases, resulting in limited improvement in cross-population generalization. Challenge 4: Lack of prediction reliability assessment and fallback mechanism for abnormal performance. Existing monocular video torque prediction methods are mostly end-to-end "black box" models, outputting a prediction result for any input, lacking a quantitative assessment of the reliability of the prediction results. When the input video quality is poor (e.g., occlusion, blurring) or the subject's motion pattern exceeds the training data distribution range, the model may output erroneous results that violate basic physical laws (e.g., joint torque exceeding physiological limits), without the user's knowledge. This "blind prediction" may pose serious risks in safety-sensitive scenarios such as rehabilitation medicine. Currently, there is no effective reliability quantification system or automatic fallback scheme for low reliability.

[0003] In existing technologies, monocular video joint torque prediction suffers from insufficient generalization ability, sensitivity to monocular depth noise, and a single domain adaptation mechanism lacking credibility assessment and fallback mechanisms in cross-crowd scenarios. To address these issues, we propose a cross-domain generalization enhancement method and system for predicting generalized mechanical parameters. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies in monocular video joint torque prediction, such as insufficient generalization ability in cross-crowd scenarios, sensitivity to monocular depth noise, single domain adaptation mechanism, and lack of credibility assessment and fallback mechanism. The invention proposes a cross-domain generalization enhanced generalized mechanical parameter prediction method and system.

[0005] The cross-domain generalization enhancement method and system for predicting generalized mechanical parameters provided in this application adopt the following technical solution: A method for predicting generalized mechanical parameters with cross-domain generalization enhancement includes the following steps: S1: Obtain the human kinematic sequence of the target to be predicted. The kinematic sequence includes at least the time-series data of joint angles, joint angular velocities, and joint angular accelerations. The joint angular velocities and joint angular accelerations are obtained by performing noise-resistant numerical difference calculation on the joint angles. The noise-resistant numerical difference calculation is designed to address the specific technical challenge of large gait signal noise in monocular video acquisition scenarios and is used to suppress the amplification effect of naive numerical difference on high-frequency gait noise. S2: Input the kinematic sequence into at least two heterogeneous domain adaptation prediction sub-models. The at least two heterogeneous domain adaptation prediction sub-models are constructed based on a prediction paradigm of physical constraint embedding and pre-trained on source domain population data. The training loss function embeds human musculoskeletal physical constraints. The prediction paradigm of physical constraint embedding includes, but is not limited to, any one or more of the following: Physical Information Neural Network (PINN), Inverse Dynamics Transformer integrating differentiable forward dynamics, and retrieval and inference model based on physical constraints. The first sub-model achieves global distribution alignment by minimizing the difference in second-order statistics between source domain features and target domain features. The second sub-model extracts domain-invariant features through gradient inversion layer adversarial training to eliminate local domain-specific artifacts. The first and second sub-models adopt a mutually orthogonal distribution correction mechanism and output at least two sets of candidate joint torque prediction values ​​respectively. S3: Based on Monte Carlo simulation, the dispersion, physical constraint pass rate, and domain adaptation residual of the at least two sets of candidate joint torque prediction values ​​are quantified to generate a cross-population credibility score. S4: Adaptively weightedly fuse the predicted values ​​of the at least two candidate joint moments based on the cross-population credibility score, and output the final estimated value of the joint moment; when the cross-population credibility score is lower than the preset credibility threshold, trigger the downgrade prediction path, call the pre-built attitude-mechanical reference database to perform temporal feature matching and constraint disambiguation, and output the last-resort estimated value of the joint moment.

[0006] Further, the kinematic sequence generation step S1 includes: acquiring monocular video data of the target to be predicted; extracting joint three-dimensional coordinate temporal data through a pre-trained monocular three-dimensional human pose estimation model; calculating joint angles frame by frame based on the three-dimensional coordinate temporal data; performing first-order and second-order numerical differentiation on the joint angles using a Savitzky-Golay smoothing difference filter to generate joint angular velocity and joint angular acceleration temporal data respectively, wherein the window length and polynomial order of the filter are adaptively set according to the human gait sampling frequency. The mouth length is taken as 5%-15% of the gait cycle, and the polynomial order is taken as 2-4. The joint angular velocity and joint angular acceleration are normalized according to the individual height parameters. The normalization process addresses the technical challenge of huge differences in skeletal scales across different populations and eliminates the inconsistency in the dimensions of angular velocity caused by differences in bone length between different individuals. The joint angle, the normalized joint angular velocity, and the joint angular acceleration are concatenated along the feature dimension, and the temporal context windows of N frames before and after are concatenated in the time dimension, where N is a positive integer, to form a spatiotemporal enhanced feature matrix as the kinematic sequence.

[0007] Furthermore, the physical constraints on human musculoskeletal system include at least one or more of the following constraint forms: Newton-Euler equation residual constraints, muscle physiological limit constraints, joint range of motion constraints, energy conservation constraints, momentum conservation constraints, and angular momentum conservation constraints; the deviation degree of the physical constraints is quantified as a loss term, which together with the data fitting loss forms a hybrid training loss function, and the model parameters are optimized through backpropagation; the weight coefficients of the physical constraints adopt a progressive scheduling strategy during training, gradually increasing from the initial low weight to the target weight, to avoid the physical constraints dominating training too early and causing a loss in data fitting performance.

[0008] Furthermore, the first sub-model uses one of the following methods to achieve second-order statistic alignment: covariance alignment, correlation alignment, or maximum mean difference; the second sub-model uses one of the following methods to achieve local domain invariant feature extraction: gradient inversion layer adversarial domain adaptation, adversarial spectral kernel matching, or Transformer contextual domain hints; the prediction results of the first and second sub-models exhibit complementary error characteristics in cross-population scenarios.

[0009] Furthermore, the weight calculation formula for the adaptive weighted fusion is as follows: ; in, For the first The fusion weights of the predicted joint torque values ​​for the group of candidates For the first The distribution difference between the candidate predicted values ​​of the group and the source domain training data is normalized to the [0,1] interval to avoid negative weights. For the first Physical constraint pass rate of candidate predicted values. For the first Cross-population reliability score of candidate predicted values ​​for a group The total number of candidate predicted values; the number of samples in the Monte Carlo simulation is not less than the minimum number of samples required for the confidence interval half-width of the fusion weights to converge to the preset target accuracy.

