Bionic cocoon shell one-piece vamp virtual sample generation system

Abstract

The present application relates to the technical field of shoe making method, and proposes a cocoon bionic one-piece vamp virtual sample generation system for balancing the support, breathability and lightweight characteristics of one-piece vamp and quickly virtually generating one-piece vamp to shorten the research and development cycle, comprising the following steps: S1, constructing cocoon bionic process performance coupling multi-dimensional data set; S2, training cocoon bionic customized generation model; S3, generating virtual sample image under cocoon bionic structure dynamic constraint; S4, two-dimensional evaluation and cocoon bionic parameter optimization deduction.

Description

Technical Field

[0001] This invention relates to the field of shoemaking methods, and in particular to a cocoon-shell biomimetic one-piece shoe upper virtual sample generation system. Background Technology

[0002] Against the backdrop of the sportswear industry's continuous upgrade towards lightweight, integrated, and functional designs, one-piece uppers have become the mainstream upper solution for high-end running shoes and casual shoes due to their advantages such as seamless molding, high fit, and simplified processes. The cocoon-like biomimetic structure, utilizing the lightweight, high-toughness, breathability, and structural support characteristics of natural silkworm cocoons, has been incorporated into the one-piece upper design. By flexibly adjusting the spacing, path offset, and local overlapping bonding of the hot-melt thread loops, it is possible to balance support, breathability, and lightweight properties on a single fabric base, overcoming the bottleneck of traditional woven and knitted uppers where support and breathability are difficult to achieve simultaneously.

[0003] Current research and development of cocoon-shell biomimetic one-piece shoe uppers still relies on repeated physical prototyping, resulting in long development cycles, high material and labor costs, and difficulty in quickly balancing support and breathability. Upper molding is influenced by multiple parameters, including the distance between wrapping loops, the width of overlapping areas, and the specifications of hot-melt wires. Parameter adjustments lack quantitative basis, relying on experience-based design, and the microstructure cannot be precisely correlated with macroscopic performance. Traditional virtual sample generation only focuses on appearance rendering, without incorporating the constraints of cocoon-shell biomimetic processes, leading to significant deviations between the generated effect and the actual molding, making it unsuitable for process verification and performance prediction. Existing technologies lack a dual-dimensional evaluation system for visual and process performance, do not support parameter reverse engineering and dynamic optimization, and struggle to form a digital closed loop of design, generation, evaluation, and iteration, severely restricting the efficiency of shoe upper development and product consistency. Summary of the Invention

[0004] Therefore, in response to the above problems, this invention proposes a virtual sample generation system for cocoon-shaped bionic one-piece shoe uppers, which solves the problem that the current research and development of cocoon-shaped bionic one-piece shoe uppers still relies on repeated physical prototyping.

[0005] To solve the above-mentioned technical problems, the solution adopted by the present invention is as follows: A cocoon-shell biomimetic one-piece shoe upper virtual sample generation system, characterized in that the system establishment includes the following steps: S1. Construct a multi-dimensional dataset of cocoon shell biomimetic process performance coupling: Collect high-definition visual images and micro-winding structure data of cocoon shell biomimetic one-piece shoe upper samples, simultaneously measure the cocoon shell biomimetic forming process parameters, micro-structural features and macro-performance indicators of the samples, perform cocoon shell biomimetic process feature-guided data augmentation processing on the high-definition visual images, generate standardized labels for the process performance coupling of the samples, and construct a multi-dimensional dataset of process performance coupling that associates cocoon shell biomimetic forming process parameters, micro-winding structure features, visual images and macro-performance indicators. S2. Training the customized generative model for cocoon-like bionics: Based on the open-source stable diffusion model, an improvement is made on its U-Net architecture to construct a multi-scale attention-enhanced U-Net architecture; through hierarchical low-rank adaptive fine-tuning technology, the model is adapted to the micro-layered ring structure and macro-winding path characteristics of the cocoon-like one-piece shoe upper, realizing the accurate mapping from process parameters and performance requirements to virtual sample images. In the upsampling path of the U-Net architecture, a microstructure feature decoding branch is added. The micro-entanglement structure features are used as auxiliary supervision signals. The latent features are decoded into microstructure parameter predictions through additional convolutional layers. The mean squared error loss is calculated with the measured microstructure features and used as a component of the total loss function. S3. Generating virtual sample images under dynamic constraints of cocoon shell biomimetic structure: Receive the range of cocoon shell biomimetic process parameters or natural language description input by the user, parse and convert it into a standardized cocoon shell biomimetic generative process parameter and performance requirement joint input vector, pre-build a cocoon shell biomimetic one-piece shoe upper process feature knowledge base and generate a process feature constraint threshold set, input the joint input vector into the trained customized generation model, and perform real-time verification and dynamic correction of the process features of winding loop path, loop density, overlapping width, and functional partition during the model generation process to generate a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image; S4. Dual-dimensional evaluation and cocoon shell biomimetic parameter optimization deduction: Establish a dual-dimensional evaluation system of visual features and process performance, quantitatively evaluate the generated virtual sample images, construct a parameter performance correlation model based on the coupling of cocoon shell biomimetic process performance with multi-dimensional datasets, quantitatively analyze the influence of cocoon shell biomimetic process parameters on the macroscopic performance indicators of the shoe upper through reverse deduction of process performance, and output optimization suggestions for cocoon shell biomimetic process parameters based on evaluation results and deduction results, forming a digital closed loop of generation, evaluation and optimization.

