Physiological mechanism driving and data fusion method and system for sitting posture human body modeling

Abstract

The invention provides a physiological mechanism driving and data fusion method and system for sitting posture human body modeling, and relates to the technical field of digital twinning and biomechanics crossing. The method comprises the following steps: mapping human body macroscopic parameters into biomechanical attribute parameters; generating a basic three-dimensional human body grid based on the human body macroscopic parameters; generating a deformed three-dimensional human body grid in combination with the basic three-dimensional human body grid and the human body macroscopic parameters; solving an optimal attitude based on the vehicle geometric parameters, the task scene identifier and the biomechanical attribute parameters; finally, the deformed three-dimensional human body grid and the optimal posture are integrated, the high-fidelity sitting posture digital twinborn body is generated, end-to-end automatic modeling from macroscopic parameters to the digital twinborn body is achieved, and the core problems that the form and the posture are not matched, crowd coverage is insufficient and task adaptability is poor are solved.

Description

Technical Field

[0001] This invention relates to the field of digital twin and biomechanics interdisciplinary technology, and in particular to a physiological mechanism-driven and data fusion method and system for seated human body modeling. Background Technology

[0002] In the field of vehicle design and evaluation, digital human models are key tools for visual field analysis, accessibility assessment, and comfort prediction. However, existing technologies have significant limitations: First, posture prediction models and body shape models are usually independent, resulting in a mismatch between the generated virtual human form and posture, making it difficult to apply to detailed simulations such as seat contact pressure distribution; second, mainstream statistical body shape models are mostly built based on standing postures or data from younger populations, failing to accurately represent the unique body shapes of older individuals, especially those in supported sitting positions due to muscle relaxation and spinal morphology changes; finally, traditional posture models are mostly designed for single driving tasks, lacking adaptability to diverse operational scenarios. These shortcomings severely restrict the accuracy and value of digital models in high-end industrial applications. Existing solutions, such as marker-based motion capture systems, are poorly portable and prone to errors, while some parametric models ignore the biomechanical rationality of sitting postures.

[0003] Therefore, the industry urgently needs a solution that can automatically and accurately generate a sitting posture digital twin that matches specific driver physiological characteristics, vehicle geometry, and task scenarios. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, the first aspect of this invention proposes a physiological mechanism-driven and data fusion method for seated human body modeling, comprising: S1: Based on the input macroscopic human body parameters, the physiological mechanism-driven model is used to process the parameters and generate biomechanical property parameters including equivalent muscle activation threshold, spinal segment bending stiffness coefficient and biomechanical posture type probability vector. S2: Based on macroscopic human body parameters, a parametric morphological generation model is used to generate a basic three-dimensional human body mesh that matches the user's physiological characteristics. S3: Based on the basic 3D human body mesh and macroscopic human body parameters, a deformable 3D human body mesh integrating the nonlinear compression deformation characteristics of soft tissue under sitting posture is generated through processing by a soft tissue deformation simulation model. S4: Based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters, the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario is obtained through processing by a task adaptive posture prediction model. S5: Based on deformable 3D human body mesh and optimal posture, a high-fidelity seated digital twin is generated through system coupling model processing.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively solves the core problems of morphology-posture mismatch, incomplete population representation, and poor task adaptability mentioned in the background technology. First, the physiological mechanism-driven model generates biomechanical attribute parameters based on input macroscopic human body parameters, including equivalent muscle activation thresholds, spinal segment bending stiffness coefficients, and biomechanical posture type probability vectors. These parameters inherently reflect individual physiological characteristics, providing a personalized driving foundation for subsequent morphology generation and posture prediction, ensuring the model can adapt to users of different ages and body types, overcoming the shortcomings of traditional models in terms of insufficient population coverage. The parameterized morphology generation model generates a basic 3D human body mesh that matches the user's physiological characteristics based on macroscopic human body parameters, establishing an accurate initial morphology and laying the foundation for subsequent deformation simulation. The soft tissue deformation simulation model generates a deformable 3D human body mesh integrating the nonlinear compression deformation characteristics of soft tissue under sitting posture based on the basic 3D human body mesh and macroscopic human body parameters, accurately simulating the physical deformation of the buttocks and thighs under sitting posture, solving the problem that traditional models cannot realistically represent the supporting sitting posture. The task-adaptive posture prediction model obtains the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters. By comprehensively considering the vehicle environment and multiple task scenarios, it achieves adaptive optimization for diverse tasks, overcoming the shortcomings of traditional posture models designed for single tasks. Finally, the system coupling model generates a high-fidelity seated digital twin based on deformable 3D human body mesh and the optimal posture, achieving dynamic unity between form and posture, fundamentally eliminating the drawbacks of form-posture mismatch.

[0006] Throughout the process, each step is closely integrated: the biomechanical property parameters output by the physiological mechanism-driven model directly drive the personalized processing of the parametric morphology generation model and the task-adaptive posture prediction model, ensuring that both morphology and posture are based on the same physiological foundation; the parametric morphology generation model and the soft tissue deformation simulation model jointly construct the realistic morphology, the task-adaptive posture prediction model optimizes the posture, and the system coupling model finally integrates morphology and posture to form end-to-end automated modeling, which significantly improves the accuracy and practicality of seated human body modeling. Attached Figure Description

[0007] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0008] Figure 1The diagram shown is a flowchart of a physiological mechanism-driven and data fusion method for seated human body modeling provided by an embodiment of the present invention.

[0009] Figure 2 The diagram shown is a deep neural network structure diagram of a physiological mechanism-driven model provided in an embodiment of the present invention.

[0010] Figure 3 The diagram shown is a structural schematic of a physiological mechanism-driven and data fusion system for seated human body modeling provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0012] The specific embodiments of the present invention will be described below.

[0013] Example 1 like Figure 1 As shown, this invention proposes a physiological mechanism-driven and data fusion method for seated human body modeling, including: S1: Based on the input macroscopic human body parameters, the physiological mechanism-driven model is used to process the parameters and generate biomechanical property parameters including equivalent muscle activation threshold, spinal segment bending stiffness coefficient and biomechanical posture type probability vector. S2: Based on macroscopic human body parameters, a parametric morphological generation model is used to generate a basic three-dimensional human body mesh that matches the user's physiological characteristics. S3: Based on the basic 3D human body mesh and macroscopic human body parameters, a deformable 3D human body mesh integrating the nonlinear compression deformation characteristics of soft tissue under sitting posture is generated through processing by a soft tissue deformation simulation model. S4: Based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters, the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario is obtained through processing by a task adaptive posture prediction model. Among them, "vehicle geometry parameters" refers to the geometric dimensions of the vehicle's interior, including seat height, seat tilt angle, steering wheel position, pedal position, and instrument panel distance, which are used to define the constraints of the seating environment.

