Individualized ornament generation type design method based on cultural label vector

Abstract

The invention provides a personalized ornament generative design method based on cultural label vectors. The method comprises the following steps: constructing design parameter vectors including materials, colors, shapes, quality, processes, combinations and wearing positions; constructing a quantitative mapping relation function of the material, color, shape, quality, process and culture label vectors, and correcting through combination and wearing positions to obtain a final culture label prediction vector; taking the difference between a set five-element discrete classification vector preference target value and a final prediction vector as a label error term, and constructing a differentiable target function in combination with cost constraint, work efficiency penalty and label overflow penalty; and performing optimization solution on the target function to obtain an optimal design parameter vector, and generating a corresponding ornament digital model design scheme based on the optimal design parameter vector. Technical coupling of numeralization mapping of traditional cultural symbols and computer-aided manufacturing is realized, and an algorithm framework which can be industrially realized is provided for large-scale personalized ornament customization.

Description

Technical Field

[0001] This invention relates to the field of jewelry design technology, specifically to a personalized jewelry generative design method based on cultural tag vectors. Background Technology

[0002] As the jewelry industry moves towards personalization and mass customization, how to integrate cultural elements into designs has become an important research topic. Existing technical solutions include some literature proposing the use of cultural symbol classification tables to map the materials, shapes, and colors of jewelry to five-element symbol attributes, thus providing designers with inspiration. However, these methods typically remain at the symbolic level, lacking mathematical and functional mapping mechanisms. They can only serve as inspiration for manual design and cannot form a fundamental technical framework for large-scale personalized production.

[0003] On the other hand, some studies have attempted to apply five-element discrete classification vectors to the field of traditional Chinese medicine meridians and establish mathematical models for quantitative calculations. While these methods can achieve numerical quantification, their application is mainly limited to medical and health scenarios and cannot be directly transferred to the design and production of jewelry.

[0004] Existing jewelry personalization technologies mostly focus on appearance modeling and process optimization, lacking mechanisms for deeply integrating cultural symbols with parametric design spaces. Therefore, there remains a technological gap in how to transform traditional cultural symbols into calculable parametric models and further couple them with modern computer-aided design and production processes. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a personalized jewelry generative design method based on cultural tag vectors, comprising the following steps:

[0006] Step S1: Construct a design parameter vector including material, color, shape, quality, process, combination, and wearing position;

[0007] Step S2: Construct a quantitative mapping function between material, color, shape, quality, craftsmanship, and cultural label vectors, and obtain the final cultural label prediction vector by combining and correcting the data based on the wearing position;

[0008] Step S3: The difference between the set five-element discrete classification vector preference target value and the final cultural label prediction vector is used as the label error term, and a differentiable objective function is constructed by combining cost constraints, efficiency penalties and label overflow penalties;

[0009] Step S4: Optimize the objective function to obtain the optimal design parameter vector, and generate the corresponding digital model design scheme for the jewelry based on the optimal design parameter vector.

[0010] Preferably, the expression for the quantization mapping function between material and cultural tag vector in step S2 is:

[0011] E_mat=BaseTable(i,xmat)+Δ_mat(xmat);

[0012] In the formula, E_mat represents the material culture label output vector; BaseTable(i,xmat) represents the material-five-element discrete classification vector base mapping matrix, i∈{metal,wood,water,fire,earth}, xmat represents the material type; Δ_mat represents the training residual vector corresponding to the material.

[0013] Preferably, the expression for the quantization mapping function between color and cultural tag vector in step S2 is:

[0014] E_col = W_col·RBF(Lab);

[0015] In the formula, E_col represents the color culture label output vector; W_col represents the color-five-element discrete classification vector basis mapping matrix; Lab represents the three-dimensional vector in the CIE-Lab space; and RBF represents the radial basis function mapping centered on white, cyan, black, red, and yellow.

[0016] Preferably, the expression for the quantization mapping function between shape and cultural tag vector in step S2 is:

[0017] E_shp = W_shp·f;

[0018] In the formula, E_shp represents the shape culture label output vector; W_shp represents the shape-five-element discrete classification vector base mapping matrix; f represents the geometric feature vector containing roundness or sphericity, aspect ratio or length-to-diameter ratio, sharpness and mean curvature;

[0019] Preferably, the expression for the quantization mapping function between quality and cultural tag vectors in step S2 is:

[0020] K_mass=β·SoftplusClip(xmass);

[0021]

[0022] In the formula, K_mass represents the quality amplification factor, i.e., the quality label weighting factor; β represents the scaling factor; and xmass represents the quality.