[0006] A further improvement is made to the construction of the multi-dimensional dataset coupling the biomimetic process performance of the cocoon shell described in step S1, specifically including: S11. High-precision surface scanning equipment and micro laser scanning equipment are used to collect high-definition visual images and micro winding structure data of the cocoon shell bionic one-piece shoe upper sample, respectively. The cocoon shell bionic one-piece shoe upper sample is formed by continuous winding and stacking of molten hot melt wire. S12. Determine and structure the biomimetic process parameters, microstructural characteristics and macroscopic performance indicators of the cocoon shell samples. The parameters for the biomimetic forming process of the cocoon shell include: the offset distance of the center path of the winding and overlapping ring group, the average width of the overlapping part, the ring spacing, the thickness of the hot melt wire, the material of the hot melt wire, the shape of the ring, and the forming path type; Microstructural features include: overlapping ring method, porosity, hot melt wire twist, and the density of the wound rings; Macroscopic performance indicators include: lightweight coefficient, toughness, abrasion resistance, breathability, support, fit, and durability; S13. Perform data augmentation operations guided by the biomimetic process features of the cocoon shell on the high-definition visual image. The data augmentation operations include rotation, flipping, brightness adjustment, and local stacking density perturbation and path offset operations to simulate deviations in the biomimetic cocoon shell forming process. The formal expression of the data augmentation operations is: ; In the formula, R θ For rotation and flip operators, α and β are the brightness scaling factor and brightness offset factor, respectively, and P δ Here, δ is the process deviation simulation operator, γ is the deviation coefficient, γ is the process deviation weighting coefficient, and N(0,σ²) is Gaussian noise. S14. The data-enhanced sample generates a standardized label for the coupling of cocoon shell biomimetic process performance. The standardized label contains the correlation information of cocoon shell biomimetic molding process parameters, micro-entanglement structure features, and macro-performance indicators, realizing deep coupling of image, parameters, and performance.

[0007] The further improvement is that the multi-scale attention-enhanced U-Net architecture and hierarchical low-rank adaptive fine-tuning technique described in step S2 specifically include: S21. In the traditional U-Net architecture, a cocoon shell biomimetic process feature attention module is added to each stage of the downsampling and upsampling paths, and a cross-scale winding structure feature fusion module is added to the intermediate bottleneck layer. The cocoon shell biomimetic process feature attention module is composed of a channel attention submodule and a spatial attention submodule connected in series. The channel attention submodule learns the channel weights of winding loops, overlapping parts, and hot-melt wire specifications through global average pooling and fully connected layers. The spatial attention submodule strengthens the spatial positioning of the cocoon shell biomimetic winding path and loop distribution through two-dimensional pooling and convolutional layers, and strengthens the extraction of core process features such as winding loop path, loop density, and overlapping part width through dual weight allocation. The cross-scale winding structure feature fusion module adopts an attention weighted fusion mechanism, which assigns differentiated weights to feature maps of different resolutions and adds them element by element to achieve accurate fusion learning of microscopic loop texture and macroscopic cocoon shell biomimetic contour. S22. The hierarchical low-rank adaptive fine-tuning technique designs differentiated low-rank matrices based on different levels of the multi-scale attention-enhanced U-Net architecture: the downsampling path is configured with a low-rank matrix of rank r = 6–8 to adapt to the complex mapping learning of macroscopic winding paths in cocoon-like bionic architecture; the upsampling path is configured with a low-rank matrix of rank r = 3–5 to adapt to the fine representation of microscopic overlapping textures; the intermediate layer is configured with a low-rank matrix of rank r = 8–10 to adapt to the coupled learning of cross-scale winding structural features; the original weights of the architecture are divided into downsampling weight reorganization, intermediate layer weight reorganization, and upsampling weight reorganization according to the hierarchy. During the forward calculation of the original weights at each level, they are updated to the sum of the original weights and the corrected value of the low-rank matrix, i.e.: ; ; In the above formula, The original weights for each level, This is the correction value for the low-rank matrix. These are the scaling factors for each level. The rank of a low-rank matrix. , The matrix is ​​a trainable low-rank matrix; during fine-tuning, the original weights are frozen, and only the parameters of the low-rank matrix at each level are updated via backpropagation. S23. The total loss function L is a weighted sum of visual loss Lvis, cocoon shell biomimetic process feature loss Lpro, and performance coupling loss Lper, and its expression is: ; In the formula, The weighting coefficients and Lvis is the sum of pixel-level reconstruction loss and perception loss, Lpro is the cosine similarity loss of cocoon shell biomimetic process feature matching, and Lper is the mean square error loss of breathability and support performance prediction. S24. After model training is complete, a dynamic weight fusion strategy is adopted. The user-input priorities of support, breathability, and lightweight requirements are expressed as percentages. The system normalizes the priority vector and converts it into a fusion ratio λ between the low-rank matrix correction value of each level and the original weights through a preset mapping function. The value of λ ranges from 0.6 to 0.9. The fusion ratio is calculated using the following formula: ; in, Based on the fusion ratio, Let the i-th performance requirement be prioritized. These are the empirical coefficients for the corresponding performance terms.

[0008] A further improvement is made to the generation of the virtual sample image under the dynamic constraints of the cocoon-like biomimetic structure in step S3, specifically including: S31. Construct a knowledge base for the process features of cocoon-shell biomimetic one-piece shoe upper. The knowledge base includes the process feature thresholds of cocoon-shell biomimetic winding molding, the matching rules of winding loop path and loop density, the correspondence between overlapping width and performance, and the correspondence between functional partitions and macroscopic performance indicators. S32. The standardized cocoon shell biomimetic process parameters and macroscopic performance index requirements are combined into an input vector and converted into a set of process feature constraint thresholds, including upper and lower limits of stacked ring density, curvature range of winding path, overlap width constraint, and functional partition position and size constraint. S33. Input the joint input vector into the trained cocoon shell biomimetic customized generation model. After the denoising operation at each time step of the model, extract the cocoon shell biomimetic process feature information in the latent space and compare it with the process feature constraint threshold set in real time. If the process feature information exceeds the constraint threshold set, generate a constraint correction vector and inject it into the denoising process of the next time step to realize the dynamic correction of winding loops, overlapping parts, and hot melt wire specifications. S34. Perform local enhancement processing on the biomimetic functional zoning features of the cocoon shell, strengthen the overlapping distribution and overlapping features of the support zone, breathable zone, and lightweight zone, ensure the consistency of functional zoning with the biomimetic molding process parameters and macroscopic performance index requirements of the cocoon shell, and complete the generation of a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image.

