A physical reinforcement neural network-based fabric constitutive parameter rapid identification method

Abstract

The present application relates to a kind of physical reinforcement neural network-based fabric constitutive parameter fast identification method, comprising the following steps: 1), the macroscopic tensile, shear and bending test are carried out to the fabric sample to be identified, while collecting microstructure image, obtain multiscale feature dataset;2), physical reinforcement neural network model forward inference: the multiscale feature dataset is input into pre-trained physical reinforcement neural network;3), the complete constitutive parameter vector of target fabric is inferred and output by physical reinforcement neural network model;4), physical consistency verification: with constitutive parameter vector as input, drive finite element simulation, compare quantitatively the mechanical response curve of simulation output and measured curve, if error exceeds threshold value, then trigger adaptive fine-tuning process;5), physical reinforcement neural network model is continuously trained and updated.The present application can realize the constitutive parameter identification of fast, small sample, high precision for new variety fabric.

Description

Technical Field

[0001] This invention relates to a method for rapid identification of fabric constitutive parameters based on a Physics-Augmented Neural Network (PANN), which is suitable for rapid digital modeling of fabric materials in fast-response garment manufacturing scenarios and belongs to the interdisciplinary field of intelligent textile manufacturing and machine learning. Background Technology

[0002] Fast-response garment manufacturing requires the production system to complete the entire manufacturing process from style design to finished product shipment in a very short time (within 14 days). The key lies in relying on digital technology to quickly and accurately model the mechanical constitutive properties of new fabrics, thereby simulating the real response of the fabric under deformation conditions such as stretching, shearing and bending in a computer environment, providing support for automatic cutting process planning and intelligent fabric laying control.

[0003] Fabrics are flexible materials with significant multi-scale structural characteristics. Their mechanical behavior is influenced not only by macroscopic deformation conditions but also by the microscopic yarn interweaving structure and microscopic fiber material properties. To achieve digital modeling of the aforementioned fabric mechanical behavior, it is necessary to establish a material constitutive model of the fabric and determine the corresponding constitutive parameters. The constitutive parameters that fully describe the fabric mechanical behavior include: orthotropic elastic tensors (warp, weft, shear modulus, and Poisson's ratio, a total of 5 independent components), anisotropic bending stiffness coefficients (3 independent components), frictional anisotropy coefficients (2 directions), and viscoelastic relaxation time spectrum parameters. These parameters have high spatial dimensions and require long measurement times.

[0004] Existing constitutive parameter identification methods have the following three core limitations: 1. Traditional experimental characterization methods are time-consuming and labor-intensive. These methods employ standard tests such as fabric tensile properties, shear properties, and bending properties, analyzing the test data to obtain corresponding constitutive parameters. However, these methods require extensive experimental equipment and skilled operators; completing the full parameter determination for a single fabric variety typically takes 8–16 hours and consumes a large amount of sample fabric. Therefore, in fast-response garment manufacturing scenarios, these methods are insufficient to meet the demands for rapid digital modeling of new fabric varieties.

[0005] 2. Purely data-driven machine learning methods have limited generalization ability. Existing studies have used deep neural networks to model the mechanical behavior of fabrics end-to-end. Although they can achieve good fitting accuracy within the training distribution, they have two fatal flaws: First, generalization to new fabric varieties requires a large number of labeled samples (usually more than 2,000 mechanical test data points), and samples of new fabrics are extremely scarce in fast-response scenarios; Second, the constitutive parameters output by purely data-driven networks cannot guarantee that they satisfy the basic principles of continuum mechanics (such as energy conservation and material objectivity), which often leads to divergence or non-physical interpretations in numerical simulations.

[0006] 3. The parameter identification efficiency of physical model-driven methods is low. Although physical-driven methods, such as finite element inverse analysis, can guarantee physical consistency, they are essentially high-dimensional nonlinear optimization problems. Each iteration requires a complete solution of the forward finite element, and a single identification takes several hours, which cannot meet the needs of fast-response manufacturing for the rapid deployment of new fabric varieties.

[0007] Therefore, in order to solve the above problems, it is indeed necessary to provide a fast identification method for fabric constitutive parameters based on physical augmentation neural networks that has fast identification speed, strict physical constraints, and low sample requirements, so as to overcome the defects in the prior art. Summary of the Invention

[0008] The purpose of this invention is to provide a rapid identification method for fabric constitutive parameters based on a physical augmentation neural network. This method embeds the three-scale physical mechanisms of yarn, weave, and fabric into the neural network structure in the form of hard constraints. Under the premise of ensuring physical consistency, it achieves rapid (single-variety identification time < 2 hours), small sample (no more than 500 mechanical test data points), and high-precision (mean error ≤ 6.3%) constitutive parameter identification for new fabric varieties.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for rapid identification of fabric constitutive parameters based on a physical augmentation neural network, comprising the following steps: 1) Multi-scale mechanical test data acquisition: Macroscopic tensile, shear and bending tests are performed on the fabric sample to be identified, and microscopic images are acquired at the same time to obtain a multi-scale feature dataset including stress-strain curves, bending stiffness and fabric structure parameters. 2) Forward inference of the physical augmentation neural network model: The multi-scale feature dataset is input into the pre-trained physical augmentation neural network; the physical augmentation neural network includes a microscopic prior embedding layer, a microscopic structural constraint layer, and a macroscopic constitutive decoder; 3) Constitutive parameter output and uncertainty quantification: The complete constitutive parameter vector of the target fabric is inferred and output through the physical augmented neural network model, including the independent components of the anisotropic elastic tensor, the bending stiffness matrix components and the viscoelastic relaxation parameters, and the confidence interval of each parameter is also output. 4) Physical consistency verification: Using the constitutive parameter vector from step 3) as input, drive the finite element simulation, quantitatively compare the mechanical response curve output by the simulation with the measured curve, and if the error exceeds the threshold, trigger the adaptive fine-tuning process. 5) Online incremental learning: The constitutive parameters verified in step 4) are stored in the fabric material database along with the corresponding fabric fiber composition, fabric specifications and batch information, and the physical augmentation neural network model is continuously trained and updated based on this.