[0010] Furthermore, the formula for calculating the cross-group credibility score is as follows: ; in, The discrete standard deviation of the candidate predicted values. The mean of the candidate predicted values. Below the preset lower limit Time Alternative To avoid division by zero exceptions, The typical value is 1e-6 Nm / kg. The proportion of candidate predicted values ​​that pass all hard physical constraints. The maximum value of the domain-adapted residual. The preset domain adaptation residual threshold is used; the prediction results are divided into three levels—high confidence level, medium confidence level, and low confidence level—based on the cross-population confidence score. The low confidence level triggers the downgrade prediction path described in claim 1.

[0011] Furthermore, in the scenario of predicting human joint torques, the target to be predicted is a human subject; the kinematic sequence consists of the angles, angular velocities, and angular accelerations of the hip / knee / ankle joints; the mechanical parameters are the knee adduction torque (KAM) and / or the hip and / or ankle torques, in Nm / kg; the source domain data originates from the gold-standard inverse dynamics calculation results jointly acquired by the 3D motion capture system Vicon and the force platform Bertec, conforming to the ISO / FDIS 23471 biomechanical measurement standard, and does not include synthetic data; the physical constraints preferentially adopt a combination of Newton-Euler equation residual constraints and muscle physiological limit constraints; the attitude-mechanical reference database called by the degraded prediction path stores the gold-standard joint torque curves, and the Dynamic Time Warping (DTW) algorithm is used for time-series matching.

[0012] Furthermore, it also includes a timing consistency constraint step: performing adjacent frame torque change rate verification on the final joint torque estimate or the fallback joint torque estimate; when the adjacent frame joint torque change rate exceeds a preset physiological threshold, performing smoothing correction or marking as an abnormal data point on the current frame torque value; the preset physiological threshold is determined based on the upper limit of the physiological change rate of human joint torque.

[0013] Furthermore, it also includes observable outputs and traceability steps: The cross-domain credibility score generated by S3, the final mechanical parameter estimate or catch-all mark output by S4, and the fusion weight allocation result are written to the prediction process log file in real time and output to the outside world through a standardized data interface. The standardized data interface includes, but is not limited to, any one or more of the following: RESTful API interface, serial port debugging port, local log file, and human-machine interface display unit. The prediction process log contains a complete record chain of timestamps, output values ​​of each sub-model, fusion weights, credibility ratings, and whether a downgrade is triggered, which is used for quality traceability, third-party verification, or evidence preservation.

[0014] Furthermore, both the source population data and the gold standard joint torque data in the posture-mechanical reference database are derived from the inverse dynamics calculation results acquired by the combination of the three-dimensional motion capture system and the force table, and conform to the ISO / FDIS23471 biomechanical measurement standard. The gold standard data does not contain synthetic data.

[0015] Furthermore, the prediction paradigm of the physical constraint embedding is preferably implemented using the Physical Information Neural Network (PINN), in which at least one of the following is embedded in the loss function of the PINN: Newton-Euler equation residual constraint, muscle physiological limit constraint, and joint range of motion constraint; alternative implementations of the prediction paradigm of the physical constraint embedding include the inverse dynamics Transformer integrating differentiable forward dynamics and the retrieval and inference model based on physical constraints.

[0016] Furthermore, it also includes a gait cycle verification step: performing gait phase consistency verification on the final joint torque estimate or catch-up value, detecting whether the torque abrupt change at the moment of switching between the support phase and the swing phase conforms to the physiological pattern; when an abnormal abrupt change is detected, cubic spline interpolation is used for correction or it is marked as an abnormal frame; the physiological threshold is determined statistically based on the gait database of healthy adults.

[0017] Furthermore, in the scenario of robot operation torque prediction, the target to be predicted is the robot's end effector or joint drive unit; the kinematic sequence is the end pose (position + attitude six degrees of freedom) and its time derivative sequence; the mechanical parameters are operation contact torque and / or joint drive torque and / or end interaction force; the source domain data comes from at least one of the following: diverse operation task annotation data generated by a physics simulation engine (including but not limited to Isaac Gym, MuJoCo, PyBullet), operation datasets collected by real robots equipped with force / torque sensors, and ground-truth output of a digital twin simulation platform; the physical constraints in the robot scenario preferentially adopt a combination of contact force feasibility constraints, joint torque boundary constraints, and operation space collision avoidance constraints; the domain adaptation is for at least one of the following drift scenarios: Sim-to-Real migration, operation generalization between different object categories, and visual perception drift under different ambient lighting and texture conditions.

[0018] Furthermore, it also includes a world model prediction verification step: before S4 outputs the final estimate, a pre-trained world action model (WAM) or other operation strategy model with forward state prediction capability is called to perform forward prediction of the future multi-frame state of the current operation intention; potential collision risk, boundary risk and contact anomaly risk are detected based on the prediction sequence; if a high-risk event is detected, the cross-domain credibility score is multiplied by a safety discount factor of less than 1 and then re-enters the fusion decision process of S4.

[0019] This invention also proposes a cross-domain generalization enhancement system for predicting generalized mechanical parameters, used to implement the above method, comprising: The kinematic sequence acquisition module is used to acquire the human kinematic sequence of the target to be predicted; The heterogeneous domain adaptation prediction module, which includes a first sub-model and a second sub-model, is used to output at least two sets of candidate joint torque prediction values ​​in parallel. The credibility metric module is used to generate cross-population credibility scores based on Monte Carlo simulations; The fusion output module is used to perform adaptive weighted fusion based on the cross-population credibility score, and trigger a downgraded prediction path to call the attitude-mechanical reference database for matching when the credibility is lower than the threshold.

[0020] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the method described above.