[0023] Preferably, the quantitative mapping relationship between the process and cultural label vectors in step S2 is obtained by looking up the process-five-element discrete classification vector basic mapping matrix.

[0024] Preferably, the expression modified by combination in step S2 is:

[0025]

[0026] E i =K_mass i ·(E_mat i +E_col i +E_shp i +E_proc i );

[0027] In the formula, E total This represents the predicted vector after combination and correction; N represents the number of ornaments; α i Let represent the attenuation coefficient of the i-th accessory; xcombo[i] represents the binary mask variable of the i-th accessory, where 0 indicates no superposition and 1 indicates superposition; E i E_mat represents the cultural tag prediction vector for the i-th accessory. i E_col i E_shp i and E_proc i These are the output vectors for the material culture label, color culture label, shape culture label, and craftsmanship culture label of the i-th ornament, respectively, where K_mass is the output vector. i Let be the mass magnification factor for the i-th piece of jewelry.

[0028] Preferably, the expression for obtaining the final cultural tag prediction vector by correcting the wearing position in step S2 is:

[0029] E complement_norm(x)=clip(T_pos(xwear)·E_total,-1,1);

[0030] In the formula, E_norm(x) represents the final label prediction vector; clip represents the element-wise limiting function; and T_pos(xwear) represents the five-element discrete classification mapping vector obtained by querying the wear position correction-five-element discrete classification vector base mapping matrix based on the wearing position xwear.

[0031] Preferably, the expression for the objective function Loss in step S3 is: Loss = ||Edeficient_norm - Esupplementary_norm(x)||2 + λ·Cost(x) + μ·Wearability(x; age, generation)

[0032] er)+δ·BalancePenalty(x);

[0033] In the formula, Emissing_norm represents the preference target vector; Ecomplementing_norm(x) represents the final label prediction vector; λ·Cost(x) represents the cost constraint, and λ represents the user budget sensitivity coefficient; μ·Wearability(x; age, gender) represents the efficiency penalty, and μ represents the global penalty coefficient; δ·BalancePenalty(x) represents the label overflow penalty, and δ represents the penalty intensity coefficient.

[0034] Preferably, the jewelry design digital model generated in step S4 includes a 3D model file in STL or OBJ format.

[0035] The beneficial effects of this invention include at least the following: By constructing a functional mapping relationship between material, color, shape, quality, and process and a five-element discrete classification vector, and introducing combination superposition and wearing position correction, this invention achieves the vectorization and parameterization of cultural labels. It transforms discrete cultural labels such as material, color, and shape into calculable five-element discrete classification vectors, avoiding the limitation of traditional classification tables serving only as inspirational references, allowing cultural label attributes to directly enter the calculation process. By setting a target value for the five-element discrete classification vector preference and introducing a label error term into the objective function, it can automatically generate digital model design schemes for jewelry that match the predicted effects for different user needs, thereby achieving personalized output. Simultaneously considering cost constraints, efficiency penalties, and label overflow penalties in the objective function ensures that the digital model design scheme matches the user's required cultural labels while controlling manufacturing costs, wearing comfort, and safety, avoiding unproducible or unwearable design results. By optimizing the solution to obtain the optimal design parameter vector, and directly driving the parameterized 3D template to generate STL / OBJ models and G-code process files, it realizes a complete link from demand input to production output, possessing engineering feasibility.

[0036] The “five-element discrete classification vector” is used only as a cultural symbol identifier, and its dimension names “metal, wood, water, fire, earth” do not carry any physical or physiological meaning.