[0010] The fast fabric constitutive parameter identification method based on physical augmentation neural network of the present invention is further described in step 1) as follows: 1-1) Macroscopic mechanical testing: Macroscopic tensile, shear, and bending tests are performed on the fabric sample to be identified. Specifically, this includes: obtaining stress-strain curves in different directions by performing warp, weft, and 45° oblique uniaxial tensile tests; obtaining internal shear response by performing simple warp and weft shear tests; and obtaining bending stiffness data by performing mandrel bending tests. 1-2) Detailed parameter extraction: Obtain images of the warp and weft yarn interlacing structure using an industrial microscope or digital microscopy imaging module, and extract structural parameters such as warp density, weft density, warp and weft yarn interlacing angle, weave repeat number, and average float length. 1-3) Construction of multi-scale feature dataset: The stress-strain curves, shear response, and bending stiffness data obtained in step 1-1) are registered, normalized, and feature-stitched with the fabric structure parameters extracted in step 1-2) to form a multi-scale feature dataset for input to the physical augmented neural network model.

[0011] The fast fabric constitutive parameter identification method based on physical augmentation neural network of the present invention is further described as follows: the specific operation method of the micro prior embedding layer in step 2) is as follows: 2-1), the microscopic prior embedded layer accepts information including yarn linear density d, twist T, and fiber elastic modulus E. f and fiber cross-sectional area ratio f eigenvector X micro That is: X micro =[d,T,E f , f ]; 2-2), Constructing a weight initialization based on physical priors: Using the Voigt-Reuss homogenization model, the axial elastic modulus E of the yarn is... y Perform analytical estimation: ,in, f E represents the cross-sectional area ratio of the fiber. f E represents the fiber's elastic modulus. mThe matrix modulus; The above analytical estimates are used as the initial weights W for the corresponding yarn mechanical feature mapping channels in the network. init ,in, ; 2-3), for neurons that output scalar mechanical parameters such as the elastic modulus and stiffness of the yarn, the Softplus activation function f(Z)=ln(1+e) is used. Z This ensures that the output value is strictly positive. 2-4), the original input X is connected via residual connection. micro Directly applied to the output, forming the physically calibrated yarn-level mechanical strength characteristic H. micro And pass it to the next level, the calculation formula is: H micro =f(W init ·X micro +b)+L inear (X micro ); where b represents the bias term; L inear () refers to a linear transformation.

[0012] In step 2), the input to the microstructure constraint layer includes two paths: one is the yarn-level mechanical strength feature H from the micro-a priori embedding layer. micro The other path is the weaving structure parameter vector X extracted from the microstructure image of the fabric. meso Including warp density n w weft density n f The warp and weft yarn interlacing angle θ, the number of weave repeats R, and the average float length l f That is: X meso =[n w ,n f ,θ,R,l f ]; In step 2), the computation method for the microstructure constraint layer is as follows: First, the symmetry group to which the constitutive tensor belongs is automatically identified based on the fabric structure; then, based on the determined symmetry group, the corresponding linear constraint matrix P is applied at the output of the microstructure constraint layer to ensure that the feature tensor C is transmitted to the next layer during the feature forward propagation process. meso Satisfy: P·vec(C micro )=vec(C meso ), where P represents the material symmetry projection matrix; vec() represents the tensor column vectorization operator; C micro The input feature tensor.

[0013] In step 2), the macroscopic constitutive decoder receives the weave structure parameter vector X from the mesoscopic structural constraint layer. meso Decode it into a complete set of constitutive parameter vectors θ const .

[0014] The fast fabric constitutive parameter identification method based on physical augmentation neural network of the present invention is further described in step 3), which specifically includes: 3-1), the target fabric complete constitutive parameter vector θ output in step 2) const A hybrid strategy combining Monte Carlo Dropout and deep ensemble methods is employed to output mean estimates and 95% confidence intervals for each parameter component.

[0015] The fast fabric constitutive parameter identification method based on physical augmentation neural network of the present invention is further described in step 4), which specifically includes: 4-1), the constitutive parameter vector θ obtained in step 3) const Fill the preset fabric finite element template and automatically generate the corresponding finite element input file; 4-2), call the finite element verification module to perform quasi-static nonlinear simulation of the fabric under tension, shear, bending or suspension conditions, and obtain the simulated mechanical response curve; 4-3) The stress-strain response curves output by the simulation are quantitatively compared with the measured curves, and the normalized root mean square error ε is calculated. RMSE If ε RMSE >ε threshold , ε threshold If the default threshold is 8%, the adaptive fine-tuning process is triggered; otherwise, the current parameters are accepted and written to the fabric material database.

