Clothing style 3D intelligent simulation generation method based on model library

Abstract

The invention discloses a 3D intelligent simulation generation method for clothing styles based on a model library, and relates to the technical field of digital twinning, and the method comprises the steps: carrying out the entity relation extraction and semantic mapping of multi-source clothing data, constructing a cross-modal knowledge graph, and extracting a physical constraint rule of the cross-modal knowledge graph; performing association reasoning and cross validation on the physical constraint rule and the version library to obtain a basic version template set; inputting the basic model template set into an intelligent clothing model, performing structured feature decoupling by the parameter analysis layer, dynamically adjusting a topological structure by the geometric generation layer, and generating 3D simulated clothing; and carrying out digital modeling on the basic model template set through a digital twinning algorithm, constructing a digital twinning body, carrying out physical simulation and optimization verification on the 3D simulation clothing by using the digital twinning body, and outputting a clothing style 3D simulation scheme. According to the invention, by constructing the pattern library and the intelligent garment model, the efficiency, accuracy and individuation level of garment 3D simulation generation are improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method for intelligent 3D simulation generation of clothing styles based on a pattern library. Background Technology

[0002] In recent years, with the accelerated digital transformation of the apparel industry, computer-aided design, virtual fitting, and 3D garment modeling technologies have become core drivers of industry upgrading. Traditional apparel relies on the experience of tailors and manual pattern making, resulting in low efficiency, lagging physical compliance verification, and insufficient response to personalized needs. Against this backdrop, data-driven intelligent methods are gradually emerging, and multi-source data collection and storage technologies are maturing, providing more comprehensive information support for decision-making. The integration of artificial intelligence and physical simulation technologies is driving the transformation of the apparel industry from "experience-oriented" to "data- and rule-driven."

[0003] However, existing technologies still have some shortcomings. Existing pattern libraries are mostly based on simple storage of structured parameters, lacking deep integration and intelligent association of multimodal data, resulting in a disconnect between pattern rules and physical parameters, making it difficult to meet personalized needs in complex scenarios. 3D clothing simulation generation relies heavily on independent calculations of the physics engine, failing to form a closed-loop optimization, resulting in low iteration efficiency and difficulty in verifying the synergy between physical compliance and aesthetic appeal in real time. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for intelligent 3D simulation generation of clothing styles based on a pattern library to solve the problems of insufficient multimodal data fusion and poor aesthetic coordination.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a method for intelligent 3D simulation generation of clothing styles based on a pattern library, which includes:

[0008] Collect multi-source clothing data and historical pattern data, perform multimodal feature alignment and intelligent knowledge distillation on historical pattern data, and build a pattern library;

[0009] Entity relationships are extracted and semantics are mapped from multi-source clothing data to construct a cross-modal knowledge graph and extract the physical constraint rules of the cross-modal knowledge graph.

[0010] Perform correlation reasoning and cross-validation between physical constraint rules and pattern library to obtain a basic pattern template set;

[0011] The basic pattern template set is input into the intelligent clothing model. The parameter parsing layer decouples the structured features, and the geometry generation layer dynamically adjusts the topology to generate a 3D simulated clothing.

[0012] The basic pattern template set is digitally modeled using digital twin algorithms to construct a digital twin. The digital twin is then used to perform physical simulation and optimization verification on 3D simulated clothing, and output a 3D simulation scheme for clothing styles.

[0013] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the multi-source clothing data includes market trend data, competitor analysis data, brand creative data, and user behavior data.

[0014] The historical pattern data includes structural parameter data, process rule data, and visual element data.

[0015] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles according to the present invention, the specific steps for constructing the pattern library are as follows:

[0016] The structural parameter data is subjected to coordinate mapping and adjacency graph convolution to generate geometric feature vectors;

[0017] Semantic parsing and feature encoding are performed on process rule data to generate semantic feature vectors;

[0018] Visual element data is visually encoded, semantically labeled, and color decomposed to generate composite visual feature vectors.

[0019] Geometric feature vectors, semantic feature vectors, and composite visual feature vectors are multimodal aligned, fused, and categorized for storage to generate a template library.

[0020] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the specific steps for constructing the cross-modal knowledge graph are as follows:

[0021] Perform entity identification and attribute extraction on market trend data and brand creative data to generate entity attribute pairs;

[0022] Perform cross-modal feature association on competitor analysis data to generate composite attribute sets;

[0023] Mining time-series preferences from user behavior data and performing collaborative filtering clustering to generate a structured preference knowledge set;

[0024] Cross-modal knowledge is embedded into entity attribute pairs, composite attribute sets, and structured preference knowledge sets to generate a cross-modal knowledge graph.