[0014] "Task scenario identifiers" refer to discrete variables representing different work scenarios. For example, 0 represents a driving task, 1 represents a rest task, and 2 represents a console operation task. Each identifier corresponds to a set of posture requirements (such as hand position and line of sight angle) for adaptive posture optimization.

[0015] S5: Based on deformable 3D human body mesh and optimal posture, a high-fidelity seated digital twin is generated through system coupling model processing.

[0016] Specifically, the physiological mechanism-driven and data fusion method for seated human body modeling achieves end-to-end automated modeling through five core steps. Each step plays a specific function in the technical process and is interconnected. The physiological mechanism-driven model processes input macroscopic human body parameters to generate biomechanical property parameters including equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors. These parameters extract intrinsic physiological characteristics from readily available macroscopic features through nonlinear mapping, providing a personalized driving foundation for subsequent modules. The parametric morphology generation model processes the same macroscopic human body parameters to generate a basic 3D human body mesh that matches the user's physiological characteristics. This mesh is constructed from large-scale seated posture data through statistical learning, ensuring that the initial morphology accurately reflects the individual's anatomical structure. The soft tissue deformation simulation model processes the basic 3D human body mesh and macroscopic human body parameters to generate a deformable 3D human body mesh that integrates the nonlinear compression deformation characteristics of soft tissue under seated posture. It simulates the realistic deformation behavior of the buttocks and thighs under load through physical simulation. The task-adaptive posture prediction model processes vehicle geometric parameters, task scene identifiers, and biomechanical attribute parameters to obtain the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scene. This step uses an optimization algorithm to uniformly consider discrete task selection and continuous joint adjustment. The system coupling model processes deformable 3D human body meshes and the optimal posture to generate a high-fidelity seated digital twin, achieving dynamic fusion and visualization output of form and posture.

[0017] The synergistic effect of these steps is reflected in data flow and functional dependencies. Biomechanical property parameters output by the physiological mechanism-driven model are directly input into the task-adaptive posture prediction model, ensuring that posture optimization is based on individual physiological constraints. Simultaneously, the parametric morphology generation model and the soft tissue deformation simulation model sequentially process morphology construction, adding physical realism to the base mesh. The optimal posture generated by the task-adaptive posture prediction model is then combined with the deformed mesh through a system coupling model to form the final digital twin. This pipelined processing eliminates the problem of morphology and posture being disconnected in traditional methods because all modules share the same set of physiological input data, and the outputs are dynamically integrated in the system coupling model. Technically, this method can automatically generate highly personalized sitting posture models, adapting to the needs of users of different ages and body types, overcoming the limitation of insufficient population coverage. By integrating task scenario optimization, it supports diverse task simulations, improving adaptability. The automation of the entire process reduces manual intervention, improving modeling efficiency and consistency. Physiological mechanism-driven models provide a scientific basis for morphology and posture by capturing the intrinsic relationship between macroscopic parameters and biomechanical properties; parametric morphology generation models utilize statistical priors to ensure morphological rationality; soft tissue deformation simulation models introduce physical precision; task-adaptive posture prediction models balance comfort and task requirements through multi-objective optimization; and system coupling models maintain the unity of morphology and posture through dynamic feedback. Ultimately, this method can provide more accurate comfort assessments and performance predictions in applications such as vehicle design, reducing reliance on physical prototypes.

[0018] This method achieves overall performance improvement through seamless integration between steps. The physiological mechanism-driven model lays a personalized foundation for subsequent modules, the parametric morphology generation model establishes accurate initial morphology, the soft tissue deformation simulation model enhances physical realism, the task-adaptive posture prediction model optimizes dynamic posture, and the system coupling model ensures the consistency of the final output. This integrated processing makes the digital twin not only visually realistic but also behaves in accordance with biomechanical principles, thus providing a reliable simulation tool in industrial design. The technical benefits are reflected in the end-to-end automated workflow, where users only need to input macroscopic parameters to obtain a high-fidelity model, significantly simplifying traditional multi-step manual modeling. Simultaneously, the model can dynamically adapt to different scenarios, improving practicality and applicability. In summary, this method, through multi-module collaboration, fundamentally solves the core challenges in seated human body modeling, providing an efficient solution for digital twin applications.

[0019] The following details the process of building the above model.

[0020] 1. The process of building a physiological mechanism-driven model.

[0021] The physiological mechanism-driven model is used to map macroscopic human parameters (such as sex, age, height, body mass index, and sitting-to-height ratio) to biomechanical attribute parameters (equivalent muscle activation threshold, spinal segment flexure stiffness coefficient, and biomechanical posture type probability vector). Its construction process is as follows: Data preparation: Collect a large amount of labeled data, including macroscopic parameters of the human body and corresponding biomechanical properties. Biomechanical properties can be obtained through biomechanical experiments, such as using electromyography (EMG) to measure muscle activation thresholds, spinal curvature tests to obtain stiffness coefficients, and posture analysis to obtain posture type probabilities.

[0022] Network architecture design: A deep feedforward neural network (DFFNN) is used. The number of nodes in the input layer corresponds to the number of macroscopic parameters (e.g., 5 parameters), which are standardized (e.g., Z-score normalization). The hidden layers include at least two fully connected layers. The first fully connected layer uses the ReLU activation function to introduce non-linearity, followed by a dropout layer for regularization (dropout rate set to 0.2-0.5). The second fully connected layer further extracts features. The number of nodes in the output layer corresponds to the number of biomechanical attribute parameters (e.g., 3 parameters), and a linear activation function is used for regression tasks.

[0023] Training process: Backpropagation algorithm and optimizer (such as Adam) are used, with mean squared error (MSE) as the loss function. Training data is split into training and validation sets, and hyperparameters (such as number of layers and nodes) are adjusted through cross-validation to prevent overfitting. After the model is trained, it can be deployed as a static module, receiving macroscopic parameters and directly outputting biomechanical property parameters.

[0024] 2. The process of building a parametric morphology generation model.

[0025] The parametric morphology generation model generates a basic 3D human body mesh based on the statistical shape model (SSM). Its construction process is as follows: Statistical shape model construction: Collect large-scale seated human 3D scan data (e.g., from public datasets or custom scans), and align and normalize the mesh. Use Principal Component Analysis (PCA) for dimensionality reduction to obtain the average mesh and principal components (eigenvectors), where the principal components represent the main directions of shape variation.

[0026] Training a Multiple Linear Regression Model: A multiple linear regression model is pre-trained, taking macroscopic human parameters (such as gender, age, height, and body mass index) as input and outputting the principal component coefficient vector of the SSM (Simplified Chinese Multidimensional Structural Model). Training data includes macroscopic parameters and their corresponding principal component coefficients (obtained by projecting onto the SSM space). The regression model uses least squares fitting to ensure the accuracy of coefficient predictions.