[0016] The adaptive fine-tuning process employs a gradient-guided freeze fine-tuning strategy, which completely freezes the parameters of the microscopic prior embedding layer and the mesoscopic structural constraint layer to preserve physical priors and symmetry constraints, and only updates the gradients of the macroscopic constitutive decoder.

[0017] The fast fabric constitutive parameter identification method based on a physically enhanced neural network of the present invention further comprises: step 5), the online incremental learning method based on the physically enhanced neural network model, specifically as follows: 5-1) Store the constitutive parameters verified in step 4) along with the corresponding fiber composition, fabric specifications and batch information of the fabric into the fabric material database; whenever 20 verified varieties are added to the fabric material database, the online incremental learning trigger condition of the physical augmentation neural network model is automatically triggered. 5-2), the training of the physical augmentation neural network model uses a multi-objective loss function, calculated as follows: L total = λ1·L data + λ2·L physics + λ3·L reg ; Among them, L dataFor data fitting loss, L physics For the physical residual loss calculated by automatic differentiation, L reg λ1, λ2, and λ3 are the parameter regularization loss and the weight coefficients.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. It breaks through the material digitization bottleneck in garment rapid response manufacturing, reducing the constitutive determination time of a single fabric from 8-16 hours to less than 2 hours, and adapting to the 14-day rapid response delivery cycle requirement; 2. Significantly reduces sample requirements and fabric waste. The sample requirement is reduced by approximately 84% compared to pure data-driven methods, making it particularly suitable for constitutive testing scenarios involving high-value fabrics. 3. The recognition results natively satisfy the mechanical and physical constraints of the continuous medium, which can directly drive finite element simulation without subsequent processing, thus opening up the entire process link of "fabric digitization - process simulation - production implementation". 4. It has the ability to continuously iterate. As the fabric database accumulates, the recognition accuracy continues to improve, adapting to the production characteristics of the apparel industry, which features multiple varieties, small batches, and rapid iteration. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the Physical Augmentation Neural Network (PANN) of the present invention.

[0020] Figure 2 This is a comparison of the efficiency of the multi-scale mechanical test data acquisition process in step 1) of the present invention with that of the physical enhancement neural network method.

[0021] Figure 3 This is a schematic diagram illustrating the weight initialization constraint principle of the microscopic prior embedding layer in the physical enhancement neural network of the present invention.

[0022] Figure 4 This is a schematic diagram illustrating the implementation principle of anisotropic symmetry constraints in the microstructure constraint layer of the physical augmentation neural network of the present invention.

[0023] Figure 5 This invention relates to the parameterized viscoelastic constitutive framework and output parameter recognition accuracy of the macroscopic constitutive decoder in the physical augmentation neural network.

[0024] Figure 6 This is a comprehensive diagram of the simulation verification and engineering benefits of the physical augmentation neural network in step 2) of the present invention.

[0025] Figure 7 This is a schematic diagram of the adaptive fine-tuning process and fine-tuning efficiency in step 4) of the present invention.

[0026] Figure 8This is a diagram illustrating the online incremental learning mechanism and performance verification of the physical augmentation neural network in step 5) of the present invention.

[0027] Figure 9 This is a bar graph showing the comparative experimental results of the identification accuracy of constitutive parameters of 20 types of fabrics in this invention.

[0028] Figure 10 This is a scatter plot comparing the recognition timeliness and sample demand in this invention. Detailed Implementation

[0029] Please refer to the instruction manual appendix. Figure 1 To be continued Figure 10 As shown, this invention is a method for rapid identification of fabric constitutive parameters based on a physical augmentation neural network, which specifically includes the following process steps: 1) Multi-scale mechanical test data acquisition: Macroscopic tensile, shear and bending tests are performed on the fabric sample to be identified, and microscopic images are acquired at the same time to obtain a multi-scale feature dataset including stress-strain curves, bending stiffness and fabric structure parameters.

[0030] Please refer to the instruction manual for details. Figure 2 As shown in (a), this step is specifically as follows: 1-1) Macroscopic Mechanical Testing: Macroscopic tensile, shear, and bending tests are performed on the fabric sample to be identified. Specifically, this includes: obtaining stress-strain curves in different directions via warp, weft, and 45° oblique uniaxial tensile tests; obtaining internal shear response via warp and weft simple shear tests; and obtaining bending stiffness data via mandrel bending tests. Test conditions must cover the actual processing deformation range; for example, the strain rate should be set to be no less than 0.01 s⁻¹. -1 The temperature was controlled within the range of 15 ℃ to 35 ℃ to ensure the validity of the data.

[0031] 1-2) Detailed parameter extraction: Images of the warp and weft yarn interlacing structure are obtained using an industrial microscope or digital microscopy imaging module, and structural parameters such as warp density, weft density, warp and weft yarn interlacing angle, weave repeat number, and average float length are extracted. Among them, float length is the length of a single yarn in a woven fabric that does not interlace with yarns from another system, but instead continuously crosses two or more opposite yarns.

[0032] 1-3) Constructing a multi-scale feature dataset: The stress-strain curves, shear response, and bending stiffness data obtained in step 1-1) are registered, normalized, and feature-stitched with the fabric structure parameters extracted in step 1-2) to form a multi-scale feature dataset for input to the physical augmentation neural network model. (Combined with the appendix...) Figure 2As can be seen from the data in part (b), the traditional finite element method requires about 5,000 samples, pure data-driven method requires 2,200 samples, while this creation only requires about 350 samples to meet the input requirements, reducing the number of test samples by about 93% and reducing the total time from the traditional 24 hours to 1.8 hours.