[0037] In summary, this invention is the first to minimize the Euclidean distance between the user-input preference target vector and the prediction vector, while introducing cost, weight, and manufacturability constraints to solve for the optimal digital model design parameters of the jewelry, thereby completing the closed loop from cultural symbols to industrial documents. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0040] The material five-element discrete classification vector standard is essentially a cultural symbol system based on cultural label consensus, symbolic meaning, and attribute mapping. This invention provides a specific method and process for constructing the core function mapping table, establishing a function mapping table structure of "material / color / shape / craftsmanship - five-element discrete classification vector attributes" and an algorithm for the objective function of jewelry design. However, it does not include the specific vector numerical definitions of the material five-element discrete classification vectors; the specific values ​​are trade secret data defined by manufacturers themselves or publicly available standard reference data established by industry associations.

[0041] Based on the above, such as Figure 1 As shown, this embodiment of the invention provides a personalized jewelry generative design method based on cultural tag vectors.

[0042] The overall algorithm and implementation process include the following steps.

[0043] 1) Input layer: Receives the five-element discrete classification vector, preference target value Emiss ∈ R0, from the third-party system. 5 And optional constraints, including budget limits, wearing position for single or multiple items, age, gender, and style preference.

[0044] 2) Adaptive scaling layer: In order to prevent the difference in the numerical range of the target value E missing of the five-element discrete classification vector input by the external system from being too large, which would lead to gradient explosion or slow convergence, the system sets an adaptive scaling layer at the optimization entry point. Its normalized output E missing_norm is shown in Equation (1).

[0045] Edeficient_norm = clip(Edeficient / max(|Edeficient|,ε),–1,1) (1)

[0046] In the formula, ε=1×10 -6 , used to prevent division by zero; clip(·,–1,1) represents the element-wise truncation function, restricting the result to the closed interval [-1,1]; |·| is the element-wise absolute value; E is missing _norm∈R 5 The units are consistent with those of the subsequent modules.

[0047] 3) Function mapping table: Establish a five-element discrete classification vector function mapping table for material / color / shape / process, and superimpose a learnable residual matrix.

[0048] 4) Construction of the objective function for jewelry design: Taking the minimization of the objective value of preference as the core, and combining cost constraints, efficiency penalties, and label overflow penalties, a differentiable objective function Loss is constructed.

[0049] 5) Jewelry Parameter Optimization Engine: Based on the gradient descent algorithm, the objective function is solved and the optimal jewelry design parameter vector x is output.

[0050] 6) Customized solution generation: Convert parameter x into 3D model and production parameters, and connect to the cloud factory through RESTful API.

[0051] 7) Incremental learning layer: Collects user feedback data and updates the residual matrix and the weights of the small neural network online.

[0052] 8) NFT Certificate Module: Supports the Ethereum ERC-721 standard, binds the final parameter x with the 3D model hash and writes it to the blockchain to ensure uniqueness and copyright.

[0053] The parameterization is defined as an adjustable parameter vector x = [xmat, xcolor, xshape, xmass, xproc, xcombo, xwear]. The meaning and value range of the sub-vectors are shown in Table 1.

[0054] Table 1

[0055]

[0056]

[0057] In Table 1, R 5 This represents a 5-dimensional vector consisting of the five elements: metal, wood, water, fire, and earth. Positive values ​​indicate an increase, while negative values ​​indicate a decrease.

[0058] The function mapping table in this embodiment of the invention is as follows.

[0059] 1) Output vector E_mat of material culture label

[0060] The mathematical expression for the material culture tag output vector E_mat is:

[0061] E_mat=BaseTable(i,xmat)+Δ_mat(xmat);

[0062] Where i∈{metal,wood,water,fire,earth}; BaseTable(i,xmat) is the material-five-element discrete classification vector base mapping matrix, with a dimension of 5×N_mat, where N_mat is the total number of material types; Δ_mat is the 5×1 residual vector corresponding to the material; the residual vectors of all materials are stored as a 5×N_mat residual matrix, initialized to zero before training, and allowed to be updated during system operation through online fine-tuning algorithm; E_mat∈R5 Furthermore, it is stipulated that positive values ​​of its elements represent increases, and negative values ​​represent decreases.

[0063] The construction process of the Material-Five-Element Discrete Classification Vector BaseTable is shown in Table 2.

[0064] Table 2

[0065]

[0066]

[0067] The training mechanism for the residual matrix Δ_mat includes:

[0068] Initialization: Each material has an independent 5×1 residual vector, which is stored as a 5×N_mat residual matrix and initialized to zero before training.