[0027] Mesh generation and post-processing: A personalized 3D mesh is reconstructed using a linear combination of the predicted principal component coefficients and the SSM. Topology checks and repairs are then performed, using mesh processing tools (such as MeshLab or CGAL libraries) to detect and repair non-manifold edges, holes, and self-intersections, ensuring the mesh meets simulation requirements.

[0028] 3. The process of building a simulation model for soft tissue deformation.

[0029] A soft tissue deformation simulation model is used to simulate the nonlinear compression deformation of soft tissues in the hip and thigh regions under a sitting posture. Its construction process is as follows: ① Finite element discretization: Based on the basic 3D human body mesh, tetrahedral elements are used to generate the hip and thigh regions using finite element mesh generation tools (such as Gmsh or Abaqus). The mesh density is adjusted according to the contact area to ensure computational accuracy.

[0030] ②Building a regression model for material parameters: The first regression model takes body mass index and body fat percentage as input and outputs the shear modulus parameters of a hyperelastic material model. Training data comes from biomechanical experiments, such as obtaining the soft tissue shear modulus of different individuals through indentation tests, and a linear regression model is used to establish the mapping.

[0031] The second regression model takes body mass index and body fat percentage as input and outputs a hardening index parameter. Similarly, the regression model is trained based on experimental data (such as uniaxial compression tests) to capture nonlinear hardening behavior. Both the first and second regression models employ simple linear regression or multinomial regression, the specific form determined according to the data distribution.

[0032] ③ Nonlinear finite element solution: Integrate personalized material parameters into the finite element solver (such as FEBio or Abaqus), set up a hyperelastic material model (such as the Neo-Hookean model), apply boundary conditions (such as fixing the seat contact surface), perform quasi-static analysis to solve the displacement field, and finally output the deformed three-dimensional mesh.

[0033] 4. The process of building a task-adaptive attitude prediction model.

[0034] The task-adaptive pose prediction model is used to generate the optimal pose that matches the user's physiological characteristics, vehicle environment, and task scenario. Its construction process is as follows: Lookup table construction: Based on biomechanical research literature or experimental data, predefine the mapping relationship between biomechanical posture types (such as "relaxed", "forward lean", "backward lean") and joint angles in neutral postures. For example, collect joint angle data for different postures using a motion capture system to construct a lookup table.

[0035] Mixed-integer programming model design: The objective function includes a posture deviation penalty term (quantifying the difference between the current posture and the neutral posture, such as the Euclidean distance between joint angles) and a muscle activation penalty term (estimating muscle effort based on biomechanical property parameters, such as using a muscle model). Constraints include joint range of motion and vehicle geometry limitations (such as seat position and steering wheel distance). The model is formalized using optimization modeling tools (such as Pyomo or Gurobi).

[0036] Optimize solver configuration: Employ branch-and-bound algorithms to handle integer variables (such as task selection), and use parallel computing architectures (such as GPU acceleration) to evaluate candidate solutions. The optimal solution selection strategy is based on minimizing the objective function value, ensuring solution efficiency and global optimality.

[0037] 5. The process of building the system coupling model.

[0038] The system coupling model is used to integrate deformable 3D human body meshes and optimal poses to generate a high-fidelity seated digital twin. Its construction process is as follows: Skeletal skinning technology is implemented by binding a skeletal system (e.g., using joint point definitions) to a deformable 3D mesh, and using a linear blend skinning (LBS) algorithm to convert joint rotations of the optimal pose into mesh vertex displacements. Weights are determined either manually drawn or automatically calculated (e.g., using thermal diffusion).

[0039] Real-time rendering: Import the pose-adjusted mesh using a physics engine (such as Unity or Unreal Engine), configure lighting, materials, and textures, and achieve visual rendering. The engine supports real-time interaction and viewpoint switching.

[0040] Dynamic feedback mechanism: By listening for attitude change events, soft tissue deformation is recalculated (e.g., by calling the finite element solver), updating the mesh shape. The feedback loop ensures that the digital twin maintains physical consistency during attitude changes.

[0041] In some implementations, S1 includes: S1.1: Based on macroscopic human body parameters including gender, age, height, body mass index, and sitting height-to-height ratio, standardized input features are generated through the input layer of a deep feedforward neural network model. S1.2: Based on standardized input features, a high-level feature representation is generated by performing nonlinear transformation processing through multiple hidden layers of a deep feedforward neural network model. S1.3: Based on high-level feature representation, biomechanical property parameters are generated by processing the output layer of a deep feedforward neural network model.

[0042] In the physiological mechanism-driven and data fusion approach, step S1 refines the data using a deep feedforward neural network model, achieving a high-precision conversion from macroscopic parameters to biomechanical properties. The input layer of the deep feedforward neural network model processes macroscopic human parameters including gender, age, height, body mass index, and sitting-to-height ratio to generate standardized input features. This process ensures data consistency through feature scaling and normalization, providing a stable foundation for subsequent network computations. Multiple hidden layers of the deep feedforward neural network model perform nonlinear transformations on the standardized input features to generate high-level feature representations. These hidden layers introduce nonlinear relationships through activation functions, gradually abstracting the complex interactive effects in the input parameters, thereby capturing the influence of age and body composition on biomechanical properties. The output layer of the deep feedforward neural network model processes these high-level feature representations to generate biomechanical property parameters, including equivalent muscle activation thresholds, spinal segment flexural stiffness coefficients, and biomechanical posture type probability vectors. The output layer uses a linear activation function to directly regress the target value, ensuring the continuity and physical meaning of the prediction results.

[0043] This hierarchical approach enables the network to learn complex patterns from simple features. The input layer processes raw parameters to reduce noise, the hidden layers extract key features through multiple transformations, and the output layer ultimately maps to biomechanical properties. Technically, this design improves prediction accuracy and generalization ability because standardized input features avoid training instability caused by differences in data scale, the nonlinear transformations of multiple hidden layers capture the deep correlation between macroscopic parameters and biomechanical properties, and the direct regression of the output layer ensures parameter interpretability. The standardized processing of the input layer ensures the convergence of network training, the nonlinear activation function of the hidden layers simulates complex relationships in biomechanics, and the linear output of the output layer conforms to the numerical characteristics of biomechanical parameters. The entire network structure learns end-to-end, avoiding the shortcomings of manual feature engineering in traditional methods, allowing the model to automatically mine relevant patterns from the data. Furthermore, this approach enhances the model's adaptability to diverse populations because the input parameters cover key physiological dimensions; by learning the combined effects of these dimensions, the network can accurately predict the biomechanical characteristics of different individuals.