[0033] 2) Forward inference of the physics-enhanced neural network model: The multi-scale feature dataset is input into the pre-trained physics-enhanced neural network. (See attached manual.) Figure 1 As shown, the physical augmentation neural network includes a microscopic prior embedding layer, a microscopic structural constraint layer, and a macroscopic constitutive decoder. These three functional modules are connected in series. The network as a whole adopts residual connections (ResNet style) to prevent deep gradient vanishing. Each layer integrates batch normalization to accelerate convergence.

[0034] The Microscopic Prior Embedding Layer (MPEL) is the first input layer of the physical augmentation neural network, essentially a physical knowledge injection module based on the homogenization theory of micromechanics. The specific computation method of the Microscopic Prior Embedding Layer is as follows: 2-1), as attached Figure 3 As shown in (a), the microscopic prior embedded layer accepts yarn linear density d (unit: dtex), twist T (twist / 10cm), and fiber elastic modulus E. f (GPa) and fiber cross-sectional area ratio eigenvector X micro That is: X micro =[d,T,E f , f Among them, fiber cross-sectional area ratio. It refers to the ratio of the area occupied by the fibrous material on the cross-section of the fiber to the total projected area of ​​the fiber (including pores).

[0035] 2-2), Constructing a weight initialization based on physical priors: Using the Voigt-Reuss homogenization model, the axial elastic modulus E of the yarn is... y Perform analytical estimation: ,in, f E represents the cross-sectional area ratio of the fiber. f E represents the fiber's elastic modulus. m The matrix modulus.

[0036] The above analytical estimates are used as the initial weights W for the corresponding yarn mechanical feature mapping channels in the network. init ,in, This allows the prior knowledge of yarn-level physical mechanisms to be encoded into the network using hard constraints.

[0037] 2-3), for neurons that output scalar mechanical parameters such as the elastic modulus and stiffness of the yarn, the Softplus activation function f(Z)=ln(1+e) is used. Z This ensures that the output value is strictly positive, eliminating physically impossible predictions. (Combined with the appendix...) Figure 3 The validation loss curve in (b) clearly shows the advantages of this physical prior initialization: the model with physical initialization using the micro prior embedding layer can reach the convergence threshold in the 85th round, while the baseline random initialization model does not converge until the 155th round, which greatly improves the convergence speed by 45%.

[0038] 2-4), the original input X is connected via residual connection. micro Directly applied to the output, forming the physically calibrated yarn-level mechanical strength characteristic H. micro And pass it to the next level, the calculation formula is: H micro =f(W init ·X micro +b)+L inear (X micro ); where b represents the bias term, corresponding to the conventional learnable bias parameters in a neural network; L inear (·) indicates a linear transformation, representing a residual connection branch, typically a linear layer. If the original input X micro Unlike the main output dimension, it performs a linear projection to align the dimensions; if the dimensions are the same, it's equivalent to directly projecting X... micro Add the past (i.e., identity mapping).

[0039] Furthermore, the Microstructure Constraint Layer (MSCL) undertakes cross-scale feature aggregation, and is essentially a hypothesis space dimensionality reduction module based on physical priors. (See attached...) Figure 4 As shown in part (a), the input to the microstructure constraint layer includes two paths: one is the yarn-level mechanical strength feature H from the micro-prior embedding layer. micro The other path is the weaving structure parameter vector X extracted from the microstructure image of the fabric. meso Including warp density n w (threads / cm), weft density n f (threads / cm), warp and weft yarn interlacing angle θ (°), weave repeat number R, and average float length l f That is: X meso =[n w ,n f ,θ,R,l f ].

[0040] The constraint mechanism of the microstructure constraint layer is the "Material Symmetry Projection" (MSP) technique. When it outputs, it first automatically identifies the symmetry group to which the constitutive tensor belongs based on the fabric structure. For example, plain weave fabrics are orthogonally anisotropic, twill / satin weave fabrics are monoclinic symmetric, and knitted fabrics are transversely isotropic.

[0041] Next, based on the symmetric group of the decision, a corresponding linear constraint matrix P is applied at the output of the mesostructure constraint layer to ensure that the feature tensor C is passed to the next layer during feature forward propagation. meso Satisfy: P·vec(C micro )=vec(C meso ), where P represents the material symmetry projection matrix; vec() represents the tensor column vectorization operator; C micro The input is a feature tensor. This operation projects the tensor into a valid subspace that meets the requirements of a symmetric group, completely avoiding the risk of physical distortion caused by improper weight allocation in the loss function. Combined with... Figure 4 As shown in the error histogram in part (b), without the microstructure constraint layer, the prediction error of each fabric structure is as high as 9.8% to 14.1% (all far exceeding the target line of 6.3%). However, after adding the microstructure constraint layer projection constraint, the errors of all categories, including plain weave (3.2%) and knitted plain fabric (5.2%), are significantly reduced to below the target baseline.

[0042] The macroscopic constitutive decoder (MCD) is as follows: Figure 5 As shown in part (a), it serves as the terminal output core module of the hierarchical constraint architecture of the physical augmented neural network, and its essence is a thermodynamically consistent parameterized inverse mapping operator.