[0021] In summary, this application includes at least one of the following beneficial technical effects: 1. This scheme uses a Savitzky-Golay smoothed differential filter instead of naive numerical differential, effectively suppressing the high-frequency amplification effect of monocular video depth noise during the differentiation process. Furthermore, in the preferred embodiment, 2D projection angles are used to calculate joint angles. Addressing the inherent deficiency of high depth noise in monocular vision, experimental verification shows that compared to traditional 3D angle calculations, 2D projection angles can reduce the standard deviation of joint angles by more than 77.8% (89.0%~90.7% for the hip joint, 89.3%~90.1% for the knee joint, and 51.4%~55.3% for the ankle joint), thus improving the stability of kinematic characteristics from the source. 2. This scheme normalizes joint angular velocity and angular acceleration according to individual height parameters, rather than the traditional global mean-variance normalization. This effectively eliminates the differences in kinematic feature dimensions caused by differences in bone length among different individuals, making the model input features comparable across population groups and providing a good feature foundation for domain adaptation. 3. This scheme employs two heterogeneous domain adaptation mechanisms in parallel prediction: global distributed alignment (CORAL / MMD) and local domain invariant feature extraction (DANN / adversarial training). These two mechanisms are orthogonal to each other and their errors are complementary. Ablation experiment data (see detailed implementation) shows that the pass rate (physical constraint pass rate) of a single MMD-PINN is 45.3%, that of a single CORAL-PINN is 58.1%, and that of a single DANN-PINN is 55.7%. In contrast, the CORAL+DANN fusion scheme of this invention achieves a pass rate as high as 70.7%, a Pearson correlation coefficient of 0.887, and a MAE reduction to 0.056 Nm / kg, representing a qualitative leap compared to the domain-independent baseline (pass rate 12.2%).

[0022] This invention suppresses monocular noise and eliminates skeletal scale differences through collaborative preprocessing of noise-resistant differential and height normalization; it adopts global and local heterogeneous domain adaptive parallel prediction and positive interactive complementation, combined with Monte Carlo confidence quantification and adaptive weighted fusion, and automatically downgrades to a fallback when confidence is low, to achieve high accuracy and high reliability in cross-population joint torque prediction. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the structure of the cross-domain generalization enhancement method for predicting generalized mechanical parameters proposed in this invention. Figure 2 This is a block diagram of the cross-domain generalization enhancement generalized mechanical parameter prediction system proposed in this invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] It should be noted that those skilled in the art will understand that the specific models (such as Physical Information Neural Network PINN, MediaPipePose pose estimation model, DANN domain adaptation network, Recursive Newton-Euler Algorithm RNEA, etc.), specific algorithms (such as CORAL alignment, MMD matching, Fourier transform frequency domain decomposition, etc.), and specific system architectures and parameter configurations shown in the above embodiments and the embodiments below are all preferred embodiments or illustrative examples of the present invention. Their purpose is to help those skilled in the art better understand the technical solutions and technical effects of the present invention, rather than to limit the scope of protection of the present invention.

[0026] Based on the technical concept disclosed in this invention, those skilled in the art can replace the specific implementation details described in the above specific embodiments with other functionally equivalent or technically equivalent models, algorithms, or architectures without creative effort. For example: the biomechanical perception model can adopt other physically constrained embedded neural networks (such as Hamiltonian neural networks, deep operator networks, etc.) besides PINN; the pose estimation model can adopt other monocular 3D human pose estimation methods besides MediaPipePose; the domain adaptation mechanism can adopt other domain generalization methods besides DANN / CORAL; and the physics solver can adopt other multibody dynamics solution methods besides RNEA (such as Lagrange method, Hamiltonian method, etc.).

[0027] All possible equivalent substitutions, modifications, or variations described above shall be deemed to fall within the protection scope of this invention.

[0028] Example 1 Reference Figure 1 and Figure 2 A method for predicting generalized mechanical parameters with cross-domain generalization enhancement includes the following steps: S1: Obtain the human kinematic sequence of the target to be predicted. The kinematic sequence includes at least the time-series data of joint angles, joint angular velocities, and joint angular accelerations. The joint angular velocities and joint angular accelerations are obtained by performing noise-resistant numerical difference calculations on the joint angles. The noise-resistant numerical difference calculation is designed to address the specific technical challenge of large gait signal noise in monocular video acquisition scenarios, and is used to suppress the amplification effect of naive numerical difference on high-frequency gait noise. This example compares the impact of using 2D projection angle calculation versus 3D angle calculation on the time series standard deviation of joint angles in a monocular video input scenario. Noise model settings: Gaussian noise with a standard deviation of σ = 0.003 is added to the two-dimensional (x,y) coordinates (simulating MediaPipe pixel-level precision), and Gaussian noise with a standard deviation of σ = 0.03 (10 times that of xy) is added to the three-dimensional depth (z) coordinates. At the same time, low-frequency systematic drift and limb-segment related drift with an amplitude of 0.02 are superimposed. Verification results:

[0029] The pass rate was 6 / 6 (100%), and the average standard deviation decreased by 77.8%, far exceeding the preset target (>30%). This example demonstrates that the present invention achieves significantly better noise reduction performance than expected by using 2D projection angle calculation combined with SG differential filtering under monocular video input. It should be emphasized that the above-mentioned technical effects do not hold true in binocular or optical motion capture scenarios; the technical advantages of this invention are precisely limited to monocular video input scenarios. The steps for generating the kinematic sequence include: acquiring monocular video data of the target to be predicted; extracting joint 3D coordinate temporal data using a pre-trained monocular 3D human pose estimation model; calculating joint angles frame by frame based on the 3D coordinate temporal data; performing first- and second-order numerical differentiation on the joint angles using a Savitzky-Golay smoothing differential filter to generate joint angular velocity and joint angular acceleration temporal data, respectively. The filter window length and polynomial order are adaptively set according to the human gait sampling frequency, with the window length being 5%-15% of the gait cycle and the polynomial order being 2-4; normalizing the joint angular velocity and joint angular acceleration according to individual height parameters. This normalization process addresses the technical challenge of significant differences in bone scale across different populations, eliminating inconsistencies in angular velocity dimensions caused by differences in bone length between individuals; concatenating the joint angles, normalized joint angular velocity, and joint angular acceleration along the feature dimension, and concatenating the temporal context windows of the preceding and following N frames in the time dimension, where N is a positive integer, to form a spatiotemporal enhanced feature matrix as the kinematic sequence. A preferred implementation of step S1: (1) Monocular video acquisition and 3D pose estimation; The walking video of the target to be predicted is captured using a regular RGB camera (such as a mobile phone camera or webcam) at a frame rate of 30-60fps. The video frames are then input into a pre-trained MediaPipePose model (or ViTPose, HRNet, etc.) to extract the 3D coordinate temporal data of 17 or 33 key points of the human body. Due to the inherent noise of monocular depth estimation, the Z-axis coordinate error is significantly larger than that of the X / Y axis. (2) Joint angle calculation; To mitigate depth noise, this embodiment prioritizes the 2D projection angle calculation method: 3D keypoints are projected onto the image plane (X / Y axes), and joint angles (such as hip flexion angle, knee flexion angle, ankle dorsiflexion angle, etc.) are calculated on the 2D plane. Specifically, for the knee flexion angle, the angle between the thigh vector and the lower leg vector on the image plane is calculated. Experiments show (see Supplementary Example 1 below) that the standard deviation of the 2D projection angle in monocular video scenes is much lower than that of the 3D angle, resulting in significant noise suppression. (3) Noise immunity numerical difference; Savitzky-Golay (SG) smoothing differential filters were used to perform first- and second-order differentiation on the joint angle time-series data. The filter parameters were set as follows: the window length was adaptively set according to the gait cycle; for example, when the gait cycle was approximately 1 second and the sampling frequency was 30Hz, the window length was set to 15–45 frames (i.e., 5%–15% of the gait cycle), and the polynomial order was set to 3. The SG filter effectively suppressed high-frequency noise amplification by simultaneously performing polynomial least-squares fitting within the sliding window, achieving both smoothing and differentiation. (4) Normalize according to individual height; Obtain the height of the target to be predicted (or the bone length estimated by a pose estimation model), multiply the joint angular velocity (in rad / s) by the height factor (e.g., height / reference height), and multiply the angular acceleration (in rad / s). 2 Multiply by the square of the height factor to eliminate the dimensional inconsistency caused by differences in skeletal scale. This normalization strategy originates from the conversion relationship between angular velocity and linear velocity in rigid body kinematics and has a clear physical meaning. (5) Enhanced spatiotemporal features; The joint angle, normalized angular velocity, and angular acceleration are concatenated along the feature dimension, and a context window is constructed in the time dimension with N frames before and after (N is 7) to form a spatiotemporal enhanced feature matrix of (time step × feature dimension), which is used as the input of the subsequent prediction model. S2: Input the kinematic sequence into at least two heterogeneous domain-adaptive prediction sub-models. Both of the at least two heterogeneous domain-adaptive prediction sub-models are constructed based on a prediction paradigm of physical constraint embedding and pre-trained on source domain population data. The training loss function embeds human musculoskeletal physical constraints. The prediction paradigm of physical constraint embedding includes, but is not limited to, any one or more of the following: Physical Information Neural Network (PINN), Inverse Dynamics Transformer integrating differentiable forward dynamics, and retrieval and inference model based on physical constraints. The first sub-model achieves global distribution alignment by minimizing the difference in second-order statistics between source domain features and target domain features. The second sub-model extracts domain-invariant features through gradient inversion layer adversarial training to eliminate local domain-specific artifacts. The first and second sub-models adopt a mutually orthogonal distribution correction mechanism and output at least two sets of candidate joint torque prediction values ​​respectively. The physical constraints of human musculoskeletal system include at least one or more of the following constraint forms: Newton-Euler equation residual constraints, muscle physiological limit constraints, joint range of motion constraints, energy conservation constraints, momentum conservation constraints, and angular momentum conservation constraints. The deviation of the physical constraints is quantified as a loss term, which together with the data fitting loss forms a hybrid training loss function, and the model parameters are optimized through backpropagation. The weight coefficients of the physical constraints adopt a progressive scheduling strategy during training, gradually increasing from the initial low weights to the target weights, to avoid the physical constraints dominating the training too early and causing a loss in data fitting performance. The prediction paradigm with physical constraint embedding is preferably implemented using the Physical Information Neural Network (PINN). The loss function of PINN embeds at least one of the following: Newton-Euler equation residual constraints, muscle physiological limit constraints, and joint range of motion constraints. Alternative implementations of the prediction paradigm with physical constraint embedding include the inverse dynamics Transformer that integrates differentiable forward dynamics and the retrieval and inference model based on physical constraints. The first sub-model uses one of the following methods to achieve second-order statistic alignment: covariance alignment, correlation alignment, or maximum mean difference. The second sub-model uses one of the following methods to achieve local domain invariant feature extraction: gradient inversion layer adversarial domain adaptation, adversarial spectral kernel matching, or Transformer contextual domain hints. The prediction results of the first and second sub-models exhibit complementary error characteristics in cross-population scenarios. Specific implementations of the first and second sub-models: Source domain data preparation: The K2MUSE public dataset (or an internally acquired laboratory-grade dataset) was used. This dataset contains dynamic data of 30 healthy adult subjects walking on a treadmill, simultaneously acquired via Vicon optical motion capture (8 cameras, 100Hz) and a Bertec force plate. Joint torques were calculated using Plug-inGait inverse dynamics to obtain the gold standard. The data conforms to the ISO / FDIS 23471 standard. Source domain data was used to pre-train two sub-models. Prediction paradigms with physically constrained embeddings (taking PINN as an example): Using kinematic sequences (joint angles, angular velocities, and angular accelerations) as input and joint torques as output, a fully connected neural network is constructed. The loss function consists of data fitting loss (MSE of predicted torque versus gold standard torque) and physical constraint loss. The physical constraints include at least one or more of the following: Newton-Euler equation residual constraint: the difference between the theoretical torque and the predicted torque calculated based on joint angles, angular velocities, angular accelerations and inertial parameters; Muscle physiological limit constraint: The predicted torque should not exceed the physiological maximum value of the joint; Joint range of motion constraints: The joint angle corresponding to the predicted torque should be within the normal range of motion; Energy conservation / momentum conservation constraints; The deviation from physical constraints is quantified as a loss term, which is then weighted and summed with the data fitting loss. A progressive scheduling strategy is used for the weight coefficients: the physical constraint weight is initially set to 0.01 and gradually increased to 0.1 to prevent physical constraints from dominating too early and causing underfitting of the data. First sub-model (global distribution alignment): Building upon PINN, a CORAL (covariance alignment) loss term is added to minimize the covariance difference between the source domain features and the target domain (the domain containing the current test sample). Specifically, the Frobenius norm distance between the source domain feature covariance matrix and the target domain feature covariance matrix is ​​calculated in an intermediate layer of the network and added as an additional loss term during training. After training, this model can globally narrow the feature distributions of the source and target domains. Second sub-model (locally invariant features): Based on PINN, a gradient inversion layer (GRL) is introduced for adversarial training. Specifically, a domain classifier (binary classification, determining whether a feature comes from the source domain or the target domain) is connected after the feature extractor. Through GRL, the feature extractor learns the feature representation of the confused domain classifier, thereby extracting domain-invariant features. This model can eliminate local domain-specific artifacts (such as feature biases caused by lighting, clothing, shooting angle, etc.). The two sub-models adopt a mutually orthogonal distribution correction mechanism (global covariance alignment vs. local adversarial features) and output candidate moment prediction values ​​respectively. The two exhibit complementary error characteristics in cross-population scenarios. S3: Based on Monte Carlo simulation, the dispersion, physical constraint pass rate, and domain adaptation residual of at least two sets of candidate joint torque prediction values ​​are quantified to generate a cross-population credibility score. Monte Carlo simulation: Apply small-amplitude Gaussian noise (standard deviation is 5% of the feature standard deviation) to the input kinematic sequence, and repeat the forward propagation 50-100 times to obtain multiple sets of predicted moment sequences. Calculate the mean and standard deviation of each set of candidate predicted values, as well as the physical constraint pass rate (i.e., the proportion of predicted values ​​that satisfy all hard constraints), and simultaneously calculate the domain adaptation residuals (such as CORAL loss and DANN domain classifier loss). S4: Adaptively weighted fusion of at least two candidate joint moment predictions based on cross-population credibility scores, and output the final joint moment estimate; when the cross-population credibility score is lower than the preset credibility threshold, a downgrade prediction path is triggered, and a pre-built attitude-mechanical reference database is called to perform temporal feature matching and constraint disambiguation, and output the fallback joint moment estimate. The formula for calculating the weights in adaptive weighted fusion is: ; in, For the first The fusion weights of the predicted joint torque values ​​for the group of candidates For the first The distribution difference between the candidate predicted values ​​of the group and the source domain training data is normalized to the [0,1] interval to avoid negative weights. For the first Physical constraint pass rate of candidate predicted values. For the first Cross-population reliability score of candidate predicted values ​​for a group The total number of candidate predicted values; the number of samples in the Monte Carlo simulation is not less than the minimum number of samples required for the confidence interval half-width of the fusion weights to converge to the preset target accuracy; The formula for calculating cross-group credibility score is: ; in, The discrete standard deviation of the candidate predicted values. The mean of the candidate predicted values. Below the preset lower limit Time Alternative To avoid division by zero exceptions, Typical values ​​are 1e-6 Nm / kg and are given in the specific implementation details in the specification. The proportion of candidate predicted values ​​that pass all hard physical constraints. The maximum value of the domain-adapted residual. The preset domain adaptation residual threshold is used (e.g., twice the mean residual on the source domain validation set). Based on the cross-population confidence score, the prediction results are divided into three levels: high confidence level, medium confidence level, and low confidence level. The low confidence level triggers a downgrade prediction path. A score ≥ 0.7 is high confidence level, 0.4~0.7 is medium confidence level, and < 0.4 is low confidence level. Downgrade fallback plan: When the confidence score is below 0.4, a downgrade path is triggered. The system retrieves the most similar temporal features to the current kinematic sequence from a pre-built attitude-mechanical reference database (using Dynamic Time Warping (DTW) distance), and returns the joint moment curves in the corresponding database as fallback estimates. The reference database contains gold standard moment data from the source domain and multiple representative target populations (different ages, genders, heights, etc.). In this embodiment, the target to be predicted is a human subject; the kinematic sequence is the angle, angular velocity, and angular acceleration of the hip / knee / ankle joints; the mechanical parameters are the knee joint adduction torque KAM and / or the hip joint torque and / or the ankle joint torque, in Nm / kg; The source domain data comes from the gold standard inverse dynamics calculation results jointly acquired by the 3D motion capture system Vicon and the force table Bertec, and conforms to the ISO / FDIS 23471 biomechanical measurement standard, and does not include synthetic data; Physical constraints are preferably a combination of Newton-Euler equation residual constraints and muscle physiological limit constraints. The attitude-mechanics reference database called by the downgrade prediction path stores the gold standard joint moment curves, and the dynamic time warping (DTW) algorithm is used for timing matching.