[0025] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the physical constraint rules are extracted in the following specific steps.

[0026] Multi-source association mining is performed on cross-modal knowledge graphs to generate candidate rule sets;

[0027] The candidate rule set is statistically verified and logically integrated to generate physical constraint rules.

[0028] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the specific steps for obtaining the basic pattern template set are as follows:

[0029] The pattern library is semantically standardized, and the semantically standardized pattern library is semantically matched with the physical constraint rules to generate an association list;

[0030] Perform cross-validation and structured processing on the associated list to obtain a set of basic templates.

[0031] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the intelligent clothing model is constructed using the following specific steps.

[0032] Call and initialize the Transformer-PINN hybrid architecture to build the parameter parsing layer and geometry generation layer;

[0033] By using the mean squared error loss function, backpropagation and gradient updates are performed on the parameter analysis layer and the geometry generation layer to generate an intelligent clothing model.

[0034] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles according to the present invention, the specific steps for generating 3D simulated clothing are as follows:

[0035] The basic pattern template set is input into the intelligent clothing model. The parameter parsing layer decouples the structured features through the explicit constraint decoupling method to generate a set of physical constraint parameters.

[0036] The geometry generation layer dynamically adjusts the topology using a topology optimization method to output a 3D clothing geometry structure.

[0037] By geometrically fusing the set of physical constraint parameters with the 3D clothing geometry, a 3D simulated clothing is generated.

[0038] As a preferred embodiment of the pattern library-based 3D intelligent simulation generation method for clothing styles described in this invention, the specific steps for constructing the digital twin are as follows:

[0039] Physical constraints are bound to the basic template set and structural parameters are standardized to construct a digital twin framework;

[0040] Multimodal data fusion and behavioral rule mapping are performed on the digital twin framework to generate a digital twin.

[0041] As a preferred embodiment of the pattern library-based intelligent 3D simulation generation method for clothing styles according to the present invention, the specific steps for outputting the 3D simulation scheme of clothing styles are as follows:

[0042] Multiphysics coupling simulation of 3D simulated clothing based on digital twins is performed to generate a simulation result set;

[0043] The simulation results set is subjected to physical rule verification and geometric correction to generate a 3D simulation scheme for clothing styles.

[0044] The beneficial effects of this invention are as follows: by constructing a pattern library, deep fusion and intelligent association of multimodal data are realized, solving the problem of the disconnect between pattern rules and physical parameters; by simultaneously constructing an intelligent clothing model to decouple structured features and dynamically adjust the topology, closed-loop optimization of physical constraints and geometric generation is achieved, improving iteration efficiency and accuracy. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a method for intelligent 3D simulation generation of clothing styles based on a pattern library.

[0047] Figure 2 A flowchart for building a template library and generating a cross-modal knowledge graph.

[0048] Figure 3 A flowchart for generating 3D simulated clothing.

[0049] Figure 4 A flowchart for optimizing verification of digital twins. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for intelligent 3D simulation generation of clothing styles based on a pattern library, including the following steps:

[0054] S1. Collect multi-source clothing data and historical pattern data, perform multimodal feature alignment and intelligent knowledge distillation on historical pattern data, and build a pattern library;

[0055] S1.1 Multi-source apparel data includes market trend data, competitor analysis data, brand creative data, and user behavior data;

[0056] It should be noted that market trend data is obtained through web crawling of fashion platforms and industry reports, social media monitoring, and purchasing reports from third-party agencies; competitor analysis data is obtained through e-commerce web crawling of product details, obtaining sample manuals from trade shows, crawling competitor information from official websites and social media, and purchasing sales data from data service providers; brand creative data is obtained by extracting sketches and virtual samples from the pattern library exported from the internal PLM; and user behavior data is obtained through tracking tools to analyze website and APP behavior, collecting preferences through online questionnaires, and customer service inquiries.

[0057] S1.2 Historical pattern data includes structural parameter data, process rule data, and visual element data;

[0058] It should be noted that the structural parameter data, including the vertex coordinates of the facets, edge connection relationships, thickness, fabric elastic modulus, and seam density, are obtained by extracting historical pattern records; the process rule data comes from the company's process manual and technical documents; and the visual element data, including the patterns, colors, and decorative details of the garments, are extracted from the brand archives.