[0044] The application of deep feedforward neural network models makes physiological mechanism-driven approaches more reliable. The input layer ensures data consistency, the hidden layers enhance feature representation through multi-layered processing, and the output layer directly generates practical parameters. This approach not only simplifies the data preprocessing process but also improves the model's robustness, enabling it to maintain stable predictions even in the face of noise or missing data. The technical benefits are reflected in the more accurate generation of biomechanical property parameters; the equivalent muscle activation threshold reflects individual differences in muscle function; the spinal segment flexure stiffness coefficient quantifies the spinal mechanical properties; and the biomechanical posture type probability vector provides a classification basis for posture preferences. In principle, the network learns a mapping from macroscopic to microscopic levels through training, utilizing statistical regularities in big data, thus performing well even with unknown data. Ultimately, this step provides a solid driving core for the entire method, ensuring that subsequent morphology and posture modules are based on scientifically sound physiological inputs, improving the realism and practicality of the digital twin. Through this automated processing, the method reduces its dependence on external biomechanical measurements, making the modeling process more efficient and scalable.

[0045] refer to Figure 2 , Figure 2 This diagram illustrates a deep feedforward neural network structure, comprising an input layer, two hidden layers (i.e., fully connected layers), and an output layer. The input layer receives standardized macroscopic human parameters (such as gender, age, height, body mass index, and sitting-height ratio), with the number of nodes corresponding to the number of parameters (e.g., 5 nodes). The fully connected layers, directly connected to the input layer, process the input features using a non-linear activation function (such as ReLU) to generate preliminary feature representations. The fully connected layers, directly connected to the output layer, further refine the features to generate higher-level feature representations. The output layer directly regresses biomechanical property parameters, including equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors; the number of nodes corresponds to the number of output parameters (e.g., 3 nodes). Figure 2 The middle arrow indicates the data flow direction, from the input layer through the fully connected layer to the output layer, reflecting the nonlinear transformation process of features and ensuring end-to-end mapping from macroscopic human body parameters to biomechanical properties.

[0046] In some implementations, S1.2 includes: S1.2.1: Based on the standardized input features, the first hidden layer features are generated by processing them through the first fully connected layer of the deep feedforward neural network model; S1.2.2: Based on the features of the first hidden layer, regularization is performed by dropping the layer to generate hidden layer features that are resistant to overfitting; S1.2.3: Based on the hidden layer features that resist overfitting, high-level feature representations are generated by processing them through the second fully connected layer of the deep feedforward neural network model.

[0047] In the refinement of step S1, the deep feedforward neural network model further optimizes the feature extraction process through specific hidden layer processing, ensuring the stability and accuracy of biomechanical attribute prediction. The first fully connected layer of the deep feedforward neural network model processes standardized input features to generate the first hidden layer features. This layer uses multiple neurons and activation functions to perform initial feature transformation, capturing the linear and nonlinear relationships between input parameters and providing a rich feature base for subsequent processing. The dropout layer performs regularization processing based on the first hidden layer features, generating overfitting-resistant hidden layer features. This layer randomly ignores some neuron connections during training, reducing the model's over-reliance on training data, thereby improving generalization ability and preventing performance degradation on unknown data. The second fully connected layer of the deep feedforward neural network model processes overfitting-resistant hidden layer features to generate high-level feature representations. This layer refines and combines features through further nonlinear transformations to form a more abstract representative vector, ultimately serving the parameter generation of the output layer.

[0048] This sequential processing ensures a gradual deepening of feature learning. The first fully connected layer extracts basic features, the dropout layer introduces regularization to control complexity, and the second fully connected layer completes advanced feature integration. Technically, this structure improves the model's robustness and generalization performance because the first fully connected layer broadens the feature space, capturing more latent patterns; the dropout layer reduces overfitting risk through randomness, making the model more adaptable to diverse data distributions; and the second fully connected layer strengthens the interaction between features, ensuring that the output feature representation comprehensively reflects biomechanical properties. The width design of the first fully connected layer allows the network to learn multiple combinations of input parameters, the regularization mechanism of the dropout layer simulates the variability of biomechanical data, and the depth processing of the second fully connected layer corresponds to high-level causal relationships in biomechanical properties. The entire process, through hierarchical regularization and feature optimization, ensures the network remains stable in complex data environments, avoiding the shortcomings of traditional single-layer models.

[0049] This processing makes biomechanical property predictions more reliable. The first fully connected layer ensures the diversity of initial features, the discard layer maintains model simplicity, and the second fully connected layer extracts key information. The technical effects are reflected in the fact that the generated high-level feature representations accurately drive subsequent outputs, the prediction of equivalent muscle activation thresholds more closely matches individual physiological states, the estimation of spinal segment flexural stiffness coefficients more consistent with mechanical principles, and the classification of biomechanical posture type probability vectors more accurate. In principle, the multi-layered processing simulates the hierarchical structure of the human physiological system, thus incorporating more biomechanical priors into the prediction. Ultimately, this refinement step enhances the efficiency of the entire physiological mechanism-driven module, providing a more consistent physiological basis for the sitting posture digital twin and supporting more accurate comfort and performance evaluation in practical applications. By avoiding overfitting, the model maintains high performance even with limited or noisy data, expanding the applicability and practicality of the method.

[0050] In some implementations, the parametric morphology generation model includes a statistical shape model, and S2 includes: S2.1: Based on macroscopic human body parameters, the principal component coefficient vector of the statistical shape model is generated by processing the data through a pre-trained multiple linear regression model. S2.2: Based on the principal component coefficient vector, a personalized basic 3D human body mesh is generated through linear combination processing using a statistical shape model. S2.3: Perform topology checks and repairs on the personalized basic 3D human body mesh to generate a basic 3D human body mesh that meets the requirements of the manifold structure.

[0051] In the parametric morphology generation model, the application of the statistical shape model enables efficient conversion from macroscopic parameters to 3D meshes, ensuring that the generated basic human body mesh conforms to individual characteristics and possesses anatomical rationality. The parametric morphology generation model includes a statistical shape model, built upon large-scale seated posture scan data. Principal component analysis (PCA) compresses the high-dimensional shape space into a low-dimensional representation, thereby capturing morphological variation patterns within the population. In step S2, a pre-trained multiple linear regression model processes macroscopic human body parameters to generate the principal component coefficient vector of the statistical shape model. This process maps macroscopic parameters to weights in the shape space through regression analysis, allowing individual physiological characteristics to directly guide mesh generation. The statistical shape model performs linear combination processing based on the principal component coefficient vector to generate a personalized basic 3D human body mesh. This mesh is synthesized using a weighted average shape and principal component displacement field, ensuring that the output is consistent with the training data at the vertex level. Topology checks and repairs are performed on the personalized basic 3D human body mesh to generate a basic 3D human body mesh that meets manifold structure requirements. This step uses algorithms to detect and correct mesh defects such as holes or self-intersections, ensuring the integrity and usability of the mesh.