[0043] In this embodiment, the macroscopic constitutive decoder receives the weaving structure parameter vector X from the mesoscopic structure constraint layer. meso Decode it into a complete set of constitutive parameter vectors θ const This vector encompasses 14 independent components across multiple physical fields, including anisotropic elastic tensors (such as the meridional elastic modulus E1, the zonal elastic modulus E2, and the in-plane shear modulus G). 12 (etc.), bending stiffness matrix (such as meridional bending stiffness B) 11 1. Weft bending stiffness B 22 (etc.) and viscoelastic relaxation parameters, where relaxation is the continuous stress response behavior of the fabric under strain from a short time (instantaneous elasticity) to a long time (molecular chain rearrangement), such as relaxation strength and relaxation time, as detailed in Table 1 below: Table 1

[0044] Furthermore, the macroscopic constitutive decoder ensures that the output satisfies the laws of thermodynamics through hard constraints at the algebraic level: the elastic modulus parameters (E, G) are guaranteed to be positive definite using the Softplus activation function; Poisson's ratio (ν) 12 The Sigmoid function is used and linearly mapped to the interval (0, 0.5) to satisfy the thermodynamic stability inequality; the viscoelastic relaxation strength uses the Softmax activation operator to ensure that its summation is ≤1, satisfying the second law of thermodynamics; the bending stiffness matrix (B) is parameterized through Cholesky decomposition to ensure positive definiteness. (See attached diagram) Figure 5 As shown in the error statistics of part (b), the independent parameters of the MCD output (hyperelastic parameter C) 10 C 01 C 20 Modulus G 12 Poisson's ratio ν 12 The average error of all of them (etc.) is between 2.7% and 5.9%, which fully meets the high precision requirements of the project.

[0045] 3) Constitutive parameter output and uncertainty quantification: The complete constitutive parameter vector of the target fabric is inferred and output through the physical augmented neural network model, including the independent components of the anisotropic elastic tensor, the bending stiffness matrix components and the viscoelastic relaxation parameters, and the confidence interval of each parameter is also output. Among them, the target fabric complete constitutive parameter vector θ output in step 2) const A hybrid strategy combining Monte Carlo Dropout and deep ensemble methods is employed to output mean estimates and 95% confidence intervals for each parameter component.

[0046] For example, for plain cotton woven fabrics, the constitutive parameters and 95% confidence intervals of the physical reinforcement neural network output are as follows: E1 is 7.82 MPa [7.41, 8.23], E2 is 5.64 MPa [5.31, 5.97], G... 12 The value is 0.43 MPa [0.38, 0.48], ν 12 B is 0.327 [0.305, 0.349]. 11 2.14 × 10 -5 N·m [1.98, 2.30]×10 -5 These uncertainties will be passed to the downstream fabric control planning module for adaptive adjustment of the control strategy's safety margin. (See attached...) Figure 6 The comprehensive diagram of simulation verification and engineering benefits of the physical augmented neural network is shown in the figure. Figure 6(a) reflects the high degree of consistency between the finite element overhang simulation driven by PANN output parameters (PANN-FEM) and the actual test, with a normalized position REMS of only 0.73 cm (compared to 3.05 cm in the traditional finite element method). (See attached image.) Figure 6 (b) shows a summary of finite element simulation errors, including peak stress and wrinkle morphology. (See attached image.) Figure 6 (c) indicates that the simulation error of this method is less than 5% (as low as 2.1%) in all five macroscopic tasks: suspension height, peak stress, wrinkle morphology, bending stiffness, and friction coefficient. (See attached diagram.) Figure 6 According to industrial verification statistics (d), this method reduces costs and increases efficiency by 87% in the parameter identification stage compared with the traditional solution, and shortens the overall R&D cycle by 42%.

[0047] 4) Physical consistency verification: Using the constitutive parameter vector from step 3) as input, drive the finite element simulation, quantitatively compare the mechanical response curve output by the simulation with the measured curve, and if the error exceeds the threshold, trigger the adaptive fine-tuning process.

[0048] This step is specifically as follows: 4-1), the constitutive parameter vector θ obtained in step 3) const Fill the preset fabric finite element template and automatically generate the corresponding finite element input file.

[0049] 4-2), call the finite element verification module to perform quasi-static nonlinear simulation of the fabric under tension, shear, bending or suspension conditions, and obtain the simulated mechanical response curve.

[0050] 4-3) The stress-strain response curves output by the simulation are quantitatively compared with the measured curves, and the normalized root mean square error ε is calculated. RMSE If ε RMSE >ε threshold , ε threshold If the default threshold is 8%, the adaptive fine-tuning process is triggered; otherwise, the current parameters are accepted and written to the fabric material database.

[0051] The adaptive fine-tuning process is described in the appendix to the specification. Figure 7 As shown in part (a), a gradient-guided freeze fine-tuning strategy is employed, completely freezing the parameters of the microscopic prior embedding layer and the mesoscopic structural constraint layer to preserve physical priors and symmetry constraints, while only updating the gradient of the macroscopic constitutive decoder. This gradient signal originates from the finite element simulation residuals and is implemented using differentiable finite element techniques. (Appendix) Figure 7 Part (b) visually demonstrates the fine-tuning efficiency: for a specific high-error batch, G 2The FT method only needs to add 3 new test points to make the recognition error drop sharply to below the target baseline (≤6.3%); while the full retraining method and the random fine-tuning method still have difficulty converging completely when 10 new test points are added.