[0069] Training objective: Minimize the "prediction error of the preference target value" in user feedback.

[0070] Update frequency: Batch updates at 2 AM daily, learning rate η = 0.001.

[0071] Regularization: L2 regularization λ = 0.01, to prevent overfitting of niche materials.

[0072] Specifically, the calculation example is as follows.

[0073] The scenario is set as a user experiencing water shortage. The base mapping matrix of the material-five-element discrete classification vector is looked up in the table. The material enumeration xmat = 7 → "Aquamarine", resulting in BaseTable(i,7) = [-0.10,-0.05,+0.75,-0.05,-0.05], where i∈{metal,wood,water,fire,earth}. The residual vector of aquamarine after training is Δ_mat = [+0.02,-0.01,+0.04,0,-0.01]. The final E_mat calculation is shown in Table 3.

[0074] Table 3

[0075] Five-element discrete classification vector BaseTable +Δ_mat =E_mat gold -0.10 +0.02 -0.08 Wood -0.05 -0.01 -0.06 water +0.75 +0.04 +0.79 fire -0.05 0 -0.05 earth -0.05 -0.01 -0.06

[0076] The output is: E_mat=[-0.08,-0.06,+0.79,-0.05,-0.06], with the added label vector water value.

[0077] 2) The color culture label output vector E_col is a differentiable function that transforms the continuous color space input into a five-element discrete classification vector. The calculation formula is as follows:

[0078]

[0079] W_col∈R 5x5 This is a trainable matrix, constructed using the same process as the material-five-element discrete classification vector base mapping matrix, and supports online fine-tuning; Lab represents the three-dimensional column vector [L,a,b] of color in the CIE-Lab uniform color space. T .

[0080] The calculation process for the radial basis function (RBF) (Lab) includes: pre-fixing five centers C1…C5 in the Lab space, corresponding to the traditional primary colors white→gold, cyan→wood, black→water, red→fire, and yellow→earth respectively; setting the bandwidth uniformly to σ=8 (a configurable constant); and adding a small ε=1×10 -6 Preventing zero gradient:

[0081]

[0082] in The order corresponds to metal, wood, water, fire, and earth, respectively. The final output is E_col∈R. 5 Each component corresponds to a discrete classification vector of five elements: metal, wood, water, fire, and earth. Positive values ​​indicate an increase, and negative values ​​indicate a decrease. They can be directly used in subsequent differentiable optimization.

[0083] The implementation process includes: converting the RGB color values ​​of the ornament into contributions to the five-element discrete classification vector of "metal, wood, water, fire, and earth". The process consists of the following four steps: ① RGB → Lab (uniform color space) → ② 5 RBF (radial basis function) center activation → ③ 5×5 trainable matrix W_col linear transformation → ④ Output 5-dimensional vector E_col∈R 5 .

[0084] The initialization and update of the trainable matrix W_col includes: During initialization, the weights of the five-element discrete classification vector corresponding to each color center k are set as follows: the weights of the main attributes corresponding to each color center are placed near the kth row and kth column, the weights of the main diagonal are set to 0.85-0.90, the weights of the corresponding attribute columns are set to 0.05-0.10, and the weights of other positions are set to small random values ​​of ±0.01.

[0085]

[0086] Each row here corresponds to a five-element discrete classification vector with center reinforcement: Row 1: Metal, representing white center reinforcement; Row 2: Wood, representing cyan center reinforcement; Row 3: Water, representing black center reinforcement; Row 4: Fire, representing red center reinforcement; Row 5: Earth, representing yellow center reinforcement.

[0087] It shares the same loss function with the material residual, with a learning rate η = 0.001 and L2 regularization λ = 0.01.

[0088] Example verification is as follows.

[0089] Color: Deep Sea Blue #0077BE→Lab≈(42,-9,-46), with fixed center containing the corresponding order of five-element discrete classification vectors, as shown in Table 4.

[0090] Table 4

[0091] center Lab coordinates Five-element discrete classification vector green (75,-40,20) Wood red (50,80,70) fire yellow (90,-10,90) earth white (95,0,0) gold black (45,-15,-55) water

[0092] RBF activation value calculation (bandwidth σ = 8), formula: The calculation results are shown in Table 5.