[0059] S1.3 Perform coordinate mapping and adjacency graph convolution on the structural parameter data to generate geometric feature vectors;

[0060] It should be noted that the vertex coordinates of the facets in the structural parameter data are mapped to a standardized 3D spatial mesh through affine transformation to unify the scale; an adjacency matrix is ​​constructed based on the edge connection relationship of the vertex coordinates of the facets, where the nodes of the adjacency matrix are the facet vertices and the edge weights of the adjacency matrix are the lengths of the connecting edges; the adjacency matrix is ​​traversed through the adjacent nodes of each node, and geometric features such as local curvature and adjacency density are obtained by acquiring the mean coordinates, variance of the normal vector, and average thickness of the adjacent vertices; the geometric features of all facet vertices are integrated into a fixed-dimensional geometric feature vector through average pooling.

[0061] S1.4 Perform semantic parsing and feature encoding on the process rule data to generate semantic feature vectors;

[0062] It should be noted that word segmentation and entity recognition are performed on the process rule data to extract structured semantic elements such as process type (e.g., "double needle sewing"), process parameters (e.g., "tensile force ≥ 50N"), and process purpose (e.g., "reinforce the joint") from the process rule data; one-hot encoding is performed on the process type to generate a type unique vector; the process parameters are normalized and mapped to a standard interval to generate a process parameter vector; and the process purpose is converted into a semantic embedding vector (e.g., "reinforcement → durability"); the type unique vector, process parameter vector, and semantic embedding vector are concatenated according to their dimensions to generate a fixed-dimensional semantic feature vector.

[0063] S1.5 Perform visual encoding, semantic annotation, and color decomposition on the visual element data to generate a composite visual feature vector;

[0064] It should be noted that edge detection extracts shape, texture, and contour features from element images (such as stripes and prints) in the visual element data, and integrates them to generate visual features. Shape features include line width and pattern period, texture features include roughness and repeating patterns, and contour features include the closure and curvature of the element image. The cosine similarity between visual features is used as feature similarity. The feature similarities are sorted, and visual features with similar feature similarity are grouped into the same cluster to classify element categories (such as "high-frequency thin stripes", "low-frequency thick stripes", "hand embroidery patterns"), generating semantic label vectors. The element images are converted from RGB color space to HSV space, and the dominant color (such as red H=0°), auxiliary color (such as beige H=30°), and accent color (such as gold H=45°) are extracted. The area ratio of each color is calculated to generate a color distribution vector. The visual features, semantic label vector, and color distribution vector are concatenated along a fixed dimension to generate a composite visual feature vector.

[0065] It should also be noted that the expression for obtaining cosine similarity is:

[0066] ;

[0067] in, Represents cosine similarity. Representing vectors (For example, visual features in this case) the first One element, Representing vectors The first in One element, This represents the total number of elements contained in each vector;

[0068] S1.6 Perform multimodal alignment, fusion, classification, and storage of geometric feature vectors, semantic feature vectors, and composite visual feature vectors to generate a pattern library;

[0069] It should be noted that geometric feature vectors, semantic feature vectors, and composite visual feature vectors are mapped to the same shared semantic space to obtain the cosine similarity between different feature vectors. The cosine similarity between different feature vectors is nonlinearly transformed through an attention mechanism to obtain the attention weights of different feature vectors. Different feature vectors are associated, aligned, and fused in the same shared semantic space to generate three-dimensional feature vectors. According to the clothing type (such as tops, bottoms, and dresses) and the corresponding season, the three-dimensional feature vectors are labeled with classification tags (such as "dress - spring / summer - fitted waist"). The three-dimensional feature vectors are associated, stored, and structured with the corresponding classification tags to generate a pattern library.

[0070] S2. Extract entity relationships and perform semantic mapping on multi-source clothing data, construct a cross-modal knowledge graph, and extract the physical constraint rules of the cross-modal knowledge graph;

[0071] S2.1 Perform entity identification and attribute extraction on market trend data and brand creative data to generate entity attribute pairs;

[0072] It should be noted that regular expressions are used to match high-frequency patterns in market trend data (e.g., the style term for the time phrase "Summer 2025" is "retro style" and the group term is "Generation Z"). Based on the high-frequency patterns in the domain, entity boundaries are located, and entity information in the market trend data is extracted. Dependency syntax is used to analyze the syntactic relationship between entity information and adjacent attribute words (e.g., "popularity index"). Attribute values ​​are generated by combining the suffix values ​​and descriptions of entity information (e.g., "85%", "French or American"), forming "entity-attribute-value" triples (e.g., <retro style, popularity index, 85%>).