[0009] A further improvement is made to the reverse inference of the dual-dimensional evaluation and cocoon shell biomimetic process performance described in step S4, specifically including: S41. The visual feature evaluation indicators include peak signal-to-noise ratio, structural similarity index, and color similarity based on improved K-means clustering. The improved K-means clustering optimizes the main color extraction process by introducing cocoon-shell biomimetic process feature weights: First, process feature vectors of each sample are extracted from the cocoon-shell biomimetic process performance coupled multi-dimensional dataset. The correlation between each process feature and color distribution is statistically analyzed, and the main color influence weight of each process feature is calculated. Then, during the K-means clustering iteration process, the Euclidean distance between the pixel and the cluster center is multiplied by the corresponding process feature weight coefficient, so that the clustering process is tilted towards the color region strongly correlated with the key process features, thereby improving the process sensitivity of main color extraction and thus improving the correlation between color similarity evaluation and process performance. S42. The evaluation indicators of the process performance include the fidelity of the cocoon shell bionic process features and the accuracy of performance prediction. The accuracy of performance prediction is the quantitative matching degree between the macro performance indicators such as breathability, support, and lightweight predicted by the cocoon shell bionic process performance coupling model and the user's performance requirements. The process performance coupling model is a lightweight gradient boosting regression model independent of the generation model. It is trained based on the cocoon shell bionic process performance coupling multi-dimensional dataset constructed in step S1. The input is the cocoon shell bionic generative process parameters, and the output is the macro performance indicators of the shoe upper. S43. A gradient boosting tree algorithm is used to construct a cocoon shell biomimetic parameter performance correlation model. Based on the cocoon shell biomimetic process performance coupled multi-dimensional dataset, the single factor influence and interactive factor influence of the offset distance of the winding and overlapping ring group, the overlap width, the ring spacing, and the hot melt wire specification on the macroscopic performance indicators of the shoe upper are quantitatively analyzed, and a cocoon shell biomimetic parameter sensitivity heat map is generated. S44. If the dual-dimensional evaluation results of the virtual sample image do not meet the user's needs, the system outputs targeted optimization suggestions for the biomimetic modeling process parameters of the cocoon shell based on the biomimetic parameter sensitivity heatmap. The suggestions specify the direction of parameter adjustment, the adjustment range, and the expected improvement in macroscopic performance indicators. The optimization suggestions are output in structured text format, including parameter name, current value, suggested adjustment value, adjustment range, expected performance indicators to be improved, and improvement range. The user modifies the input parameters in the system interface based on the optimization suggestions, and the model completes the secondary generation. The system also supports a "one-click optimization" mode, which automatically updates the input based on the suggested values ​​and triggers regeneration, forming a digital closed loop of generation, evaluation, and optimization.

[0010] By adopting the aforementioned technical solution, the beneficial effects of the present invention are: 1. The cocoon-shell biomimetic one-piece upper is formed by continuously molten hot-melt wire wound at equal intervals along the outer contour of the upper, consisting of multiple sets of wound overlapping loops, with overlapping portions of adjacent loops forming a fixed bond. This structure has the following beneficial effects: Integrated molding simplifies the process and reduces costs: One-piece winding eliminates cumbersome processes such as multi-piece cutting, splicing, and sewing, significantly shortening the production cycle; no fabric cutting waste is produced, reducing raw material waste. It uses molten hot-melt wire for winding, automatic bonding and fixing, and subsequent overall hot pressing, resulting in high molding precision and structural stability.

[0011] Seamless construction enhances comfort and durability: Eliminating the foreign object feeling caused by seams and glue layers, the overall upper is as thin and flexible as a cocoon; at the same time, it avoids wear and tear and seam opening at the seams, significantly improving the durability and lifespan of the upper.

[0012] Bionic contour fit for excellent fit: The wrapping rings are distributed at equal intervals inward along the outer contour of the shoe upper, with the central path consistent with the contour of the shoe upper. Utilizing the three-dimensional wrapping and uniform force characteristics of the bionic cocoon shell, the overall contour and support structure of the shoe upper are precisely controlled, perfectly fitting the curve of the foot, so that the feet are not stuffy or pressured when worn.

[0013] Cocoon-level balance between breathability and support: Composed of continuously overlapping rings, the gaps between the rings precisely replicate the micro-breathing channels of a cocoon, achieving efficient breathability while ensuring support; adjacent wrapped overlapping rings are fixed and bonded together through overlapping parts, combined with the characteristics of heat-melting thread material, achieving a cocoon-level balance between breathability and support, solving the pain point of traditional shoe uppers that are difficult to balance breathability and support.

[0014] Adjustable parameters allow for quick adaptation to various scenarios: the offset distance of the wrapping and overlapping rings, the width of the overlapping part, the distance between the rings, the thickness of the wires, and the material can all be flexibly adjusted. Different performance uppers such as balanced comfort, lightweight and breathable, and strong support and durability can be quickly customized to meet the wearing needs of various scenarios such as leisure and sports.

[0015] 2. This invention also proposes a virtual sample generation system that matches the aforementioned cocoon-shell biomimetic one-piece shoe upper. The system is constructed around the multi-dimensional coupling relationship of "process-performance-vision" of the cocoon-shell biomimetic structure, realizing the accurate mapping from process parameters and performance requirements to high-fidelity virtual sample images, and has the following beneficial effects: This technology revolutionizes the traditional trial-and-error development model, significantly shortening the R&D cycle and reducing costs. Based on a deep learning model, it achieves precise mapping between molding process parameters, performance requirements, and virtual sample images, eliminating the need for repeated physical prototyping. Real-world application verification shows that the R&D cycle for a single one-piece shoe upper has been reduced from the traditional 15 days to 5-7 days; virtual sample generation can replace more than 80% of the physical prototyping process, resulting in a cost reduction of 40%-50%.

[0016] A multi-dimensional dataset coupling process performance is constructed to lay the foundation for model learning. Molding process parameters (offset distance, overlap width, loop spacing, wire specifications, etc.), microstructural features (porosity, hot-melt wire twist, overlapping method), and macroscopic performance indicators (breathability, support, lightweight coefficient, etc.) are deeply coupled with visual images, breaking through the single-dimensional limitation of traditional datasets that only associate images with parameters. Based on this dataset, the model can accurately capture and reproduce the core process features of the cocoon-shell biomimetic one-piece shoe upper (such as winding path, overlapping density, overlapping part distribution, functional zoning, etc.), possessing strong process feature reproduction accuracy and performance prediction capabilities.