[0030] In this embodiment, a gait cycle verification step is also included: Gait phase consistency verification is performed on the final joint torque estimate or catch-up value to detect whether the torque abrupt change at the moment of switching between the support phase and the swing phase conforms to the physiological pattern; when an abnormal abrupt change is detected, cubic spline interpolation is used for correction or it is marked as an abnormal frame; the physiological threshold is determined based on the gait database of healthy adults.

[0031] In this embodiment, it is applied to a robot operating torque prediction scenario. The target to be predicted is the robot's end effector or joint drive unit; the kinematic sequence is the end pose (position + attitude six degrees of freedom) and its time derivative sequence; the mechanical parameters are the operating contact torque and / or joint drive torque and / or end interaction force; The source domain data comes from at least one of the following: diverse operation task annotation data generated by physical simulation engines (including but not limited to Isaac Gym, MuJoCo, PyBullet), operation datasets collected by real robots equipped with force / torque sensors, and ground-truth output of digital twin simulation platforms; In robotic scenarios, physical constraints are preferably a combination of contact force feasibility constraints, joint moment boundary constraints, and operation space collision avoidance constraints. Domain adaptation addresses at least one of the following drift scenarios: Sim-to-Real migration, operation generalization between different object categories, and visual perception drift under different ambient lighting and texture conditions.