[0073] The brand creative data is categorized and extracted. Specifically, for tabular data in the brand creative data, entities are extracted through column name tags (such as "process type"), where the attribute values ​​are other attribute contents of the corresponding column (such as "durability ≥500 friction cycles"). For semi-structured documents, entities and nested attributes are extracted through tag parsing. The categorized extraction results are integrated to generate a structured knowledge set (such as <double needle sewing, durability, ≥500 friction cycles>). The "entity-attribute-value" triples are integrated with the structured knowledge set to generate entity-attribute pairs.

[0074] S2.2 Perform cross-modal feature association on competitor analysis data to generate composite attribute sets;

[0075] It should be noted that, in extracting multimodal features from competitor analysis data, the text modality of competitor analysis data is extracted through word frequency statistics and sentiment analysis to extract related text (such as "slimming") and sentiment tendencies; the image modality is extracted through YOLO object detection and color quantization to extract pattern elements (such as "waist-slimming cut") and visual attributes; and the numerical attributes of competitor analysis data are extracted to generate structured data.

[0076] To establish cross-modal association rules, the specific operation involves aligning the multimodal features of competitors in the same competitor analysis data based on product ID and release timestamp; semantically mapping the pattern elements with visual attributes (e.g., "waist-cinching" in the text corresponds to the visual attribute "waist narrower by 10cm") to generate semantic association pairs; statistically analyzing the correlation between associated text, sentiment, and numerical attributes (e.g., the positive correlation between the frequency of the keyword "slimming" and "monthly sales of 5000+") to generate statistical association pairs; and establishing cross-modal association rules based on semantic and statistical association pairs, and combining the multimodal features of competitors according to the cross-modal association rules to generate a composite attribute set.

[0077] S2.4 Mine time-series preferences from user behavior data and perform collaborative filtering clustering to generate a structured preference knowledge set;

[0078] It should be noted that user behavior data undergoes preprocessing such as deduplication and missing value completion to complete data cleaning. The basic fields of the cleaned user behavior data are extracted, including user ID, timestamp, behavior type, and product ID. The basic fields are arranged in ascending order by timestamp to generate a user behavior sequence. The user behavior sequence is segmented using a fixed-time user time window, and the behavioral characteristics within each user time window are statistically analyzed, including browsing frequency, add-to-cart conversion rate, and purchase frequency. Combined with the time context, a time-series preference vector is generated (e.g., <User A, weekday evening, dress browsing frequency = 8 times / week>).

[0079] The system counts the number of times each user interacts with each product and assigns interaction weights based on behavior type (e.g., purchase = 3 points, add to cart = 2 points, and browse = 1 point), forming an interaction matrix with users as rows, products as columns, and interaction weights as values. Each row in the interaction matrix is ​​treated as a user vector, and the cosine similarity between user vectors is obtained. The cosine similarity is dynamically adjusted based on the timestamps of user behaviors (e.g., recent behavior weight 0.7, long-term 0.3) to obtain temporal similarity. Based on temporal preference vectors and temporal similarity, user groups and product categories are divided, such as into clusters like "high-frequency short-term preference type" and "low-frequency long-term stable type," and categories like "seasonal best-selling type" and "classic lasting type." The data is stored in a four-tuple format of "user cluster-time-product-preference intensity," generating a structured preference knowledge set.

[0080] S2.5. Embed entity attribute pairs, composite attribute sets, and structured preference knowledge sets into cross-modal knowledge to generate a cross-modal knowledge graph;

[0081] It should be noted that data standardization processing is performed on entity attribute pairs, composite attribute sets, and structured preference knowledge sets to unify the logical structure and data specifications, eliminate heterogeneity, align entities through semantic matching, align attributes through statistical co-occurrence frequency, and integrate the aligned entity attribute pairs, composite attribute sets, and structured preference knowledge sets to obtain "entity-relationship-attribute" triples. Using the entities and attributes in the "entity-relationship-attribute" triples as nodes and cross-modal association rules as edges, a cross-modal association graph is constructed. The cross-modal association graph is then integrated with the "entity-relationship-attribute" triples to generate a cross-modal knowledge graph.

[0082] S2.6 Perform multi-source association mining on the cross-modal knowledge graph to generate a candidate rule set;

[0083] It should be noted that a path search algorithm is used to traverse the cross-modal knowledge graph. Starting from any entity node in the cross-modal knowledge graph, the algorithm expands outward along relationship types such as "influence" and "preference" to obtain multi-hop association paths (such as "style entity → influence → pattern element entity → association → sales attribute"). The frequency of occurrence of all multi-hop association paths is statistically analyzed according to entity type, relationship type, and attribute type. The frequency of occurrence is used as the support of the multi-hop association path. The ratio of the actual number of times each entity node and attribute node co-occurs in the multi-hop association path to the total possible number of times they co-occurs is used as the co-occurrence confidence. All multi-hop association paths and their corresponding support and collinearity confidence are integrated to generate a candidate rule set.