[0052] 5) Online incremental learning: The constitutive parameters verified in step 4) are stored in the fabric material database along with the corresponding fabric fiber composition, fabric specifications and batch information, and the physical augmentation neural network model is continuously trained and updated based on this.

[0053] The online incremental learning method for the physical augmentation neural network model is as follows: 5-1) Store the constitutive parameters verified in step 4) along with the corresponding fabric's fiber composition, fabric specifications, and batch information into the fabric material database. (See attached...) Figure 8 As shown in part (a), if incremental learning is performed using a pure data method (red line) without protection mechanisms, "catastrophic forgetting" (error rebounds to over 12%) will occur when the accumulated data volume reaches several hundred.

[0054] Whenever 20 verified fabric material varieties are added to the database, the online incremental learning trigger condition of the physical augmentation neural network model is automatically triggered. Incremental learning employs an elastic weight consolidation (EWC) strategy. (Combined with...) Figure 8 As can be seen from the principle in (b), EWC applies regularized elastic constraint penalties to the important weights of historical tasks through the Fisher information matrix. With the continuous expansion of the database (such as...), Figure 8 As shown by the blue line in (a), the error of the PANN+EWC mechanism not only did not rebound, but also stably broke through the target baseline when the number of accumulated data points reached N=200, and further approached 4.0% as data accumulated, effectively ensuring the system's ability to continuously evolve.

[0055] 5-2), The training of the physical augmentation neural network model uses a multi-objective loss function: L total = λ1·L data + λ2·L physics + λ3·L reg ; Among them, L data For data fitting loss, L physics For the physical residual loss calculated by automatic differentiation, L reg Let λ1, λ2, and λ3 be the parameter regularization loss, and λ1, λ2, and λ3 be the weight coefficients. The weight coefficients λ1, λ2, and λ3 are adaptively adjusted by a dynamically weighted strategy (ALB), with an emphasis on L1 during the initial training phase. physics To quickly establish physical awareness, the later stages of training will emphasize L. dataTo refine the fitting accuracy, an active learning strategy is adopted for training set construction. The most representative historical samples are selected from the fabric material database according to the information entropy maximization criterion to reduce the cost of training data collection.

[0056] To verify the generalizability and ultimate performance of this invention, see attached... Figure 9 As shown, this invention sets up three sets of controls: traditional inverse finite element method, pure data-driven method, and the PANN method of this invention. The constitutive parameters of 20 typical fabrics (cotton woven / polyester woven / silk / nylon / elastic fabrics, etc.) are identified. The results show that the average error of the constitutive parameters is less than 6.3%, and the overall accuracy is significantly better than the traditional finite element method and the data-driven method. In addition to the identification accuracy, this invention also has significant advantages in terms of sample requirements and identification timeliness, as shown in the attached figure. Figure 10 As shown, compared to traditional standard testing methods and pure data-driven methods, the data points of this invention are highly concentrated within the "target region" of the coordinate axis, vividly demonstrating that this invention achieves a leapfrog breakthrough by reducing the sample size by 93% and increasing the speed by 8 times.

[0057] Example 1: Plain cotton woven fabric (surface weight 180 g / m²) 2 Constitutive parameter identification This example focuses on plain cotton woven fabrics, the most commonly used fabric in the apparel industry, to verify the basic recognition accuracy and timeliness of this invention.

[0058] 1) Sample preparation. Take 10 plain weave cotton woven fabric samples from the same batch, of which 7 samples are used for macroscopic mechanical testing and 3 samples are reserved for finite element verification.

[0059] 2) PANN forward inference. Input the collected data into PANN (inference takes about 8 seconds) and output the constitutive parameter vector.

[0060] 3) Finite element verification. A rectangular fabric drape simulation (80cm × 80cm, center support) was performed using the identified parameters. The Hausdorff distance between the simulated profile and the actual drape profile was 4.3mm, and the relative drape height deviation was 3.2%, which is below the preset threshold of 8%, thus the verification passed. The entire identification process took 1 hour and 42 minutes.

[0061] Example 2: Elastic polyester woven fabric (surface weight 240 g / m²) 2 Constitutive parameters identified by (containing 10% spandex) This example focuses on a highly nonlinear elastic fabric to verify the adaptability of the present invention to fabrics with complex mechanical behavior.

[0062] 1) Test data acquisition. In the standard tensile test, the strain range was extended to 80%, and the shear test adopted the off-axis tensile method; the microscopic image showed that the spandex was embedded in the warp yarn in the form of a cover yarn, and the cover angle parameter β = 23.5° was extracted as the MSCL input.

[0063] 2) PANN automatically activates the hyperelastic mode. After detecting this structural feature, MSCL automatically selects a composite constitutive framework of Yeoh hyperelasticity and linear viscoelasticity, and outputs the corresponding hyperelastic parameters and additional stiffness correction terms, without manual intervention.

[0064] 3) Finite element verification. Simulation of large deformation garment conditions was performed using identified parameters. The peak error of the simulated stress field was 7.1%, and the verification was successful. The entire identification process took 1 hour and 58 minutes, a 75% improvement in efficiency compared to traditional methods.