[0093] Table 5

[0094]

[0095]

[0096] therefore: calculate Calculate line by line (round to 2 decimal places):

[0097] Gold: [0.85,0.01,0.01,0.09,0.01]·[0.00,0.00,0.14,0.00,0.00] T =0.85·0 + 0.01·0 + 0.01·0.14 + 0.09·0 + 0.01·0 = 0.0014

[0098] Wood: [0.01,0.88,0.01,0.01,0.01]·[0.00,0.00,0.14,0.00,0.00] T =0.01·0 + 0.88·0 + 0.01·0.14 + 0.01·0 + 0.01·0 = 0.0014

[0099] Water: [0.01,0.01,0.90,0.01,0.08]·[0.00,0.00,0.14,0.00,0.00] T =0.01·0 + 0.01·0 + 0.90·0.14 + 0.01·0 + 0.08·0 = 0.1260

[0100] Fire: [0.01,0.06,0.01,0.87,0.01]·[0.00,0.00,0.14,0.00,0.00] T =0.01·0 + 0.06·0 + 0.01·0.14 + 0.87·0 + 0.01·0 = 0.0014

[0101] Soil: [0.06,0.01,0.01,0.01,0.89]·[0.00,0.00,0.14,0.00,0.00] T =0.06·0 + 0.01·0 + 0.01·0.14 + 0.01·0 + 0.89·0 = 0.0014

[0102] Therefore, E_col≈[0.00,0.00,0.13,0.00,0.00], indicating that only the cultural label water dimension has a significant positive increase, while the other five discrete classification vectors are approximately 0, which is consistent with the expectation of deep sea blue → water enhancement.

[0103] 3) The shape culture label output vector E_shp has the following applicable range as shown in Table 6.

[0104] Table 6

[0105] Dimension Data format Typical jewelry 2D Closed polygon (vertices sequence) Flat pendant, laser-cut brooch 3D Triangular mesh (STL / OBJ) 3D rings, teardrop pendants, openwork spheres

[0106] The formula for calculating the shape culture tag output vector E_shp is:

[0107] E_shp = W_shp·f;

[0108] Where f = [f1, f2, f3, f4] T ∈R 4 , which is the normalized 4-dimensional geometric feature vector; as shown in Table 7.

[0109] Table 7

[0110]

[0111] A quick reference table of unified symbols and parameters is shown in Table 8.

[0112] Table 8

[0113]

[0114]

[0115] W_shp∈R 5x4 It is a trainable matrix with 5 rows corresponding to "metal, wood, water, fire, and earth" and 4 columns corresponding to f1-f4 geometric features. The construction process is the same as the material-five-element discrete classification vector basic mapping matrix construction process above, which will not be described here. It supports subsequent online fine-tuning.

[0116] Specific examples are shown in Table 9, where 2D and 3D share the same dimension.

[0117] Table 9

[0118] Dimension <![CDATA[Overall shape integrity f1]]> <![CDATA[Degree of stretch f2]]> <![CDATA[Sharpness f3]]> <![CDATA[Curvature f4]]> gold +0.80 -0.10 -0.05 +0.10 Wood -0.10 +0.75 -0.05 -0.05 water +0.05 -0.05 -0.05 +0.80 fire -0.05 +0.05 +0.85 -0.10 earth +0.10 +0.05 -0.10 +0.15

[0119] E_shp∈R 5 Each component corresponds to a discrete classification vector of five elements: metal, wood, water, fire, and earth. Positive values ​​indicate an increase, and negative values ​​indicate a decrease.

[0120] Example verification is as follows.

[0121] 2D example: Laser-cut circular brooch.

[0122] Geometric data:

[0123] Area = 314.16mm 2 ;L=62.83mm; Length=Width=20mm; κ_iall≈0.1mm -1 (No sharp corners)

[0124] eigenvectors;

[0125] f1 = 4π·314.16 / 62.83 2 =1.000;

[0126] f2 = 20 / 20 = 1.000;

[0127] f3 = 0 (no sharp corners);

[0128] f4=(1 / 62.83)∫|κ|ds=0.100;

[0129] After normalization, f = [1.00, 1.00, 0.00, 0.10] T .