[0084] S2.7 Perform statistical verification and logical integration on the candidate rule set to generate physical constraint rules;

[0085] It should be noted that multi-hop association paths with support higher than the mean support and co-occurrence confidence higher than the confidence threshold are selected from the candidate rule set to generate a theoretical rule set. Duplicate and redundant multi-hop association paths are removed from the theoretical rule set. The processed theoretical rule set is then structured to generate physical constraint rules.

[0086] It should also be noted that the confidence threshold is defined based on the co-occurrence confidence distribution of multi-hop associated paths, with an example value of (0.7, 1).

[0087] S3. Perform correlation reasoning and cross-validation on the physical constraint rules and the pattern library to obtain a set of basic pattern templates;

[0088] S3.1. Perform semantic standardization on the pattern library, and then perform semantic matching between the semantically standardized pattern library and the physical constraint rules to generate an association list;

[0089] It should be noted that the multimodal features in the pattern library, such as geometric feature vectors, semantic feature vectors, and composite visual feature vectors, are standardized to unify heterogeneous representations. The entity, attribute, and relation names (such as "style entity X", "pattern element Y", and "sales Z") in the physical constraint rules are parsed. The standardized pattern library is matched with the physical constraint rules (such as "retro style" matching "style entity X"). Entities in the pattern library that satisfy the physical constraint rules (such as "style-retro style") are identified as pattern entities. At the same time, the association between pattern entities and physical constraint rules is recorded and an association list is generated.

[0090] S3.2 Perform cross-validation and structured processing on the associated list to obtain a set of basic templates;

[0091] It should be noted that, based on the pattern entities in the association list and the matching physical constraint rules, the system cross-compares historical matching records in the pattern library to detect whether the pattern entities appear repeatedly in other physical constraint rules, thereby verifying the reliability of the association list and removing pattern entities that only match in a single physical constraint rule. The verified association list is then structured to generate a set of basic pattern templates, such as "Style - Retro Style - Waist-cinching Cut - Sales - Monthly Sales 5000+ - Constraint - Durability ≥ 1000 times".

[0092] S4. Input the basic pattern template set into the intelligent clothing model, the parameter parsing layer decouples the structured features, and the geometry generation layer dynamically adjusts the topology to generate a 3D simulated clothing.

[0093] S4.1 Call and initialize the Transformer-PINN hybrid architecture to build the parameter parsing layer and geometry generation layer;

[0094] It should be noted that in the PyTorch framework, the Transformer encoder is called via nn.TransformerEncoder, and a differential equation solver for a physical information neural network is embedded in the Transformer encoder. The input layer dimension is set according to the type and quantity of the basic pattern template set; the hidden layer is configured based on the Transformer's multi-head attention mechanism; a Gaussian integral node is connected after the Transformer encoder for numerical discretization to achieve high-precision fusion of multimodal features, and a nonlinear transformation is performed through the GELU function to enhance the ability to represent style and pattern element related features, thus completing the construction of the parameter analysis layer; the PINN ordinary differential equation solver is called via torchdiffeq.odeint, and the input dimension of the physical constraint equation is set according to the clothing physical parameters (such as elastic modulus and sheet thickness) in the basic pattern template set, and the number of network stacking layers is set according to the stability requirements of the ordinary differential equation solver; a gated recurrent network is connected after the ordinary differential equation solver to enhance geometric sequence features to capture the topological temporal characteristics of sheet connections, and a bidirectional gating mechanism is applied for forward and reverse geometric constraint analysis, thus completing the construction of the geometric generation layer.

[0095] S4.2. Using the mean squared error loss function, backpropagation and gradient update are performed on the parameter analysis layer and the geometry generation layer to generate an intelligent clothing model.

[0096] It should be noted that forward propagation is performed on the parameter parsing layer and the geometry generation layer to obtain the prediction error matrix. Backpropagation is then performed on the prediction error matrix to generate the gradient tensor of each layer. The gradient tensor of each layer is updated using the momentum gradient descent method to obtain the updated network weights. Based on the updated network weights, depthwise separable convolution is used to perform hyperparameter tuning on the parameter parsing layer and the geometry generation layer. The model is then stacked across layers through residual connections to output the constructed smart clothing model.