[0017] The multi-scale attention-enhanced U-Net architecture and hierarchical low-rank adaptive fine-tuning precisely adapt to the dual feature generation of "macro-scale process + micro-structure". At each stage of the U-Net downsampling and upsampling paths, a cocoon-like biomimetic process feature attention module (channel attention + spatial attention) is added to strengthen the extraction of core process features such as winding loop paths, loop density, and overlap width. A cross-scale winding structure feature fusion module is added to the intermediate bottleneck layer to achieve accurate fusion learning of micro-loop textures and macro-cocoon-like biomimetic contours. The hierarchical low-rank adaptive fine-tuning technology designs differentiated ranks according to different architecture levels: the downsampling path (r=6-8) adapts to the complex mapping of macro-winding paths, the upsampling path (r=3-5) adapts to the fine expression of micro-loop textures, and the intermediate layer (r=8-10) adapts to cross-scale feature coupling learning. A microstructure feature decoding branch is added to the upsampling path, using microstructure features such as porosity and hot-melt wire twist as auxiliary supervision signals. This ensures that the latent space of the generated model contains microstructure information, guaranteeing that the generated image is consistent with the actual molding at both the visual and microstructure levels. A dynamic weighted fusion strategy is adopted, dynamically adjusting the fusion ratio (0.6-0.9) based on the user's input priority of support, breathability, and lightweight requirements (in percentage form) through a preset mapping function. This enables customized generation of different types of shoe uppers, such as balanced comfort, lightweight breathability, and strong support durability.

[0018] A dynamic constraint mechanism for process features is introduced during the generation process to ensure engineering feasibility. A knowledge base of process features for a cocoon-like biomimetic one-piece shoe upper is constructed (including process feature thresholds, winding path and density matching rules, the correspondence between overlapping width and performance, and the correspondence between functional zones and performance). After denoising at each time step of the model, process feature information of the latent space is extracted, compared with the constraint threshold set in real time, and dynamically corrected. At the same time, the local feature expression of the support area, breathable area, and lightweight area is strengthened. This fundamentally avoids generating results with infeasible processes, ensuring that the virtual sample is highly consistent with the actual production process parameters and performance requirements, directly guiding actual production implementation.

[0019] A closed loop of dual-dimensional evaluation (visual features + process performance) and parameter back-drafting is established to achieve data-driven iterative optimization. Visual feature evaluation includes peak signal-to-noise ratio, structural similarity index, and color similarity based on improved K-means clustering (optimizing the primary color extraction process by introducing biomimetic process feature weights from cocoon shells). Process performance evaluation includes the fidelity of biomimetic process features from cocoon shells (quantitative matching degree between the extracted interlocking loops, overlapping parts, and hot-melt wire specifications and target parameters); and performance prediction accuracy (predicting indicators such as air permeability, support, and lightweighting using a lightweight gradient boosting regression model independent of the generation model). A gradient boosting tree is used to construct a parameter performance correlation model, quantifying the single and interactive influence of each process parameter on macroscopic performance, and generating a parameter sensitivity heatmap. If the evaluation results do not meet requirements, the system outputs targeted parameter optimization suggestions in structured text format (parameter name, current value, suggested adjustment value, adjustment range, and expected improvement effect), supporting manual modification by users or automatic iteration with "one-click optimization," forming a digital closed loop of "generation-evaluation-optimization." Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the three-dimensional structure of the cocoon-shell biomimetic one-piece shoe upper in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the unfolded planar structure of the cocoon-shell biomimetic one-piece shoe upper in an embodiment of the present invention.

[0022] Figure 3 This is a diagram showing the center path trajectory of each winding and overlapping ring group in the cocoon-shell biomimetic one-piece shoe upper in this embodiment of the invention.

[0023] Figure 4 yes Figure 2 Enlarged view of the local structure at point A in the middle. Detailed Implementation

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0025] refer to Figures 1 to 4 The present invention discloses a cocoon-shell biomimetic one-piece shoe upper virtual sample generation system.

[0026] A cocoon-shell biomimetic one-piece shoe upper includes a shoe upper body 10, which is composed of multiple sets of winding and overlapping rings 11. The winding and overlapping rings 11 adopt a cocoon-shell biomimetic layered winding structure. The center path 12 of the winding and overlapping rings 11 is consistent with the outer contour shape of the shoe upper body 10. Several winding and overlapping rings 11 are distributed inward at equal intervals along the outer contour of the shoe upper body 10. Adjacent winding and overlapping rings 11 form a fixed and adhered overlapping part 13. The offset distance between the center paths 12 of adjacent winding and overlapping rings 11 is 10mm, and the average width of the overlapping part 13 is 5mm. The winding and overlapping rings 11 are composed of several rings 14 continuously stacked along the center path 12. The ring spacing between adjacent rings 14 is 0.2mm. The rings 14 are circular or elliptical, and the stacking shape of the rings 14 resembles a biomimetic cocoon-shell silk loop layered structure. The material of the annular ring 14 is a hot melt wire, specifically a TPU hot melt wire, and the thickness of the hot melt wire is 100D.