[0130] E_shp=W_shp·f, and the calculation results are shown in Table 10.

[0131] Table 10

[0132]

[0133] The result is E_shp≈[+0.71,+0.65,+0.08,-0.01,+0.17] → the round gold and wood label vector is the strongest.

[0134] 3D example: a teardrop pendant made of obsidian.

[0135] Geometric data:

[0136] V = 3077 mm 3 (3.077cm 3 S = 1560 mm 2 a = 28 mm; b = 18 mm; c = 18 mm; mean κ_max 0.32 mm -1 Sharp corner ratio: 7%.

[0137] Feature vector:

[0138] f1 = 36π·3077 2 / 1560 3 ≈0.18;

[0139] f2 = 28 / 18 = 1.56, normalized to 0.78;

[0140] f3 = 0.07, representing sharpness;

[0141] f4=(1 / 1560)∫∫|H|dA=0.19;

[0142] After normalization, f = [0.18, 0.78, 0.07, 0.19] T .

[0143] E_shp=W_shp·f, and the calculation results are shown in Table 11.

[0144] Table 11

[0145]

[0146]

[0147] The result is E_shp≈[+0.08,+0.55,+0.12+0.07,+0.08], which yields the label vector distribution of the teardrop-shaped wood-main and water-auxiliary elements.

[0148] 4) Mass amplification K_mass, which maps physical mass to a scalable scaling factor of the intensity of a five-element discrete classification vector.

[0149] The formula for calculating the mass magnification factor K_mass is:

[0150] K_mass=β·SoftplusClip(xmass);

[0151] Where β is the scaling factor, with a default value of 0.05, which can be updated through backpropagation during the training phase; xmass∈[1g,20g] is a continuously differentiable optimization variable representing physical mass; Softplus Clip(xmass) is a piecewise differentiable function;

[0152] The complete expression for the piecewise function SoftplusClip(xmass) is:

[0153]

[0154] The expression for Softplus(xmass) is:

[0155] Softplus(xmass)=ln(1+e^xmass);

[0156] The meaning is that Softplus(0)=ln2≈0.693, which ensures that the function value and the first derivative are continuous at xmass=1g and xmass=20g, thus avoiding gradient vanishing caused by hard truncation.

[0157] The final output K_mass is a positive real number, which is directly used as a scaling factor for the strength of the subsequent five-element discrete classification vectors and participates in the overall differentiable optimization process.

[0158] The default values ​​and gradient characteristics are as follows: within the interval 1g≤xmass≤20g, SoftplusCl ip(xmass)=xmass, therefore K_mass=β·xmass, and the gradient is always β.

[0159] Example: β = 0.05, xmass = 8g (obsidian) yields K_mass = 0.4; xmass = 20g (aquamarine) yields K_mass = 1.0. When xmass < 1g or xmass > 20g, Soft plusClip saturates the compression gradient to prevent the formation of extreme masses that are unmanufacturable or uncomfortable to wear.

[0160] In this embodiment, β is set to either a fixed value or an online learnable scalar, initialized to 0.05 during learning, and L2 regularized to 1×10. -4 It is updated synchronously with the residual matrix.

[0161] 5) The process increment E_proc is calculated using the following formula:

[0162] E_proc(i)=BaseProc(i,xproc);

[0163] Where i ∈ {metal, wood, water, fire, earth}; BaseProc(i, xproc) is the process-five-element discrete classification vector base mapping matrix with a dimension of 5×N_proc, where N_proc is the total number of process types; xproc is the current process category index variable, with a value range of 1…N_proc; E_proc ∈ R 5 Each component corresponds to a discrete classification vector of five elements: metal, wood, water, fire, and earth; positive values ​​indicate an increase, and negative values ​​indicate a decrease.

[0164] The final output E_proc is determined solely by the process category and is independent of color, quality, and shape. It is output directly using a lookup table and can replace the BaseProc matrix online as the process library is updated.

[0165] The process for constructing the basic mapping matrix of the process-five-element discrete classification vector is the same as that for constructing the basic mapping matrix of the material-five-element discrete classification vector, and will not be elaborated further here.

[0166] Typical process items are shown in Table 12.