[0052] This process automates and accurately generates morphology. The statistical shape model provides shape priors, multiple linear regression converts parameters to coefficients, linear combination generates the mesh, and topology repair ensures quality. Technically, this method can quickly generate meshes that match the user's physiological characteristics because the statistical shape model, based on real sitting posture data, accurately reflects the geometric characteristics under supported posture. The multiple linear regression model utilizes the statistical correlation between macroscopic parameters and shape, simplifying the mapping process. Linear combination processing maintains the smoothness and continuity of the mesh. Topology checking and repair eliminate potential errors and improve the stability of the mesh in simulation. Principal component analysis of the statistical shape model reduces data dimensionality, making the generation process efficient. The multiple linear regression model learns the linear relationship between macroscopic parameters and shape coefficients, ensuring prediction accuracy. Linear combination uses orthogonal basis vectors to maintain shape rationality. Topology processing ensures mesh usability through computational geometry methods. The entire process combines statistical learning and geometric processing, thus achieving a balance between generation speed and quality.

[0053] The parametric morphology generation model achieves high personalization through statistical shape models, while a pre-trained multiple linear regression model ensures reliable coefficient predictions. Linear combination processing generates a visual mesh, and topology repair enhances practicality. The technical benefits are reflected in the accurate representation of individual sitting posture morphology by the basic 3D human body mesh, including spinal curvature and limb proportions, providing a realistic basis for subsequent soft tissue deformation. Simultaneously, automated processing reduces manual modeling time and improves overall efficiency. In principle, this method utilizes the statistical regularities of large-scale data, ensuring that the generated mesh is not only visually realistic but also conforms to biomechanical constraints. Ultimately, this step provides reliable morphological input for the entire digital twin process, supporting more accurate seat interaction simulation and comfort analysis, and reducing reliance on physical scanning in industrial design. Integrated topology checking ensures mesh quality, preventing computational errors in subsequent simulations, thereby improving the overall robustness and application value of the method.

[0054] The process of constructing the multiple linear regression model is as follows.

[0055] The multiple linear regression model is a component of the parametric morphological generative model (SSM) and is used to predict the principal component coefficients of the SSM. Its construction process is as follows: Data preparation: Collect macroscopic parameters of the human body and the corresponding principal component coefficients of the SSM (obtained by projecting 3D scan data into the SSM space).

[0056] Model training: A multiple linear regression algorithm is used, with the form of coefficient = W × human macroscopic parameters + b, where W is the weight matrix and b is the bias vector. The parameters are optimized using the least squares method or gradient descent.

[0057] Deployment: After the model is trained, it is integrated into the system, receiving macroscopic parameters and directly outputting principal component coefficients.

[0058] In some implementations, S3 includes: S3.1: Based on the basic 3D human body mesh, tetrahedral element meshes for the buttocks and thigh regions are generated through finite element discretization. S3.2: Based on macroscopic human body parameters, the personalized material parameters of the hyperelastic material model are determined through material parameter regression model. S3.3: Based on tetrahedral element meshes and personalized material parameters, a deformable three-dimensional human body mesh is generated through quasi-static analysis using a nonlinear finite element solver.

[0059] In the soft tissue deformation simulation model, step S3 simulates the physical response of soft tissue under sitting posture using finite element analysis, ensuring that the deformable mesh accurately reflects the individual's morphological changes under load. Finite element discretization is performed based on the basic 3D human body mesh, generating tetrahedral element meshes for the hip and thigh regions. This process divides continuous geometry into discrete elements, facilitating numerical calculations, and refines the mesh for the sitting contact area to capture local deformation. The material parameter regression model is based on macroscopic human body parameters, determining personalized material parameters for the hyperelastic material model. This model maps physiological characteristics such as body mass index and body fat percentage to material constants through regression analysis, allowing mechanical properties to dynamically adjust with individual differences. The nonlinear finite element solver performs quasi-static analysis based on the tetrahedral element mesh and personalized material parameters, generating a deformable 3D human body mesh. The solver calculates the displacement field under equilibrium conditions using an iterative algorithm and superimposes the results onto the basic mesh, outputting the final form integrating deformation characteristics.

[0060] This approach achieves physically accurate simulation of soft tissue deformation. Finite element discretization provides the computational framework, material parameter regression reverts to individualized mechanical properties, and a nonlinear solver handles large deformation behavior. Technically, this method can accurately predict the compression deformation of the hips and thighs in a seated position because finite element discretization captures geometric details, material parameter regression ensures that mechanical properties are consistent with individual physiology, and the nonlinear solver handles the hyperelastic response of soft tissue. Finite element discretization transforms a continuous problem into a discrete system through mesh generation, making simulation feasible; material parameter regression utilizes the correlation between physiological data and mechanical tests, making the model parameters realistic and reliable; and the nonlinear solver solves the equilibrium equations, simulating the deformation process under quasi-static loading. The entire process combines computational mechanics and physiological data, thus incorporating individualized characteristics into the simulation.

[0061] The soft tissue deformation simulation model enhances the realism of the digital twin through finite element analysis. Finite element discretization ensures the mesh is suitable for simulation, and a material parameter regression model introduces personalized material properties. A nonlinear finite element solver calculates the deformation results. The technical benefits are reflected in the fact that the deformable 3D human body mesh can accurately simulate the contact pressure distribution between the seat and the human body, providing a direct basis for comfort assessment. Simultaneously, personalized material parameters consider the influence of body fat and weight, making the deformation more consistent with biomechanical principles. In principle, this method describes the nonlinear behavior of soft tissue based on hyperelastic constitutive relations, thus reproducing real physical phenomena in the simulation. Ultimately, this step provides the system with high-fidelity morphological output, supporting more reliable engineering design decisions and reducing the need for experimental measurements in applications. By integrating quasi-static analysis, the deformation simulation controls computational costs while maintaining accuracy, making the method applicable to real-time or near-real-time scenarios, improving overall practicality and efficiency.

[0062] In some implementations, S3.2 includes: S3.2.1: Based on the body mass index and body fat percentage in the macroscopic parameters of the human body, the shear modulus parameter in the hyperelastic material model is generated by processing through the first regression model; S3.2.2: Based on the body mass index and body fat percentage in the macroscopic parameters of the human body, the hardening index parameter in the hyperelastic material model is generated by processing through the second regression model; S3.2.3: Integrate shear modulus parameters and hardening index parameters to generate personalized material parameters.