[0027] A virtual sample generation system for a cocoon-shell biomimetic one-piece shoe upper, the system establishment includes steps S1-S4: S1. Construct a multi-dimensional dataset of cocoon shell biomimetic process performance coupling: Collect high-definition visual images and micro-winding structure data of cocoon shell biomimetic one-piece shoe upper samples, simultaneously measure the cocoon shell biomimetic forming process parameters, micro-structural features and macro-performance indicators of the samples, perform cocoon shell biomimetic process feature-guided data augmentation processing on the high-definition visual images, generate standardized labels for the process performance coupling of the samples, and construct a multi-dimensional dataset of process performance coupling that associates cocoon shell biomimetic forming process parameters, micro-winding structure features, visual images and macro-performance indicators. The construction of the multi-dimensional dataset coupling the performance of the cocoon shell biomimetic process specifically includes: S11. High-precision surface scanning equipment and micro laser scanning equipment are used to collect high-definition visual images and micro winding structure data of the cocoon shell bionic one-piece shoe upper sample, respectively. The cocoon shell bionic one-piece shoe upper sample is formed by continuous winding and stacking of molten hot melt wire. S12. Determine and structure the biomimetic process parameters, microstructural characteristics and macroscopic performance indicators of the cocoon shell samples. The parameters for the biomimetic cocoon shell forming process include: the offset distance of the center path of the winding and overlapping rings, the average width of the overlapping part, the ring spacing, the thickness of the hot-melt wire, the material of the hot-melt wire, the shape of the rings, and the forming path type; the microstructural characteristics include: the overlapping method of the rings, porosity, the twist of the hot-melt wire, and the density of the winding and overlapping rings; the macroscopic performance indicators include: lightweight coefficient, toughness, wear resistance, air permeability, support, fit, and durability; S13. Perform data augmentation operations guided by the biomimetic process features of the cocoon shell on the high-definition visual image. The data augmentation operations include rotation, flipping, brightness adjustment, and local stacking density perturbation and path offset operations to simulate deviations in the biomimetic cocoon shell forming process. The formal expression of the data augmentation operations is: ; In the formula, R θ For rotation and flip operators, α and β are the brightness scaling factor and brightness offset factor, respectively, and P δ Here, δ is the process deviation simulation operator, γ is the deviation coefficient, γ is the process deviation weighting coefficient, and N(0,σ²) is Gaussian noise. S14. The data-enhanced sample generates a standardized label for the coupling of cocoon shell biomimetic process performance. The standardized label contains the correlation information of cocoon shell biomimetic molding process parameters, micro-entanglement structure features, and macro-performance indicators, realizing deep coupling of image, parameters, and performance.

[0028] S2. Training the customized biomimetic generative model for cocoon shells: Based on open-source stable diffusion models (such as StableDiffusion v2.1, which is pre-trained on a large-scale general image dataset and has strong texture and structure generation capabilities), we improve upon its U-Net architecture to construct a multi-scale attention-enhanced U-Net architecture; through hierarchical low-rank adaptive fine-tuning technology, we adapt the model to the micro-layered ring structure and macro-winding path features of the one-piece biomimetic shoe upper of cocoon shells, and achieve accurate mapping from process parameters and performance requirements to virtual sample images; In the upsampling path of the U-Net architecture, a microstructure feature decoding branch is added. Microstructure features such as porosity, hot-melt wire twist, and overlapping ring methods are used as auxiliary supervision signals. These latent features are decoded into predicted microstructure parameters through an additional convolutional layer. The mean squared error loss is calculated by comparing these predicted and actual microstructure features, and this error is incorporated into the total loss function. This mechanism ensures that the latent space of the generative model contains microstructure information, guaranteeing that the generated virtual sample image not only visually conforms to the process characteristics but also maintains consistency with the actual molded structure at the microstructure level.

[0029] The multi-scale attention-enhanced U-Net architecture and hierarchical low-rank adaptive fine-tuning technique specifically include: S21. In the traditional U-Net architecture, a cocoon shell biomimetic process feature attention module is added to each stage of the downsampling and upsampling paths, and a cross-scale winding structure feature fusion module is added to the intermediate bottleneck layer. The cocoon shell biomimetic process feature attention module is composed of a channel attention submodule and a spatial attention submodule connected in series. The channel attention submodule learns the channel weights of winding loops, overlapping parts, and hot-melt wire specifications through global average pooling and fully connected layers. The spatial attention submodule strengthens the spatial positioning of the cocoon shell biomimetic winding path and loop distribution through two-dimensional pooling and convolutional layers, and strengthens the extraction of core process features such as winding loop path, loop density, and overlapping part width through dual weight allocation. The cross-scale winding structure feature fusion module adopts an attention weighted fusion mechanism, which assigns differentiated weights to feature maps of different resolutions and adds them element by element to achieve accurate fusion learning of microscopic loop texture and macroscopic cocoon shell biomimetic contour. S22. The hierarchical low-rank adaptive fine-tuning technique designs differentiated low-rank matrices based on different levels of the multi-scale attention-enhanced U-Net architecture: the downsampling path is configured with a low-rank matrix of rank r = 6–8 to adapt to the complex mapping learning of macroscopic winding paths in cocoon-like bionic architecture; the upsampling path is configured with a low-rank matrix of rank r = 3–5 to adapt to the fine representation of microscopic overlapping textures; the intermediate layer is configured with a low-rank matrix of rank r = 8–10 to adapt to the coupled learning of cross-scale winding structural features; the original weights of the architecture are divided into downsampling weight reorganization, intermediate layer weight reorganization, and upsampling weight reorganization according to the hierarchy. During the forward calculation of the original weights at each level, they are updated to the sum of the original weights and the corrected value of the low-rank matrix, i.e.: ; ; In the above formula, The original weights for each level, This is the correction value for the low-rank matrix. These are the scaling factors for each level. The rank of a low-rank matrix. , The matrix is ​​a trainable low-rank matrix; during fine-tuning, the original weights are frozen, and only the parameters of the low-rank matrix at each level are updated via backpropagation. The improved U-Net architecture, its original weights Initialize the weights to pre-trained values ​​from an open-source stable diffusion model (such as StableDiffusion v2.1). During fine-tuning, freeze these pre-trained weights and only adjust the newly added low-rank matrix parameters at each level. , Random initialization and training are performed.

[0030] S23. The total loss function L is a weighted sum of visual loss Lvis, cocoon shell biomimetic process feature loss Lpro, and performance coupling loss Lper, and its expression is: ; In the formula, The weighting coefficients and Lvis is the sum of pixel-level reconstruction loss and perception loss, Lpro is the cosine similarity loss of cocoon shell biomimetic process feature matching, and Lper is the mean square error loss of breathability and support performance prediction. S24. After model training is complete, a dynamic weight fusion strategy is adopted. The user-input priorities for support, breathability, and lightweight are expressed as percentages (e.g., support 50%, breathability 30%, lightweight 20%). The system normalizes the priority vector and converts it into a fusion ratio λ between the low-rank matrix correction value of each level and the original weights through a preset mapping function. The value of λ ranges from 0.6 to 0.9. The fusion ratio is calculated using the following formula: ; in, Based on the fusion ratio, Let the i-th performance requirement be prioritized. The empirical coefficients for the corresponding performance items (obtained from statistical analysis of the dataset) are dynamically adjusted to enable the model to customize the generation of different types of cocoon-shaped bionic uppers, such as balanced comfort, lightweight breathability, and strong support and durability.