[0032] In this embodiment, a world model prediction verification step is also included: Before S4 outputs the final estimate, a pre-trained World Action Model (WAM) or other operation strategy model with forward state prediction capability is invoked to perform forward prediction of the future multi-frame state of the current operation intention; potential collision risks, boundary risks, and contact anomaly risks are detected based on the prediction sequence; if a high-risk event is detected, the cross-domain credibility score is multiplied by a safety discount factor of less than 1 and then re-enters the S4 fusion decision process.

[0033] It also includes a temporal consistency constraint step: For the final joint moment estimate or the fallback joint moment estimate, an adjacent frame moment change rate check is performed. When the adjacent frame joint moment change rate exceeds a preset physiological threshold, the current frame moment value is smoothed or marked as an abnormal data point. The preset physiological threshold is determined based on the upper limit of the physiological rate of change of human joint moment. After outputting the final moment estimate, the adjacent frame moment change rate check is performed. The preset physiological threshold is determined based on the upper limit of the physiological rate of change of human joint moment; for example, the rate of change of knee flexion moment does not exceed 50 Nm / kg / s per frame (based on 100 Hz sampling). If the current frame moment change rate exceeds the threshold, median filtering or cubic spline interpolation is used for smoothing correction, and the point is marked as an abnormal data point for subsequent analysis reference. The gold standard joint torque data in the source domain population data and posture-mechanical reference database are all derived from the inverse dynamics calculation results collected by the combination of the three-dimensional motion capture system and the force table, and conform to the ISO / FDIS23471 biomechanical measurement standard. The gold standard data does not include synthetic data. This example uses the OpenCapLabValidation dataset (9 subjects × 6 walks = 54 trials) to verify the MAE (mean absolute error) of the 2D angle calculation method proposed in this invention with the OpenSim3D gold standard angle, and compares it with the QMT benchmark (international open method accuracy benchmark).

[0034] Results analysis: The knee flexion MAE of 3.71° is better than the QMT benchmark of 4°, proving that the 2D angle calculation of this invention achieves the same accuracy as the internationally published method in the estimation of major joint angles in the sagittal plane. The reason for the hip / ankle / lumbar spine exceeding the standard is the systematic deviation between the 2D angle definition and the OpenSim3D definition (such as the difference in anatomical coordinate system). This is a definition deviation rather than random noise, which can be eliminated by projection plane calibration and coordinate alignment. It does not affect the core innovation of this invention (noise-resistant numerical difference and domain-adaptive fusion method). The source domain data and the gold standard joint torque data in the reference database used in this invention are both derived from the K2MUSE dataset, as evidenced below:

[0035] Original text of K2MUSE paper: "The kinematic and ground reaction force data were collected using a Vicon motion capture system and an instrumented treadmill with embedded force plates... The dynamic Plug-in Gait model subsequently calculated joint kinematics and kinetics by integrating motion capture data with participant-specific anthropometric characteristics."; The above chain of evidence proves that the K2MUSE gold standard does indeed come from Vicon + force table + Plug-inGait inverse dynamics, and the ISO / FDIS23471 statement of claim 8 of this invention is completely legal; The validation dataset for this invention is a walking scenario of healthy adults. The prediction effect for pathological populations such as osteoarthritis (OA) needs additional validation, but this does not affect the application of the method of this invention in healthy populations and populations that can be generalized.

[0036] This embodiment also proposes a generalized mechanical parameter prediction system with cross-domain generalization enhancement to implement the above method, including: The kinematic sequence acquisition module is used to acquire the human kinematic sequence of the target to be predicted, including a video input interface, a pose estimation unit, an SG differential filter, a height normalization unit, and a feature splicing unit. The heterogeneous domain adaptation prediction module includes a first sub-model and a second sub-model, which is used to output at least two sets of candidate joint moment prediction values ​​in parallel. The two models run in parallel and output candidate moment prediction values. The credibility quantification module is used to generate cross-population credibility scores based on Monte Carlo simulation. It includes a Monte Carlo sampling unit, a physical constraint testing unit, a domain adaptation residual calculation unit, and a credibility scoring unit. The fusion output module is used to perform adaptive weighted fusion based on cross-population confidence scores, and triggers the downgrade prediction path to call the attitude-mechanical reference database matching when the confidence is lower than the threshold. It includes a weighted fusion unit, a downgrade judgment unit, and an attitude-mechanical reference database matching unit (the database is stored locally or in the cloud), and finally outputs the joint torque estimate. Each module can be implemented in software and deployed on edge computing devices or cloud servers to receive real-time video streams or offline video files.

[0037] This embodiment also proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, the above-described method is implemented. The storage medium includes, but is not limited to, ROM, RAM, hard disk, solid-state drive, USB flash drive, CD-ROM, etc.

[0038] Example 2 The difference between this embodiment and Embodiment 1 is that it also includes a physical constraint verification module, used to verify the physical feasibility of the candidate joint torque prediction values ​​output by the heterogeneous domain adaptation prediction module and the final joint torque estimate values ​​output by the fusion output module. The physical constraint verification module embeds a Newton-Euler equation residual calculation unit, a muscle physiological limit determination unit, and a joint range of motion determination unit. These units are used to calculate the residual between the predicted torque and the theoretical torque derived from the kinematic sequence and inertial parameters, determine whether the predicted torque exceeds the physiological maximum value of the corresponding joint, and determine whether the joint angle corresponding to the predicted torque is within the normal physiological range of motion. When the verification finds that any physical constraint deviation exceeds a preset tolerance threshold, a physical alarm flag is generated and fed back to the credibility quantification module as a correction factor for cross-population credibility scoring, thereby reducing the fusion weight of prediction results that deviate from physical constraints or directly marking them as unreliable candidate values.