[0063] In determining the personalized material parameters for the hyperelastic material model, step S3.2 uses two independent regression models to handle key material constants, ensuring a precise correlation between soft tissue mechanical properties and individual physiological characteristics. The first regression model processes body mass index and body fat percentage from macroscopic human parameters to generate the shear modulus parameter in the hyperelastic material model. This model establishes a mathematical relationship between physiological indicators and the material shear modulus through multiple linear regression, allowing the initial stiffness of soft tissue to adaptively adjust with changes in body composition. The second regression model, also based on body mass index and body fat percentage, generates the hardening index parameter in the hyperelastic material model. This model also employs regression analysis to specifically capture the influence of body fat distribution on the nonlinear hardening behavior of the material, thus accurately characterizing the stress response characteristics of soft tissue under large deformations. Integrating the shear modulus and hardening index parameters to generate personalized material parameters combines the outputs of the two regression models into a complete set of material parameters, which is directly used for defining constitutive relations in subsequent finite element analysis.

[0064] This separate approach allows for specialized modeling of different mechanical properties of the material. The first regression model focuses on the shear modulus, which controls small deformation behavior, while the second regression model describes the hardening characteristics during large deformation. The integration process ensures the integrity and consistency of the parameter set. Technically, this method can more accurately simulate the soft tissue deformation behavior of individuals with different body types because the shear modulus parameter is directly related to the initial hardness of the soft tissue, while the hardening index parameter determines the stiffness change during deformation. The combination of these two parameters allows the hyperelastic model to comprehensively reflect the mechanical response throughout the entire process from mild compression to large deformation. The first regression model utilizes the known physiological correlation between body mass index and body fat percentage and tissue stiffness, while the second regression model is based on the influence mechanism of adipose tissue distribution on strain stiffening effects. The integration step ensures the physical consistency between the parameters. Through this hierarchical regression strategy, material parameters are not only correlated with individual characteristics but also maintain the inherent physical consistency of the hyperelastic constitutive relation.

[0065] The generation of personalized material parameters makes soft tissue deformation simulation more realistic and reliable. The first regression model accurately predicts the basic stiffness characteristics, the second regression model captures nonlinear hardening behavior, and parameter integration ensures the integrity of the constitutive model. This process significantly improves the accuracy of predicting sitting pressure distribution because the material parameters are based entirely on individual physiological characteristics rather than population averages, allowing the simulation to distinguish the mechanical differences caused by different body fat distributions. In principle, this method establishes a direct mapping between physiological parameters and continuous medium mechanical parameters, transforming macroscopic anthropometric data into microscopic mechanical properties, achieving cross-scale modeling. Ultimately, this step provides highly personalized material input for finite element analysis, enabling the deformable 3D human body mesh to realistically reflect the soft tissue response of a specific individual in a sitting posture, providing a more reliable simulation basis for seat comfort assessment and ergonomic design. Through this refined processing, the consistency between simulation results and real human behavior is significantly improved, effectively supporting decision optimization in the product design process.

[0066] In some implementations, S4 includes: S4.1: Based on the biomechanical posture type probability vector in the biomechanical attribute parameters, the individualized neutral posture joint angles are determined by processing through a predefined lookup table. S4.2: Based on individualized neutral posture joint angles, vehicle geometric parameters, and task scenario identifiers, a mixed integer programming model is used to construct and process the objective function, which includes posture deviation penalty term and muscle activation penalty term. S4.3: Based on the objective function and constraints, the optimal pose is obtained by processing the data through an optimization solver.

[0067] In the task-adaptive posture prediction process, step S4 systematically combines biomechanical attributes with task requirements to achieve highly personalized posture optimization. The biomechanical posture type probability vector is processed based on a predefined lookup table to determine individualized neutral posture joint angles. This lookup table establishes a correspondence between posture types and ideal joint angles through biomechanical research, allowing the dominant type in the probability vector to be directly mapped to specific joint angle configurations, providing a personalized comfort benchmark for posture optimization. A mixed-integer programming model is constructed based on individualized neutral posture joint angles, vehicle geometry parameters, and task scenario identifiers. It establishes an objective function containing posture deviation penalty terms and muscle activation penalty terms. The posture deviation penalty term quantifies the difference between the current posture and the individualized neutral posture, while the muscle activation penalty term estimates the muscle effort required to maintain the posture based on biomechanical attribute parameters. Both are weighted to balance comfort and energy consumption requirements.

[0068] This construction process formalizes the posture optimization problem into a structured mathematical programming problem. The lookup table transforms discrete biomechanical classifications into continuous joint angle references, while the mixed-integer programming model unifies continuous posture variables and discrete task selection within the same optimization framework. Technically, this method generates optimal postures that conform to both individual physiological characteristics and specific task requirements. This is because individualized neutral posture joint angles provide a comfort benchmark based on biomechanical principles, posture deviation penalties drive the optimization result closer to this benchmark, and muscle activation penalties minimize muscle load. The comprehensive consideration of the objective function ensures a balance between comfort, functionality, and efficiency in the posture. The lookup table, built upon extensive biomechanical experimental data, ensures the scientific validity of the neutral posture angles; the mixed-integer programming model, through mathematical formalization, transforms the complex multi-objective optimization problem into a solvable standard form, enabling the algorithm to systematically explore the solution space.

[0069] Task-adaptive posture prediction achieves accurate personalized posture generation through this structured approach. A lookup table provides an individualized baseline, a mixed-integer programming model constructs the complete optimization problem, and the objective function balances multiple optimization objectives. This approach significantly enhances the practicality of digital twins in diverse scenarios because the optimized posture respects individual physiological preferences, meets the functional requirements of specific tasks, and considers the spatial constraints within the vehicle. In principle, this method encodes biomechanical knowledge into the optimization problem, giving the algorithm's decision-making process clear physical meaning and a physiological basis. Ultimately, this step provides a highly adaptive posture output for the seated digital twin, supporting accurate comfort prediction and performance evaluation. During product design, it effectively simulates the natural posture responses of different users in various operating scenarios. Through this systematic optimization framework, posture prediction no longer relies on empirical rules but is based on scientific optimization principles, significantly improving the reliability and consistency of prediction results.

[0070] In some implementations, S4.3 includes: S4.3.1: Based on the objective function and constraints, a branch and bound algorithm is used to generate a candidate solution space; S4.3.2: Based on the candidate solution space, a parallel computing architecture is used to accelerate the evaluation of multiple possible solutions; S4.3.3: Based on the accelerated evaluation results, the optimal solution selection strategy is adopted to obtain the optimal pose.