[0031] S3. Generating virtual sample images under dynamic constraints of cocoon shell biomimetic structure: Receive the range of cocoon shell biomimetic process parameters or natural language description input by the user, parse and convert it into a standardized cocoon shell biomimetic generative process parameter and performance requirement joint input vector, pre-build a cocoon shell biomimetic one-piece shoe upper process feature knowledge base and generate a process feature constraint threshold set, input the joint input vector into the trained customized generation model, and perform real-time verification and dynamic correction of the process features of winding loop path, loop density, overlapping width, and functional partition during the model generation process to generate a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image; The generation of virtual sample images under the dynamic constraints of the cocoon-shell biomimetic structure specifically includes: S31. Construct a knowledge base for the process features of cocoon-shell biomimetic one-piece shoe upper. The knowledge base includes the process feature thresholds of cocoon-shell biomimetic winding molding, the matching rules of winding loop path and loop density, the correspondence between overlapping width and performance, and the correspondence between functional partitions and macroscopic performance indicators. S32. The standardized cocoon shell biomimetic process parameters and macroscopic performance index requirements are combined into an input vector and converted into a set of process feature constraint thresholds, including upper and lower limits of stacked ring density, curvature range of winding path, overlap width constraint, and functional partition position and size constraint. S33. Input the joint input vector into the trained cocoon shell biomimetic customized generation model. After the denoising operation at each time step of the model, extract the cocoon shell biomimetic process feature information in the latent space and compare it with the process feature constraint threshold set in real time. If the process feature information exceeds the constraint threshold set, generate a constraint correction vector and inject it into the denoising process of the next time step to realize the dynamic correction of winding loops, overlapping parts, and hot melt wire specifications. S34. Perform local enhancement processing on the biomimetic functional zoning features of the cocoon shell, strengthen the overlapping distribution and overlapping features of the support zone, breathable zone, and lightweight zone, ensure the consistency of functional zoning with the biomimetic molding process parameters and macroscopic performance index requirements of the cocoon shell, and complete the generation of a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image.

[0032] S4. Dual-dimensional evaluation and cocoon shell biomimetic parameter optimization deduction: Establish a dual-dimensional evaluation system of visual features and process performance, quantitatively evaluate the generated virtual sample images, construct a parameter performance correlation model based on the coupling of cocoon shell biomimetic process performance with multi-dimensional datasets, quantitatively analyze the influence of cocoon shell biomimetic process parameters on the macroscopic performance indicators of the shoe upper through reverse deduction of process performance, and output optimization suggestions for cocoon shell biomimetic process parameters based on evaluation results and deduction results, forming a digital closed loop of generation, evaluation and optimization.

[0033] The dual-dimensional evaluation and reverse inference of the cocoon shell biomimetic process performance specifically include: S41. The visual feature evaluation indicators include peak signal-to-noise ratio, structural similarity index, and color similarity based on improved K-means clustering. The improved K-means clustering optimizes the main color extraction process by introducing biomimetic process feature weights for cocoon shells: First, process feature vectors (including offset distance of winding and overlapping rings, width of overlapping parts, ring spacing, etc.) of each sample are extracted from the multi-dimensional dataset coupled with biomimetic process performance of cocoon shells. The correlation between each process feature and color distribution is statistically analyzed, and the main color influence weight of each process feature is calculated. Then, during the K-means clustering iteration process, the Euclidean distance between the pixel and the cluster center is multiplied by the corresponding process feature weight coefficient, so that the clustering process is tilted towards the color region that is strongly correlated with the key process features, thereby improving the process sensitivity of main color extraction and thus improving the correlation between color similarity evaluation and process performance. S42. The evaluation indicators for the process performance include the fidelity of the cocoon-like bionic process features and the accuracy of performance prediction. The accuracy of performance prediction is the quantitative matching degree between the macroscopic performance indicators such as breathability, support, and lightweighting predicted by the cocoon-like bionic process performance coupling model and the user's performance requirements. The process performance coupling model is a lightweight gradient boosting regression model (such as XGBoost) independent of the generation model. It is trained based on the cocoon-like bionic process performance coupling multi-dimensional dataset constructed in step S1. The input is the cocoon-like bionic generative process parameters (offset distance of the winding and overlapping ring group, width of the overlapping part, ring spacing, hot melt wire specifications, etc.), and the output is the macroscopic performance indicators of the shoe upper (breathability, support, lightweight coefficient, etc.). This model is used to predict the performance of the process parameters extracted from the generated image, supporting the evaluation of the process performance dimension in the two-dimensional evaluation.