[0065] Example 3: Lightweight silk power spinning (38 g / m²) 2 Constitutive parameter identification and adaptive fine-tuning This example demonstrates the adaptive fine-tuning capability of the present invention by testing a lightweight silk fabric, which is difficult to test.

[0066] 1) A special testing scheme was adopted, using a frame-type clamping fixture supplemented with a weakly adhesive retainer to fix the edges, reducing the tensile strain rate to 0.001 s⁻¹. -1 Microscopic imaging is acquired when the fabric is naturally laid out to reduce stress concentration at clamping points and boundary disturbances.

[0067] 2) Initial Identification and Fine-tuning. Due to the systematic differences between the silk fiber structure and the basic training set, the drape profile deviation reached 8.7% during the initial PANN identification output constitutive parameter-driven finite element verification, exceeding the threshold and triggering the adaptive fine-tuning process. After additionally acquiring four off-axis shear test data points, G... 2 The FT strategy was used to fine-tune the parameters of the last two layers of the MCD, reducing the recognition error to 5.2%, and the re-verification was successful.

[0068] 3) Benefit analysis. The traditional method takes about 14 hours to determine the complete constitutive structure of electric spinning, while the total time including fine-tuning of this invention is 2 hours and 23 minutes, which is about 5.9 times more efficient.

[0069] Example 4: Knitted jacquard fabric (weft-knitted plain weave + jacquard structure, face weight 320 g / m²) 2 Constitutive parameter identification This example focuses on knitted jacquard fabrics with significant structural nonlinear characteristics, verifying the invention's ability to identify the complex anisotropic mechanical behavior of knitted fabrics.

[0070] 1) Microstructural Characterization and Constraint Fitting. By acquiring microstructural images of the knitted jacquard fabric, the loop geometric parameters (loop length l) were extracted. loop = 3.24 mm, etc.). MSCL automatically adapts to the transverse isotropic symmetry mode and automatically infers the anisotropy ratio.

[0071] 2) Constitutive parameter identification results. In the output constitutive parameters, E2 / E1 = 6.5, and the tension-bending coupling term B... 12 The result is non-zero, consistent with the mechanical properties of knitted fabrics. Finite element analysis shows an average identification error of 5.6%, meeting the accuracy requirements set by this invention.

[0072] 3) Efficiency Analysis. Traditional methods for calibrating the complete constitutive parameters of this type of knitted jacquard fabric usually require a long time for testing and repeated fitting. Using the method of this invention, the entire process from data acquisition and parameter identification to finite element verification takes 1 hour and 51 minutes, which is more than 77% faster than traditional methods.

[0073] Example 5: System-level deployment verification in a garment rapid response manufacturing enterprise This example focuses on identifying scenarios in a continuous industrial production environment to verify the feasibility of system-level deployment and engineering application effects of the present invention in garment rapid response manufacturing enterprises.

[0074] 1) Deployment Scenario and Operation Mode. This invention was deployed at the industrial site in a fast-response garment manufacturing enterprise. Over 30 consecutive working days, constitutive parameter identification, finite element verification, and necessary adaptive fine-tuning were performed on each of the 47 new fabric varieties that entered the factory to test the system's stable operation capability under real production rhythm.

[0075] 2) System recognition performance. Operational results show that the average recognition error for the 47 varieties was 5.8%, the average recognition time was 1 hour and 53 minutes, and the first-time pass rate for finite element verification was 91.5%. For the few varieties that failed the first verification, the system automatically triggered G... 2 FT adaptive fine-tuning process, via G 2 All tests passed after FT's minor adjustments.

[0076] 3) Industrial application benefits. After obtaining accurate constitutive parameters, the interlayer alignment accuracy of the fabric laying process was improved from ±1.8 mm to ±0.4 mm, and the configuration time for changeover parameters was shortened to 28 minutes, effectively adapting to the production needs of fast-response manufacturing and verifying that the invention has good industrial application value.

[0077] The above-described specific embodiments are merely preferred embodiments of this invention and are not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network, characterized in that: Includes the following steps: 1) Multi-scale mechanical test data acquisition: Macroscopic tensile, shear and bending tests are performed on the fabric sample to be identified, and microscopic images are acquired at the same time to obtain a multi-scale feature dataset including stress-strain curves, bending stiffness and fabric structure parameters. 2) Forward inference of the physical augmentation neural network model: The multi-scale feature dataset is input into the pre-trained physical augmentation neural network; the physical augmentation neural network includes a microscopic prior embedding layer, a microscopic structural constraint layer, and a macroscopic constitutive decoder; 3) Constitutive parameter output and uncertainty quantification: The complete constitutive parameter vector of the target fabric is inferred and output through the physical augmented neural network model, including the independent components of the anisotropic elastic tensor, the bending stiffness matrix components and the viscoelastic relaxation parameters, and the confidence interval of each parameter is also output. 4) Physical consistency verification: Using the constitutive parameter vector from step 3) as input, drive the finite element simulation, quantitatively compare the mechanical response curve output by the simulation with the measured curve, and if the error exceeds the threshold, trigger the adaptive fine-tuning process. 5) Online incremental learning: The constitutive parameters verified in step 4) are stored in the fabric material database along with the corresponding fabric fiber composition, fabric specifications and batch information, and the physical augmentation neural network model is continuously trained and updated based on this.

2. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 1, characterized in that: Step 1) specifically refers to: 1-1) Macroscopic mechanical testing: Macroscopic tensile, shear, and bending tests are performed on the fabric sample to be identified. Specifically, this includes: obtaining stress-strain curves in different directions by performing warp, weft, and 45° oblique uniaxial tensile tests; obtaining internal shear response by performing simple warp and weft shear tests; and obtaining bending stiffness data by performing mandrel bending tests. 1-2) Detailed parameter extraction: Obtain images of the warp and weft yarn interlacing structure using an industrial microscope or digital microscopy imaging module, and extract structural parameters such as warp density, weft density, warp and weft yarn interlacing angle, weave repeat number, and average float length. 1-3) Constructing a multi-scale feature dataset: Register, normalize and feature-join the stress-strain curves, shear response and bending stiffness data obtained in step 1-1) with the fabric structure parameters extracted in step 1-2) to form a multi-scale feature dataset for input to the physical augmented neural network model.

3. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 1, characterized in that: In step 2), the specific computation method of the microscopic prior embedding layer is as follows: 2-1), the microscopic prior embedded layer accepts information including yarn linear density d, twist T, and fiber elastic modulus E. f and fiber cross-sectional area ratio f eigenvector X micro That is: X micro =[d,T,E f , f ]; 2-2), Constructing a weight initialization based on physical priors: Using the Voigt-Reuss homogenization model, the axial elastic modulus E of the yarn is... y Perform analytical estimation: ,in, f E represents the cross-sectional area ratio of the fiber. f E represents the fiber's elastic modulus. m The matrix modulus; The above analytical estimates are used as the initial weights W for the corresponding yarn mechanical feature mapping channels in the network. init ,in, ; 2-3), for neurons that output scalar mechanical parameters such as the elastic modulus and stiffness of the yarn, the Softplus activation function f(Z)=ln(1+e) is used. Z This ensures that the output value is strictly positive. 2-4), the original input X is connected via residual connection. micro Directly applied to the output, forming the physically calibrated yarn-level mechanical strength characteristic H. micro And pass it to the next level, the calculation formula is: H micro =f(W init ·X micro +b)+L inear (X micro ); where b represents the bias term; L inear () refers to a linear transformation.

4. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 3, characterized in that: The input to the microstructure constraint layer includes two paths: one is the yarn-level mechanical strength feature H from the micro-prior embedding layer. micro The other path is the weaving structure parameter vector X extracted from the microstructure image of the fabric. meso Including warp density n w weft density n f The warp and weft yarn interlacing angle θ, the number of weave repeats R, and the average float length l f That is: X meso =[n w ,n f ,θ,R,l f ]; When the microstructure constraint layer outputs, firstly, it automatically identifies the symmetry group to which the constitutive tensor belongs based on the fabric structure; then, based on the determined symmetry group, it applies the corresponding linear constraint matrix P at the output of the microstructure constraint layer to ensure that the feature tensor C is passed to the next layer during feature forward propagation. meso Satisfy: P·vec(C) micro )=vec(C meso ), where P represents the material symmetry projection matrix; vec() represents the tensor column vectorization operator; C micro The input feature tensor.

5. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 4, characterized in that: The macroscopic constitutive decoder receives the weaving structure parameter vector X from the mesoscopic structural constraint layer. meso Decode it into a complete set of constitutive parameter vectors θ const .

6. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 1, characterized in that: In step 3), the target fabric complete constitutive parameter vector θ output in step 2) const A hybrid strategy combining Monte Carlo Dropout and deep ensemble methods is employed to output mean estimates and 95% confidence intervals for each parameter component.

7. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 1, characterized in that: Step 4) specifically refers to: 4-1), the constitutive parameter vector θ obtained in step 3) const Fill the preset fabric finite element template and automatically generate the corresponding finite element input file; 4-2), call the finite element verification module to perform quasi-static nonlinear simulation of the fabric under tension, shear, bending or suspension conditions, and obtain the simulated mechanical response curve; 4-3) The stress-strain response curves output by the simulation are quantitatively compared with the measured curves, and the normalized root mean square error ε is calculated. RMSE If ε RMSE > ε threshold , ε threshold If the default threshold is 8%, the adaptive fine-tuning process is triggered; otherwise, the current parameters are accepted and written to the fabric material database.

8. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 7, characterized in that: The adaptive fine-tuning process employs a gradient-guided freeze fine-tuning strategy, which completely freezes the parameters of the microscopic prior embedding layer and the mesoscopic structural constraint layer to preserve physical priors and symmetry constraints, and only updates the gradients of the macroscopic constitutive decoder.

9. The method for rapid identification of fabric constitutive parameters based on a physically reinforced neural network as described in claim 1, characterized in that: The online incremental learning method for the physical augmentation neural network model is as follows: 5-1) Store the constitutive parameters verified in step 4) along with the corresponding fiber composition, fabric specifications and batch information of the fabric into the fabric material database; whenever 20 verified varieties are added to the fabric material database, the online incremental learning trigger condition of the physical augmentation neural network model is automatically triggered. 5-2), The training of the physical augmentation neural network model uses a multi-objective loss function: L total = λ1·L data + λ2·L physics + λ3·L reg ; Among them, L data For data fitting loss, L physics For the physical residual loss calculated by automatic differentiation, L reg λ1, λ2, and λ3 are the parameter regularization loss and the weight coefficients.

Images