[0097] The basic template set is divided into a sample set, a training set, and a validation set based on data type (e.g., in a 7:2:1 ratio). The sample set is augmented using random affine transformation to generate augmented standard samples. The Adam optimizer is used to dynamically adjust the learning rate of the augmented standard samples, and early stopping is applied simultaneously to monitor the training process and obtain intermediate parameters. The prediction error of the intermediate parameters is quantified using the MSE loss function. Training is terminated when the validation set error exceeds the convergence threshold, and the trained smart clothing model is output.

[0098] It should also be noted that the convergence threshold is defined based on the rate of change of the prediction error on the validation set, and its value ranges from 0.0001 to 0.01.

[0099] S4.3 Input the basic pattern template set into the intelligent clothing model. The parameter parsing layer performs structured feature decoupling through explicit constraint decoupling method to generate a set of physical constraint parameters.

[0100] It should be noted that, based on physical constraints, the basic pattern template set is decomposed into independent feature subsets, including style feature subsets, pattern element feature subsets, and physical parameter feature subsets. Each independent feature subset is then standardized to generate a set of physical constraint parameters.

[0101] It should also be noted that the style feature subset includes style type (such as "retro style") and style keywords (such as "French style"). The steps to obtain the style feature subset are as follows: extract the "style type" field from the brand creative data of the basic pattern template set, combine it with market trend data to perform keyword matching, and filter out the style feature subset related to the physical constraint rules.

[0102] The pattern element feature subset contains specific garment details (such as "waist shrinkage rate of the fitted cut"). The pattern element feature subset is obtained by reading the "pattern element type" field from the visual element data of the basic pattern template set, extracting the contour coordinate points of the element image through edge detection, and obtaining the waist shrinkage rate; and then filtering by combining physical constraint rules (such as retaining fitted cuts with a shrinkage rate ≤30%) to generate a pattern element feature subset containing pattern element size and shape parameters.

[0103] The physical parameter feature subset includes physical properties such as fabric elastic modulus, sheet thickness, seam density, and tensile strength. The physical parameter feature subset is obtained by extracting process parameters from the process rule data of the basic pattern template set, obtaining the sheet thickness through the sheet vertex coordinates of the structural parameter data, and filtering the process parameters in combination with physical constraints (such as retaining fabric parameters with elastic modulus ≥ 200MPa) to generate the physical parameter feature subset.

[0104] S4.4 The geometry generation layer dynamically adjusts the topology using the topology optimization method to output the 3D clothing geometry.

[0105] It should be noted that the set of physical constraint parameters is input into the geometry generation layer. Through the topology optimization algorithm, the topology of the basic pattern template set is dynamically adjusted with the optimization objectives of the maximum deformation range under the elastic modulus limit, the seam strength within the tensile value range, and the waist shrinkage rate of the waist-tightening cut in the feature subset of the pattern elements. Specifically, under the premise of ensuring geometric continuity, the topology is made to conform to the style type (such as the smooth lines of a retro style) by iteratively optimizing vertex positions, merging or splitting redundant faces, and adjusting edge connection lengths, and outputting the 3D clothing geometry.

[0106] S4.5. Geometrically fuse the set of physical constraint parameters with the 3D clothing geometry to generate a 3D simulated clothing.

[0107] It should be noted that the physical parameters in the physical parameter feature subset are bound to the positional information of the 3D garment geometry through coordinate mapping. The physical parameters in the physical parameter feature subset include the sheet thickness, elastic modulus, and tensile strength. The deformation range of the sheet under stress is obtained based on the maximum deformation range under the elastic modulus limit and the seam strength within the tensile strength range. At the same time, the waist shrinkage rate of the waist-cinching cut is used as a geometric constraint to adjust the vertex coordinates of the waist sheet. The 3D garment geometry is optimized in detail by combining style type, generating a 3D simulated garment that integrates physical parameters, 3D garment geometry, and style features.

[0108] S5. Digitally model the basic pattern template set using the digital twin algorithm to construct a digital twin. Use the digital twin to perform physical simulation and optimization verification on the 3D simulated clothing, and output a 3D simulation scheme for the clothing style.

[0109] S5.1. Physically constrain and standardize the structural parameters of the basic template set to construct a digital twin framework;

[0110] It should be noted that, based on physical constraint rules, the physical parameters and geometric feature vectors in the basic pattern template set are aligned with entities through the "entity-relationship-attribute" triples of the cross-modal knowledge graph, establishing a bound geometric structure of "physical parameters-geometric structure-pattern rules"; the geometric parameters are standardized, including the vertex coordinates of the facets, the edge connection length, and the facet thickness; the pattern element parameters, including the waist shrinkage rate of the waist-cinching cut, are discretized according to the physical constraint rules; the style features are standardized and labeled using semantic tags (such as "line type: soft") of composite visual feature vectors; the bound geometric structure, the standardized geometric parameters, and the style features are fused through a semantic space (such as the "clothing type-season-style" semantic space) to generate a three-dimensional feature vector and label it with classification tags (such as "dress-spring / summer-retro waist-cinching style"), and stored together to form a digital twin framework.