[0034] S43. A gradient boosting tree algorithm is used to construct a cocoon shell biomimetic parameter performance correlation model. Based on the cocoon shell biomimetic process performance coupled multi-dimensional dataset, the single factor influence and interactive factor influence of the offset distance of the winding and overlapping ring group, the overlap width, the ring spacing, and the hot melt wire specification on the macroscopic performance indicators of the shoe upper are quantitatively analyzed, and a cocoon shell biomimetic parameter sensitivity heat map is generated. S44. If the dual-dimensional evaluation results of the virtual sample image do not meet the user's needs, the system outputs targeted optimization suggestions for the biomimetic modeling process parameters of the cocoon shell based on the biomimetic parameter sensitivity heatmap. The suggestions specify the direction of parameter adjustment, the adjustment range, and the expected improvement in macroscopic performance indicators. The optimization suggestions are output in structured text format, including parameter name, current value, suggested adjustment value, adjustment range, expected performance indicators to be improved, and improvement range. The user modifies the input parameters in the system interface based on the optimization suggestions, and the model completes the secondary generation. The system also supports a "one-click optimization" mode, which automatically updates the input based on the suggested values ​​and triggers regeneration, forming a digital closed loop of generation, evaluation, and optimization.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions above are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A cocoon-shell biomimetic one-piece shoe upper virtual sample generation system, characterized in that, The system setup includes the following steps: S1. Construct a multi-dimensional dataset of cocoon shell biomimetic process performance coupling: Collect high-definition visual images and micro-winding structure data of cocoon shell biomimetic one-piece shoe upper samples, simultaneously measure the cocoon shell biomimetic forming process parameters, micro-structural features and macro-performance indicators of the samples, perform cocoon shell biomimetic process feature-guided data augmentation processing on the high-definition visual images, generate standardized labels for the process performance coupling of the samples, and construct a multi-dimensional dataset of process performance coupling that associates cocoon shell biomimetic forming process parameters, micro-winding structure features, visual images and macro-performance indicators. S2. Training the customized generative model for cocoon-like bionics: Based on the open-source stable diffusion model, an improvement is made on its U-Net architecture to construct a multi-scale attention-enhanced U-Net architecture; through hierarchical low-rank adaptive fine-tuning technology, the model is adapted to the micro-layered ring structure and macro-winding path characteristics of the cocoon-like one-piece shoe upper, realizing the accurate mapping from process parameters and performance requirements to virtual sample images. In the upsampling path of the U-Net architecture, a microstructure feature decoding branch is added. The micro-entanglement structure features are used as auxiliary supervision signals. The latent features are decoded into microstructure parameter predictions through additional convolutional layers. The mean squared error loss is calculated with the measured microstructure features and used as a component of the total loss function. S3. Generating virtual sample images under dynamic constraints of cocoon shell biomimetic structure: Receive the range of cocoon shell biomimetic process parameters or natural language description input by the user, parse and convert it into a standardized cocoon shell biomimetic generative process parameter and performance requirement joint input vector, pre-build a cocoon shell biomimetic one-piece shoe upper process feature knowledge base and generate a process feature constraint threshold set, input the joint input vector into the trained customized generation model, and perform real-time verification and dynamic correction of the process features of winding loop path, loop density, overlapping width, and functional partition during the model generation process to generate a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image; S4. Dual-dimensional evaluation and cocoon shell biomimetic parameter optimization deduction: Establish a dual-dimensional evaluation system of visual features and process performance, quantitatively evaluate the generated virtual sample images, construct a parameter performance correlation model based on the coupling of cocoon shell biomimetic process performance with multi-dimensional datasets, quantitatively analyze the influence of cocoon shell biomimetic process parameters on the macroscopic performance indicators of the shoe upper through reverse deduction of process performance, and output optimization suggestions for cocoon shell biomimetic process parameters based on evaluation results and deduction results, forming a digital closed loop of generation, evaluation and optimization.

2. The cocoon-shell biomimetic one-piece shoe upper virtual sample generation system according to claim 1, characterized in that: The construction of the multi-dimensional dataset coupling the biomimetic process performance of the cocoon shell described in step S1 specifically includes: S11. High-precision surface scanning equipment and micro laser scanning equipment are used to collect high-definition visual images and micro winding structure data of the cocoon shell bionic one-piece shoe upper sample, respectively. The cocoon shell bionic one-piece shoe upper sample is formed by continuous winding and stacking of molten hot melt wire. S12. Determine and structure the biomimetic process parameters, microstructural characteristics and macroscopic performance indicators of the cocoon shell samples. The parameters for the biomimetic forming process of the cocoon shell include: the offset distance of the center path of the winding and overlapping ring group, the average width of the overlapping part, the ring spacing, the thickness of the hot melt wire, the material of the hot melt wire, the shape of the ring, and the forming path type; Microstructural features include: overlapping ring method, porosity, hot melt wire twist, and the density of the wound rings; Macroscopic performance indicators include: lightweight coefficient, toughness, abrasion resistance, breathability, support, fit, and durability; S13. Perform data augmentation operations guided by the biomimetic process features of the cocoon shell on the high-definition visual image. The data augmentation operations include rotation, flipping, brightness adjustment, and local stacking density perturbation and path offset operations to simulate deviations in the biomimetic cocoon shell forming process. The formal expression of the data augmentation operations is: ； In the formula, R θ For rotation and flip operators, α and β are the brightness scaling factor and brightness offset factor, respectively, and P δ Here, δ is the process deviation simulation operator, γ is the deviation coefficient, γ is the process deviation weighting coefficient, and N(0,σ²) is Gaussian noise. S14. The data-enhanced sample generates a standardized label for the coupling of cocoon shell biomimetic process performance. The standardized label contains the correlation information of cocoon shell biomimetic molding process parameters, micro-entanglement structure features, and macro-performance indicators, realizing deep coupling of image, parameters, and performance.

3. The cocoon-shell biomimetic one-piece shoe upper virtual sample generation system according to claim 1, characterized in that: The multi-scale attention-enhanced U-Net architecture and hierarchical low-rank adaptive fine-tuning technique described in step S2 specifically include: S21. In the traditional U-Net architecture, a cocoon shell biomimetic process feature attention module is added to each stage of the downsampling and upsampling paths, and a cross-scale winding structure feature fusion module is added to the intermediate bottleneck layer. The cocoon shell biomimetic process feature attention module is composed of a channel attention submodule and a spatial attention submodule connected in series. The channel attention submodule learns the channel weights of winding loops, overlapping parts, and hot-melt wire specifications through global average pooling and fully connected layers. The spatial attention submodule strengthens the spatial positioning of the cocoon shell biomimetic winding path and loop distribution through two-dimensional pooling and convolutional layers, and strengthens the extraction of core process features such as winding loop path, loop density, and overlapping part width through dual weight allocation. The cross-scale winding structure feature fusion module adopts an attention weighted fusion mechanism, which assigns differentiated weights to feature maps of different resolutions and adds them element by element to achieve accurate fusion learning of microscopic loop texture and macroscopic cocoon shell biomimetic contour. S22. The hierarchical low-rank adaptive fine-tuning technique designs differentiated low-rank matrices based on different levels of the multi-scale attention-enhanced U-Net architecture: the downsampling path is configured with a low-rank matrix of rank r = 6–8 to adapt to the complex mapping learning of macroscopic winding paths in cocoon-like bionic architecture; the upsampling path is configured with a low-rank matrix of rank r = 3–5 to adapt to the fine representation of microscopic overlapping textures; the intermediate layer is configured with a low-rank matrix of rank r = 8–10 to adapt to the coupled learning of cross-scale winding structural features; the original weights of the architecture are divided into downsampling weight reorganization, intermediate layer weight reorganization, and upsampling weight reorganization according to the hierarchy. During the forward calculation of the original weights at each level, they are updated to the sum of the original weights and the corrected value of the low-rank matrix, i.e.: ； ； In the above formula, The original weights for each level, This is the correction value for the low-rank matrix. These are the scaling factors for each level. The rank of a low-rank matrix. , The matrix is ​​a trainable low-rank matrix; during fine-tuning, the original weights are frozen, and only the parameters of the low-rank matrix at each level are updated via backpropagation. S23. The total loss function L is a weighted sum of visual loss Lvis, cocoon shell biomimetic process feature loss Lpro, and performance coupling loss Lper, and its expression is: ； In the formula, The weighting coefficients and Lvis is the sum of pixel-level reconstruction loss and perception loss, Lpro is the cosine similarity loss of cocoon shell biomimetic process feature matching, and Lper is the mean square error loss of breathability and support performance prediction. S24. After model training is complete, a dynamic weight fusion strategy is adopted. The user-input priorities of support, breathability, and lightweight requirements are expressed as percentages. The system normalizes the priority vector and converts it into a fusion ratio λ between the low-rank matrix correction value of each level and the original weights through a preset mapping function. The value of λ ranges from 0.6 to 0.