[0039] Example 3 The difference between this embodiment and Embodiment 1 is that it also includes a temporal consistency verification module, which performs temporal smoothness checks and anomaly corrections on the final joint moment estimate output by the fusion output module or the fallback joint moment estimate output by the downgraded prediction path. The temporal consistency verification module includes a differential rate of change calculation unit and an adaptive smoothing filtering unit. The differential rate of change calculation unit calculates the first and second differences of the joint moment between adjacent time frames to obtain the temporal curve of the moment change rate. The adaptive smoothing filtering unit determines the current frame as an abnormal data point when the moment change rate of adjacent frames exceeds a preset physiological threshold, and uses median filtering or cubic spline interpolation to smooth and correct the abnormal frame moment value, while generating an anomaly marker record for subsequent analysis. The preset physiological threshold is determined based on the upper limit of the physiological rate of change of human joint moment.

[0040] Example 4 The difference between this embodiment and Embodiment 1 is that it also includes an attitude-mechanical reference database management module, used to construct, maintain, and retrieve the attitude-mechanical reference database upon which the degraded prediction path depends. The attitude-mechanical reference database management module includes a data entry unit, a feature indexing unit, and a similarity retrieval unit. The data entry unit stores the gold standard joint torque data of the source domain population and representative target population into the database according to a unified temporal format and coordinate system. The gold standard data originates from the inverse dynamics calculation results acquired by the combination of a 3D motion capture system and a force table, conforming to the ISO / FDIS 23471 biomechanical measurement standard. The feature indexing unit extracts temporal feature descriptors from the entered data and constructs an efficient index structure. The similarity retrieval unit receives the current kinematic sequence as query input when the degraded prediction path is triggered, uses the Dynamic Time Warping (DTW) algorithm to calculate the temporal feature similarity between the query sequence and each entry in the database, and returns the joint torque curves corresponding to the several entries with the highest similarity as fallback estimates to the fusion output module.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting generalized mechanical parameters with cross-domain generalization enhancement, characterized in that: Includes the following steps: S1: Obtain the human kinematic sequence of the target to be predicted. The kinematic sequence includes at least the time-series data of joint angles, joint angular velocities, and joint angular accelerations. The joint angular velocities and joint angular accelerations are obtained by performing noise-resistant numerical difference calculation on the joint angles. The noise-resistant numerical difference calculation is designed to address the specific technical challenge of large gait signal noise in monocular video acquisition scenarios and is used to suppress the amplification effect of naive numerical difference on high-frequency gait noise. S2: Input the kinematic sequence into at least two heterogeneous domain adaptation prediction sub-models. The at least two heterogeneous domain adaptation prediction sub-models are constructed based on a prediction paradigm of physical constraint embedding and pre-trained on source domain population data. The training loss function embeds human musculoskeletal physical constraints. The prediction paradigm of physical constraint embedding includes, but is not limited to, any one or more of the following: Physical Information Neural Network (PINN), Inverse Dynamics Transformer integrating differentiable forward dynamics, and retrieval and inference model based on physical constraints. The first sub-model achieves global distribution alignment by minimizing the difference in second-order statistics between source domain features and target domain features. The second sub-model extracts domain-invariant features through gradient inversion layer adversarial training to eliminate local domain-specific artifacts. The first and second sub-models adopt a mutually orthogonal distribution correction mechanism and output at least two sets of candidate joint torque prediction values ​​respectively. S3: Based on Monte Carlo simulation, the dispersion, physical constraint pass rate, and domain adaptation residual of the at least two sets of candidate joint torque prediction values ​​are quantified to generate a cross-population credibility score. S4: Adaptively weightedly fuse the predicted values ​​of the at least two candidate joint moments based on the cross-population credibility score, and output the final estimated value of the joint moment; when the cross-population credibility score is lower than the preset credibility threshold, trigger the downgrade prediction path, call the pre-built attitude-mechanical reference database to perform temporal feature matching and constraint disambiguation, and output the last-resort estimated value of the joint moment.

2. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The kinematic sequence generation step S1 includes: acquiring monocular video data of the target to be predicted; extracting joint three-dimensional coordinate temporal data through a pre-trained monocular three-dimensional human pose estimation model; calculating joint angles frame by frame based on the three-dimensional coordinate temporal data; performing first-order and second-order numerical differentiation on the joint angles using a Savitzky-Golay smoothing differential filter to generate joint angular velocity and joint angular acceleration temporal data respectively, wherein the window length and polynomial order of the filter are adaptively set according to the human gait sampling frequency, with the window length being 5%-15% of the gait cycle and the polynomial order being 2-4; normalizing the joint angular velocity and joint angular acceleration according to individual height parameters, wherein the normalization process addresses the technical challenge of significant differences in skeletal scale across populations, eliminating inconsistencies in angular velocity dimensions caused by differences in bone length between different individuals; concatenating the joint angles, normalized joint angular velocities, and joint angular accelerations along the feature dimension, and concatenating the temporal context windows of N frames before and after in the time dimension, where N is a positive integer, to form a spatiotemporal enhancement feature matrix as the kinematic sequence.

3. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The physical constraints on human muscles and bones include at least one or more of the following constraint forms: Newton-Euler equation residual constraints, muscle physiological limit constraints, joint range of motion constraints, energy conservation constraints, momentum conservation constraints, and angular momentum conservation constraints; the deviation degree of the physical constraints is quantified as a loss term, which together with the data fitting loss forms a hybrid training loss function, and the model parameters are optimized through backpropagation; the weight coefficients of the physical constraints adopt a progressive scheduling strategy during training, gradually increasing from the initial low weight to the target weight, to avoid the physical constraints dominating the training too early and causing a loss in data fitting performance.

4. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The first sub-model uses one of the following methods to achieve second-order statistic alignment: covariance alignment, correlation alignment, or maximum mean difference; the second sub-model uses one of the following methods to achieve local domain invariant feature extraction: gradient inversion layer adversarial domain adaptation, adversarial spectral kernel matching, or Transformer context domain hints. The prediction results of the first sub-model and the second sub-model exhibit complementary error characteristics in cross-population scenarios.

5. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The formula for calculating the weights in the adaptive weighted fusion is as follows: ； in, For the first The fusion weights of the predicted joint torque values ​​for the group of candidates For the first The distribution difference between the candidate predicted values ​​of the group and the source domain training data is normalized to the [0,1] interval to avoid negative weights. For the first Physical constraint pass rate of candidate predicted values. For the first Cross-population reliability score of candidate predicted values ​​for a group The total number of candidate predicted values; the number of samples in the Monte Carlo simulation is not less than the minimum number of samples required for the confidence interval half-width of the fusion weights to converge to the preset target accuracy.

6. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The formula for calculating the cross-population credibility score is as follows: ； in, The discrete standard deviation of the candidate predicted values. The mean of the candidate predicted values. Below the preset lower limit Time Alternative To avoid division by zero exceptions, The typical value is 1e-6 Nm / kg. The proportion of candidate predicted values ​​that pass all hard physical constraints. The maximum value of the domain-adapted residual. The preset domain adaptation residual threshold is used; the prediction results are divided into three levels—high confidence level, medium confidence level, and low confidence level—based on the cross-population confidence score. The low confidence level triggers the downgrade prediction path described in claim 1.

7. The cross-domain generalization enhancement method for predicting generalized mechanical parameters according to claim 1, applied to the scenario of predicting human joint torques, is characterized in that: The target to be predicted is a human subject; the kinematic sequence consists of the angles, angular velocities, and angular accelerations of the hip / knee / ankle joints; the mechanical parameters are the knee adduction torque (KAM) and / or hip and / or ankle torques, in Nm / kg; the source domain data is derived from the gold-standard inverse dynamics calculation results jointly acquired by the 3D motion capture system Vicon and the force platform Bertec, conforming to the ISO / FDIS 23471 biomechanical measurement standard, and does not include synthetic data; the physical constraints preferentially adopt a combination of Newton-Euler equation residual constraints and muscle physiological limit constraints; the attitude-mechanical reference database called by the degraded prediction path stores the gold-standard joint torque curves, and the Dynamic Time Warping (DTW) algorithm is used for time-series matching.

8. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: It also includes a timing consistency constraint step: performing adjacent frame torque change rate verification on the final joint torque estimate or the fallback joint torque estimate; when the adjacent frame joint torque change rate exceeds a preset physiological threshold, performing smoothing correction or marking as an abnormal data point on the current frame torque value; the preset physiological threshold is determined based on the upper limit of the physiological change rate of human joint torque.

9. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that, It also includes observable outputs and traceability steps: The cross-domain credibility score generated by S3, the final mechanical parameter estimate or catch-all mark output by S4, and the fusion weight allocation result are written to the prediction process log file in real time and output to the outside world through a standardized data interface. The standardized data interface includes, but is not limited to, any one or more of the following: RESTful API interface, serial port debugging port, local log file, and human-machine interface display unit. The prediction process log contains a complete record chain of timestamps, output values ​​of each sub-model, fusion weights, credibility ratings, and whether a downgrade is triggered, which is used for quality traceability, third-party verification, or evidence preservation.

10. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The source population data and the gold standard joint torque data in the posture-mechanical reference database are both derived from the inverse dynamics calculation results collected by the combination of the three-dimensional motion capture system and the force table, and conform to the ISO / FDIS23471 biomechanical measurement standard. The gold standard data does not contain synthetic data.

11. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that: The prediction paradigm of the physical constraint embedding is preferably implemented using the Physical Information Neural Network (PINN). The loss function of the PINN embeds at least one of the following: Newton-Euler equation residual constraints, muscle physiological limit constraints, and joint range of motion constraints. Alternative implementations of the prediction paradigm of the physical constraint embedding include the inverse dynamics Transformer that integrates differentiable forward dynamics and the retrieval and inference model based on physical constraints.

12. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that, It also includes a gait cycle verification step: performing gait phase consistency verification on the final joint torque estimate or bottom value, detecting whether the torque abrupt change at the moment of switching between the support phase and the swing phase conforms to the physiological pattern; when an abnormal abrupt change is detected, cubic spline interpolation is used to correct it or it is marked as an abnormal frame; The physiological thresholds were determined based on statistics from a database of healthy adult gait patterns.

13. The cross-domain generalization enhancement method for predicting generalized mechanical parameters according to claim 1, applied to robot operating torque prediction scenarios, is characterized in that: The target to be predicted is the robot's end effector or joint drive unit; the kinematic sequence is the end effector pose and its time derivative sequence; the mechanical parameters are the operating contact torque and / or joint drive torque and / or end effector interaction force; The source domain data originates from at least one of the following: diverse operation task annotation data generated by a physics simulation engine, operation datasets collected by real robots equipped with force / torque sensors, and ground-truth output from a digital twin simulation platform; the physical constraints in the robot scenario preferentially adopt a combination of contact force feasibility constraints, joint torque boundary constraints, and operation space collision avoidance constraints; the domain adaptation targets at least one of the following drift scenarios: Sim-to-Real migration, operation generalization between different object categories, and visual perception drift under different ambient lighting and texture conditions.

14. The method for predicting generalized mechanical parameters with cross-domain generalization enhancement according to claim 1, characterized in that, It also includes a world model prediction and verification step: before S4 outputs the final estimate, it calls the pre-trained world action model WAM or other operation strategy models with forward state prediction capabilities to make forward predictions of the future multi-frame states of the current operation intention; Based on the predicted sequence, potential collision risks, boundary crossing risks, and contact anomaly risks are detected; if a high-risk event is detected, the cross-domain credibility score is multiplied by a safety discount factor of less than 1 and then re-entered into the S4 fusion decision process.

15. A generalized mechanical parameter prediction system with cross-domain generalization enhancement, used to implement the method according to any one of claims 1-14, characterized in that, include: The kinematic sequence acquisition module is used to acquire the human kinematic sequence of the target to be predicted; The heterogeneous domain adaptation prediction module, which includes a first sub-model and a second sub-model, is used to output at least two sets of candidate joint torque prediction values ​​in parallel. The credibility metric module is used to generate cross-population credibility scores based on Monte Carlo simulations; The fusion output module is used to perform adaptive weighted fusion based on the cross-population credibility score, and trigger a downgraded prediction path to call the attitude-mechanical reference database for matching when the credibility is lower than the threshold.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-14.

Images