[0111] S5.2 Perform multimodal data fusion and behavior rule mapping on the digital twin framework to generate a digital twin;

[0112] It should be noted that multimodal data is extracted from the digital twin framework, including geometric feature vectors, physical parameter feature subsets, and composite visual feature vectors. An attention mechanism is used to align and fuse the features of the multimodal data, integrating the vertex coordinates of the geometric structure, the material properties of the physical parameters, and the structural elements of the style features (such as the smooth lines of a retro style) into a unified multimodal representation vector. At the same time, based on the structured preference knowledge set and cross-modal association rules, user behavior preferences (such as "looking thinner") and market trend rules (such as "Generation Z is into retro elements") are mapped to the multimodal representation vector to generate a digital twin containing geometric structure, physical attributes, style features, and user behavior rules.

[0113] S5.3. Perform multiphysics coupling simulation on 3D simulated clothing based on digital twin, and generate simulation result set;

[0114] It should be noted that the input parameters required for multiphysics coupling are extracted from the digital twin, including geometric parameters, physical parameters, and boundary conditions (such as displacement constraints in fixed areas and load values ​​at stress points). The corresponding physical field type is selected based on the characteristics of the clothing material (such as cotton and silk). The mechanical field describes the stress-strain relationship using Hooke's Law, the thermodynamic field simulates temperature and humidity diffusion using the Fourier heat conduction equation, and the fluid field describes the force of airflow on the fabric using the simplified Navier-Stokes equations. The responses of each physical field are obtained step-by-step, and the coupling effect is iteratively updated. Specifically, the strain distribution of the mechanical field under initial load (such as human body stretching) is obtained, and the temperature boundary of the thermodynamic field is adjusted based on the strain distribution results. The temperature boundary of the thermodynamic field is input into the fluid field to obtain the pressure distribution of airflow on the fabric surface. The pressure distribution of the fluid field is applied to the mechanical field to correct the strain acquisition results (such as local stretching caused by airflow pressure). The simulation results of each physical field are then integrated to generate a simulation result set.

[0115] S5.4 Perform physical rule verification and geometric correction on the simulation result set to generate a 3D simulation scheme for clothing styles.

[0116] It should be noted that, based on physical constraint rules and user behavior rules, the simulation result set is verified. Specifically, this involves checking whether the maximum strain value of the mechanical field exceeds the elastic modulus limit, whether the extreme temperature value of the thermodynamic field meets the user scenario requirements (e.g., the local temperature ≥20℃ for "winter commuting"), and whether the air resistance coefficient of the fluid field matches the aesthetic appeal (e.g., the aerodynamic distribution of "skirt swaying"). Simultaneously, the line features of the geometric structure are verified and adjusted using style tags in the composite visual feature vector. Specifically, if the mechanical strain exceeds the elastic modulus limit, the thickness of the surface piece is adjusted or the edge connection relationship is optimized (e.g., shortening the side seam length); if the thermodynamic temperature exceeds the extreme temperature value, the vertex coordinates of the surface piece are modified; if the fluid field resistance coefficient deviates from the styling requirements (e.g., excessive resistance during skirt swaying), the surface piece outline is adjusted. The corrected geometric structure, physical parameters, and style features are then integrated to output a 3D simulation scheme for the clothing style.

[0117] It should also be noted that the elastic modulus limit is based on physical constraint rules and fabric type definition. An exemplary range is that the elastic modulus limit for high elastic knitted fabrics is ≥1.5MPa, and the elastic modulus limit for traditional denim fabrics is ≥200MPa.

[0118] User scenario requirements are based on a structured set of preference knowledge and specific spatiotemporal scenario definitions. For example, the range of values ​​is 18℃-22℃ for the armpit area of ​​clothing in the "winter commuting" scenario and less than 26℃ for the armpit area of ​​clothing in the "summer outdoor" scenario.

[0119] This embodiment also provides a computer device applicable to the 3D intelligent simulation generation method for clothing styles based on a pattern library, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the 3D intelligent simulation generation method for clothing styles based on a pattern library as proposed in the above embodiment.