9. The fusion ratio is calculated using the following formula: ； in, Based on the fusion ratio, Let the i-th performance requirement be prioritized. These are the empirical coefficients for the corresponding performance terms.

4. The cocoon-shell biomimetic one-piece shoe upper virtual sample generation system according to claim 1, characterized in that: The generation of the virtual sample image under the dynamic constraints of the cocoon-like biomimetic structure in step S3 specifically includes: S31. Construct a knowledge base for the process features of cocoon-shell biomimetic one-piece shoe upper. The knowledge base includes the process feature thresholds of cocoon-shell biomimetic winding molding, the matching rules of winding loop path and loop density, the correspondence between overlapping width and performance, and the correspondence between functional partitions and macroscopic performance indicators. S32. The standardized cocoon shell biomimetic process parameters and macroscopic performance index requirements are combined into an input vector and converted into a set of process feature constraint thresholds, including upper and lower limits of stacked ring density, curvature range of winding path, overlap width constraint, and functional partition position and size constraint. S33. Input the joint input vector into the trained cocoon shell biomimetic customized generation model. After the denoising operation at each time step of the model, extract the cocoon shell biomimetic process feature information in the latent space and compare it with the process feature constraint threshold set in real time. If the process feature information exceeds the constraint threshold set, generate a constraint correction vector and inject it into the denoising process of the next time step to realize the dynamic correction of winding loops, overlapping parts, and hot melt wire specifications. S34. Perform local enhancement processing on the biomimetic functional zoning features of the cocoon shell, strengthen the overlapping distribution and overlapping features of the support zone, breathable zone, and lightweight zone, ensure the consistency of functional zoning with the biomimetic molding process parameters and macroscopic performance index requirements of the cocoon shell, and complete the generation of a high-resolution cocoon shell biomimetic one-piece shoe upper virtual sample image.

5. The cocoon-shell biomimetic one-piece shoe upper virtual sample generation system according to claim 1, characterized in that: The dual-dimensional evaluation and reverse inference of cocoon shell biomimetic process performance described in step S4 specifically include: S41. The visual feature evaluation indicators include peak signal-to-noise ratio, structural similarity index, and color similarity based on improved K-means clustering. The improved K-means clustering optimizes the main color extraction process by introducing cocoon-shell biomimetic process feature weights: First, process feature vectors of each sample are extracted from the cocoon-shell biomimetic process performance coupled multi-dimensional dataset. The correlation between each process feature and color distribution is statistically analyzed, and the main color influence weight of each process feature is calculated. Then, during the K-means clustering iteration process, the Euclidean distance between the pixel and the cluster center is multiplied by the corresponding process feature weight coefficient, so that the clustering process is tilted towards the color region strongly correlated with the key process features, thereby improving the process sensitivity of main color extraction and thus improving the correlation between color similarity evaluation and process performance. S42. The evaluation indicators of the process performance include the fidelity of the cocoon shell bionic process features and the accuracy of performance prediction. The accuracy of performance prediction is the quantitative matching degree between the macro performance indicators such as breathability, support, and lightweight predicted by the cocoon shell bionic process performance coupling model and the user's performance requirements. The process performance coupling model is a lightweight gradient boosting regression model independent of the generation model. It is trained based on the cocoon shell bionic process performance coupling multi-dimensional dataset constructed in step S1. The input is the cocoon shell bionic generative process parameters, and the output is the macro performance indicators of the shoe upper. S43. A gradient boosting tree algorithm is used to construct a cocoon shell biomimetic parameter performance correlation model. Based on the cocoon shell biomimetic process performance coupled multi-dimensional dataset, the single factor influence and interactive factor influence of the offset distance of the winding and overlapping ring group, the overlap width, the ring spacing, and the hot melt wire specification on the macroscopic performance indicators of the shoe upper are quantitatively analyzed, and a cocoon shell biomimetic parameter sensitivity heat map is generated. S44. If the dual-dimensional evaluation results of the virtual sample image do not meet the user's needs, the system outputs targeted optimization suggestions for the biomimetic modeling process parameters of the cocoon shell based on the sensitivity heat map of the cocoon shell biomimetic parameters. The suggestions clarify the direction of parameter adjustment, the adjustment range, and the expected improvement in macroscopic performance indicators. The optimization suggestions are output in structured text form, including parameter name, current value, suggested adjustment value, adjustment range, expected performance indicators to be improved, and improvement range. The user modifies the input parameters in the system interface based on the optimization suggestions, and the model completes the secondary generation. The system also supports a "one-click optimization" mode, which automatically updates the input based on the suggested values ​​and triggers regeneration, forming a digital closed loop of generation, evaluation, and optimization.

Images