[0071] In solving the mixed-integer programming problem, step S4.3 achieves efficient computation through an advanced combination of algorithms, ensuring rapid location of the optimal pose within the complex solution space. The branch-and-bound algorithm, based on the objective function and constraints, generates a candidate solution space. This algorithm systematically enumerates possible combinations of integer solutions and calculates their boundary values, gradually eliminating search regions where an optimal solution is impossible, thus decomposing the complex mixed-integer programming problem into a series of manageable subproblems. The parallel computing architecture, based on the candidate solution space, accelerates the evaluation of multiple possible solutions. This architecture utilizes the many-core computing power of the graphics processing unit to process a large number of candidate solutions simultaneously, significantly shortening the computation time for each iteration by parallel computation of objective function values ​​and constraint satisfaction. The optimal solution selection strategy, based on the accelerated evaluation results, obtains the optimal pose. This strategy compares the objective function values ​​of all evaluated solutions and selects the pose configuration with the best overall performance as the final output, ensuring the quality and feasibility of the solution.

[0072] This hierarchical solution strategy fully leverages the advantages of different computing technologies. The branch-and-bound algorithm provides a rigorous mathematical programming framework, the parallel computing architecture addresses computational bottlenecks, and the optimal solution selection strategy ensures solution quality. In terms of technical effectiveness, this method can solve high-dimensional pose optimization problems within a reasonable timeframe. This is because the branch-and-bound algorithm reduces the number of solutions to be evaluated through intelligent pruning, the parallel computing architecture significantly improves the evaluation speed of remaining solutions, and the optimal solution selection strategy ensures the global optimality of the final result. The branch-and-bound algorithm utilizes the mathematical properties of integer programming to build a search tree, gradually converging to the optimal solution; the parallel computing architecture distributes the computational load across multiple processing units through data parallelism; and the optimal solution selection strategy makes decisions based on the optimality conditions in optimization theory. The entire solution process effectively controls computational complexity while ensuring solution quality.

[0073] The efficient solution mechanism enables task-adaptive attitude prediction to meet the real-time requirements of practical applications. The branch-and-bound algorithm ensures the completeness of the search, the parallel computing architecture provides necessary computational acceleration, and the optimal solution selection strategy guarantees the optimality of the solution. This approach makes complex multi-objective attitude optimization feasible in engineering practice because the algorithm can complete calculations within seconds, supporting interactive applications and rapid iterations in the design process. In principle, this method achieves a good balance between solution quality and computational cost by combining the high reliability of accurate algorithms with the high efficiency of parallel computing. Ultimately, this step provides the entire digital twin system with fast and reliable attitude decision-making capabilities, enabling the system to respond promptly to different task scenarios and user needs, playing a vital role in real-world engineering environments. Through this optimized solution implementation, the attitude prediction module is no longer a theoretical concept but a practical engineering tool, significantly improving the overall performance and usability of the digital twin system.

[0074] In some implementations, S5 includes: S5.1: Based on the deformable 3D human body mesh and the optimal pose, the 3D human body mesh with the adjusted pose is generated through skeletal skinning technology. S5.2: Based on the 3D human body mesh after posture adjustment, a visual digital model of sitting posture is generated through real-time rendering processing using a physics engine. S5.3: Based on a visualized digital model of sitting posture, a high-fidelity digital twin of sitting posture is generated by processing it through a dynamic feedback mechanism, reflecting the soft tissue deformation state after posture changes.

[0075] In the system coupling process, step S5 achieves deep fusion of form and posture through multi-layered technology integration, completing the final construction from components to a complete digital twin. The skeletal skinning technology processes deformable 3D human meshes and optimal postures to generate a posture-adjusted 3D human mesh. This technology transfers joint transformations of the skeletal system to the vertices of the attached soft tissue mesh, allowing the base mesh to naturally deform following skeletal movement while maintaining existing soft tissue deformation characteristics. The physics engine performs real-time rendering based on the posture-adjusted 3D human mesh, generating a visualized seated digital model. The engine uses lighting calculations, material rendering, and scene management techniques to transform the geometric model into a visually realistic 3D image, providing interactive observation and analysis capabilities. The dynamic feedback mechanism processes the visualized seated digital model to generate a high-fidelity seated digital twin reflecting the soft tissue deformation state after posture changes. This mechanism monitors the impact of posture changes on the contact state, triggering necessary recalculations to update soft tissue deformation, ensuring that the mechanical response remains consistent with the current posture.

[0076] This sequential processing establishes a complete workflow from data to visualization to feedback. Skeletal skinning technology achieves geometric pose adaptation, the physics engine provides visual representation, and the dynamic feedback mechanism maintains physical realism. Technically, this method outputs a digital twin that is both visually realistic and physically accurate. Skeletal skinning technology maintains the continuity and natural appearance of the mesh, the physics engine provides high-quality visual performance, and the dynamic feedback mechanism ensures real-time consistency of mechanical behavior. Skeletal skinning technology converts joint rotation into vertex displacement through a linear hybrid skinning algorithm, the physics engine utilizes computer graphics principles to achieve realistic rendering, and the dynamic feedback mechanism establishes a causal relationship between pose and deformation based on physical simulation principles. The entire processing flow forms a complete modeling closed loop, enabling the digital twin to dynamically respond to changes in its internal state.

[0077] This integrated approach to system coupling achieves the final presentation of the digital twin. Skeletal skinning technology handles geometric deformation, the physics engine provides visualization, and a dynamic feedback mechanism ensures physical accuracy. This integration gives the generated seated digital twin not only static realism but also dynamic responsiveness. Skeletal skinning ensures natural posture changes, the physics engine provides an intuitive visualization interface, and dynamic feedback maintains the model's physical consistency across different states. In principle, this method combines computer graphics, computational mechanics, and real-time simulation techniques to create a digital counterpart that faithfully reflects real human behavior. Ultimately, this step provides usable outputs for the entire modeling process, enabling designers and engineers to intuitively evaluate the effects of different design schemes and support simulation-based design decisions and optimization processes. Through this end-to-end coupling, the digital twin is no longer an isolated collection of components but becomes an organic whole system, playing a greater role in practical applications.

[0078] Example 2 like Figure 3 As shown, in a second aspect, the present invention proposes a physiological mechanism-driven and data fusion system for seated human body modeling. The system employs the method provided in any of the above embodiments and includes: The physiological mechanism-driven module is used to generate biomechanical property parameters, including equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors, based on input macroscopic human body parameters. The parametric morphology generation module is used to generate a basic 3D human body mesh that matches the user's physiological characteristics based on macroscopic human body parameters. The soft tissue deformation simulation module is used to generate deformable 3D human body meshes that integrate the nonlinear compression deformation characteristics of soft tissues in a sitting posture, based on the basic 3D human body mesh and macroscopic human body parameters. The task-adaptive attitude prediction module is used to obtain the optimal attitude that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters. The system coupling module is used to generate a high-fidelity seated digital twin based on deformable 3D human body mesh and optimal posture.