[0120] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0121] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for intelligent simulation generation of clothing styles based on a pattern library as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0122] In summary, this invention achieves deep fusion and intelligent association of multimodal data by constructing a pattern library, thus solving the problem of the disconnect between pattern rules and physical parameters; and simultaneously constructs an intelligent clothing model to decouple structured features and dynamically adjust the topology, thereby realizing closed-loop optimization of physical constraints and geometric generation, and improving iteration efficiency and accuracy.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent 3D simulation generation of clothing styles based on a pattern library, characterized in that: include, Collect multi-source clothing data and historical pattern data, perform multimodal feature alignment and intelligent knowledge distillation on historical pattern data, and build a pattern library; Entity relationships are extracted and semantics are mapped from multi-source clothing data to construct a cross-modal knowledge graph and extract the physical constraint rules of the cross-modal knowledge graph. Perform correlation reasoning and cross-validation between physical constraint rules and pattern library to obtain a basic pattern template set; The basic pattern template set is input into the intelligent clothing model. The parameter parsing layer decouples the structured features, and the geometry generation layer dynamically adjusts the topology to generate a 3D simulated clothing. The basic pattern template set is digitally modeled using digital twin algorithms to construct a digital twin. The digital twin is then used to perform physical simulation and optimization verification on 3D simulated clothing, and output a 3D simulation scheme for clothing styles.

2. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 1, characterized in that: The multi-source apparel data includes market trend data, competitor analysis data, brand creative data, and user behavior data; The historical pattern data includes structural parameter data, process rule data, and visual element data.

3. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 2, characterized in that: The specific steps for building the template library are as follows. The structural parameter data is subjected to coordinate mapping and adjacency graph convolution to generate geometric feature vectors; Semantic parsing and feature encoding are performed on process rule data to generate semantic feature vectors; Visual element data is visually encoded, semantically labeled, and color decomposed to generate composite visual feature vectors. Geometric feature vectors, semantic feature vectors, and composite visual feature vectors are multimodal aligned, fused, and categorized for storage to generate a template library.

4. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 3, characterized in that: The specific steps for constructing the cross-modal knowledge graph are as follows. Perform entity identification and attribute extraction on market trend data and brand creative data to generate entity attribute pairs; Perform cross-modal feature association on competitor analysis data to generate composite attribute sets; Mining time-series preferences from user behavior data and performing collaborative filtering clustering to generate a structured preference knowledge set; Cross-modal knowledge is embedded into entity attribute pairs, composite attribute sets, and structured preference knowledge sets to generate a cross-modal knowledge graph.

5. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 4, characterized in that: The specific steps for extracting the physical constraint rules are as follows. Multi-source association mining is performed on cross-modal knowledge graphs to generate candidate rule sets; The candidate rule set is statistically verified and logically integrated to generate physical constraint rules.

6. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 5, characterized in that: The specific steps for obtaining the basic template set are as follows. The pattern library is semantically standardized, and the semantically standardized pattern library is semantically matched with the physical constraint rules to generate an association list; Perform cross-validation and structured processing on the associated list to obtain a set of basic templates.

7. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 6, characterized in that: The specific construction steps for the intelligent clothing model are as follows. Call and initialize the Transformer-PINN hybrid architecture to build the parameter parsing layer and geometry generation layer; By using the mean squared error loss function, backpropagation and gradient updates are performed on the parameter analysis layer and the geometry generation layer to generate an intelligent clothing model.

8. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 7, characterized in that: The specific steps for generating the 3D simulated clothing are as follows. The basic pattern template set is input into the intelligent clothing model. The parameter parsing layer decouples the structured features through the explicit constraint decoupling method to generate a set of physical constraint parameters. The geometry generation layer dynamically adjusts the topology using a topology optimization method to output a 3D clothing geometry structure. By geometrically fusing the set of physical constraint parameters with the 3D clothing geometry, a 3D simulated clothing is generated.

9. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 8, characterized in that: The specific steps for constructing a digital twin are as follows: Physical constraints are bound to the basic template set and structural parameters are standardized to construct a digital twin framework; Multimodal data fusion and behavioral rule mapping are performed on the digital twin framework to generate a digital twin.

10. The method for intelligent 3D simulation generation of clothing styles based on a pattern library as described in claim 9, characterized in that: The specific steps for the output 3D simulation solution for clothing styles are as follows. Multiphysics coupling simulation of 3D simulated clothing based on digital twins is performed to generate a simulation result set; The simulation results set is subjected to physical rule verification and geometric correction to generate a 3D simulation scheme for clothing styles.

Images