[0079] This system corresponds to the method provided in Embodiment 1 above, and will not be described in detail here.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A physiological mechanism-driven and data fusion method for seated human body modeling, characterized in that, include: Based on the input macroscopic human body parameters, the model is processed through physiological mechanisms to generate biomechanical property parameters that include equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors. Based on macroscopic human body parameters, a parametric morphological generation model is used to generate a basic three-dimensional human body mesh that matches the user's physiological characteristics. Based on the basic 3D human body mesh and macroscopic human body parameters, a deformable 3D human body mesh integrating the nonlinear compression deformation characteristics of soft tissue under sitting posture is generated through processing by a soft tissue deformation simulation model. Based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters, the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario is obtained through processing by a task adaptive posture prediction model. Based on deformable 3D human body mesh and optimal posture, a high-fidelity seated digital twin is generated through system coupling model processing.

2. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 1, characterized in that, Based on the input macroscopic human body parameters, a physiological mechanism-driven model is used to process the data and generate biomechanical property parameters, including equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors. Based on macroscopic human body parameters including gender, age, height, body mass index, and sitting height-to-height ratio, standardized input features are generated through the input layer of a deep feedforward neural network model. Based on standardized input features, a high-level feature representation is generated by nonlinear transformation through multiple hidden layers of a deep feedforward neural network model. Based on high-level feature representation, biomechanical property parameters are generated by processing the output layer of a deep feedforward neural network model.

3. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 2, characterized in that, Based on standardized input features, a high-level feature representation is generated by performing nonlinear transformations through multiple hidden layers of a deep feedforward neural network model, including: Based on standardized input features, the first hidden layer features are generated by processing them through the first fully connected layer of the deep feedforward neural network model. Based on the features of the first hidden layer, regularization is performed by dropping the layer to generate hidden layer features that are resistant to overfitting. Based on the hidden layer features that resist overfitting, high-level feature representations are generated by processing them through the second fully connected layer of a deep feedforward neural network model.

4. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 1, characterized in that, Parametric morphology generation models include statistical shape models, which are based on macroscopic human body parameters and processed through parametric morphology generation models to generate a basic 3D human body mesh that matches the user's physiological characteristics, including: Based on macroscopic human body parameters, a pre-trained multiple linear regression model is used to process the data and generate the principal component coefficient vector of the statistical shape model. Based on the principal component coefficient vector, a personalized basic 3D human body mesh is generated by linear combination processing through a statistical shape model. The personalized basic 3D human body mesh is subjected to topology inspection and repair processing to generate a basic 3D human body mesh that meets the requirements of manifold structure.

5. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 1, characterized in that, Based on a basic 3D human body mesh and macroscopic human body parameters, a deformable 3D human body mesh integrating the nonlinear compression deformation characteristics of soft tissue under sitting posture is generated through soft tissue deformation simulation model processing, including: Based on the basic 3D human body mesh, tetrahedral element meshes for the buttocks and thigh regions are generated through finite element discretization. Based on macroscopic human body parameters, personalized material parameters for the hyperelastic material model are determined through material parameter regression model processing. Based on tetrahedral element meshes and personalized material parameters, a deformable three-dimensional human body mesh is generated through quasi-static analysis using a nonlinear finite element solver.

6. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 5, characterized in that, Based on macroscopic human body parameters, personalized material parameters for the hyperelastic material model are determined through material parameter regression modeling, including: Based on the body mass index and body fat percentage in the macroscopic parameters of the human body, the shear modulus parameter in the hyperelastic material model is generated by processing through the first regression model. Based on the body mass index and body fat percentage in the macroscopic parameters of the human body, the hardening index parameter in the hyperelastic material model is generated by processing through a second regression model. By integrating shear modulus and hardening index parameters, personalized material parameters can be generated.

7. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 1, characterized in that, Based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters, an adaptive posture prediction model is used to obtain the optimal posture that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario, including: Based on the biomechanical posture type probability vector in the biomechanical attribute parameters, the individualized neutral posture joint angles are determined by processing through a predefined lookup table. Based on individualized neutral posture joint angles, vehicle geometric parameters, and task scenario identifiers, a mixed integer programming model is used to construct and process an objective function that includes posture deviation penalty terms and muscle activation penalty terms. Based on the objective function and constraints, the optimal pose is obtained by processing the data through an optimization solver.

8. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 7, characterized in that, Based on the objective function and constraints, an optimization solver is used to obtain the optimal pose, including: Based on the objective function and constraints, a branch and bound algorithm is used to generate a candidate solution space. Based on the candidate solution space, a parallel computing architecture is used to accelerate the evaluation of multiple possible solutions; Based on the accelerated evaluation results, the optimal solution selection strategy is adopted to obtain the optimal pose.

9. The physiological mechanism-driven and data fusion method for seated human body modeling according to claim 1, characterized in that, Based on deformable 3D human body meshes and optimal poses, a high-fidelity seated digital twin is generated through system coupling model processing, including: Based on the deformable 3D human body mesh and the optimal pose, the 3D human body mesh with the adjusted pose is generated by processing it with skeletal skinning technology. Based on the 3D human body mesh after posture adjustment, a visual digital model of sitting posture is generated through real-time rendering processing using a physics engine. Based on a visualized digital model of sitting posture, a high-fidelity digital twin of sitting posture is generated by processing it through a dynamic feedback mechanism, which reflects the soft tissue deformation state after posture changes.

10. A physiological mechanism-driven and data fusion system for seated human body modeling, characterized in that, The system employs the method described in any one of claims 1 to 9, the system comprising: The physiological mechanism-driven module is used to generate biomechanical property parameters, including equivalent muscle activation thresholds, spinal segment flexure stiffness coefficients, and biomechanical posture type probability vectors, based on input macroscopic human body parameters. The parametric morphology generation module is used to generate a basic 3D human body mesh that matches the user's physiological characteristics based on macroscopic human body parameters. The soft tissue deformation simulation module is used to generate deformable 3D human body meshes that integrate the nonlinear compression deformation characteristics of soft tissues in a sitting posture, based on the basic 3D human body mesh and macroscopic human body parameters. The task-adaptive attitude prediction module is used to obtain the optimal attitude that matches the user's physiological characteristics, vehicle interior geometry, and dynamic task scenario based on vehicle geometric parameters, task scenario identifiers, and biomechanical attribute parameters. The system coupling module is used to generate a high-fidelity seated digital twin based on deformable 3D human body mesh and optimal posture.

Images