[0167] Table 12

[0168]

[0169]

[0170] 6) The formula for calculating E_total, which combines and superimposes data, is as follows:

[0171]

[0172] Where i = 1, 2, ..., N is the combination order (1 is the newest, N is the last), N = 5; xcombo[i] ∈ {0, 1} is the binary mask variable of the i-th single item, 0 indicates that it does not participate in the superposition, and 1 indicates that it participates in the superposition; α ∈ (0, 1] is a configurable attenuation coefficient, the default α = 0.8, which can be dynamically adjusted in the range of 0–1 during the training or inference phase; E i Let i be the predicted tag vector for the i-th single piece of jewelry;

[0173] Single item tag prediction vector E i The expression is:

[0174] E i =K_mass i ·(E_mat i +E_col i +E_shp i +E_proc i );

[0175] Where K_mass i Let be the mass amplification factor for the i-th single item.

[0176] E_mat i E_col i E_shp i E_proc i These are the output vectors for the material culture label, color culture label, shape culture label, and process increment of the i-th single piece, respectively.

[0177] Output E_total∈R 5 Each component corresponds to a discrete classification vector of five elements: metal, wood, water, fire, and earth. Positive values ​​indicate an increase, and negative values ​​indicate a decrease. These can be directly used for subsequent differentiable optimization or decision output.

[0178] Exemplarily, the user selects 3 pieces of jewelry and makes selections in the combination order of Table 13:

[0179] Table 13

[0180]

[0181]

[0182] Calculate the total prediction vector E_total = 0.8E 1 + 0.64E 2 + 0.512E 3 ≈ [0.44, 0.03, -0.18, -0.07, 0.02].

[0183] 7) The calculation formula for the wearing position correction matrix T_pos is:

[0184] T_pos(xwear) = diag(t_gold, t_wood, t_water, t_fire, t_earth);

[0185] where xwear ∈ {1, 2,..., 10} is the enumeration value of the wearing position; when multiple pieces of jewelry are worn simultaneously, xwear can be a vector of length 5 [xwear1,..., xwear5], corresponding to the positions of each piece of jewelry in xcombo; fill the empty positions with 0 to follow the single-piece xwear; diag(·) represents a 5×5 diagonal matrix, and its diagonal elements are fixed as (t_gold, t_wood, t_water, t_fire, t_earth), all of which are real constants; T_pos(xwear) ∈ R 5 , and its components correspond to the five-element discrete classification vectors of gold, wood, water, fire, and earth in sequence; positive values indicate an increase, and negative values indicate a decrease, and it is directly output by looking up the table.

[0186] The construction process of the wearing position correction - five-element discrete classification vector basic mapping matrix is the same as that of the material - five-element discrete classification vector basic mapping matrix, and will not be elaborated here.

[0187] The constant mapping table of the enumerated xwear and the corresponding 5×1 vector t = [t_gold, t_wood, t_water, t_fire, t_earth] is shown in Table 14.

[0188] Table 14

[0189] Part xwear fixed value of vector t left wrist 1 [1.10,1.00,1.05,0.95,0.95] right wrist 2 [0.90,1.00,0.95,1.05,1.05] chest 3 [0.95,0.90,0.95,1.20,1.00] front of the neck 4 [1.00,1.00,1.00,1.00,1.00] Left fingers 5 [1.05,0.95,1.00,0.95,1.05] Right finger 6 [0.95,1.05,1.00,1.05,0.95] left ankle 7 [1.00,1.00,1.05,0.95,0.95] Right ankle 8 [1.00,1.00,0.95,1.05,1.05] front of the waist 9 [0.95,1.05,0.95,0.95,1.10] head 10 [1.05,1.00,0.95,1.10,0.95]

[0190] 8) The calculation formula for the final label prediction vector E_hat_norm(x) is:

[0191] E_hat_norm(x) = clip(T_pos(xwear) · E_total, -1, 1);

[0192] Where xwear is the index of the wearing position; T_pos(xwear)∈R 5x5 The dressing position correction matrix; E_total∈R 5 The combined superposition of the total prediction vector; the symbol · represents matrix-vector multiplication; clip(·,-1,1) is the element-wise limiting function, defined as clip(z,-1,1)=min(max(z,-1),1).

[0193] The overall calculation process is as follows:

[0194] (1) Calculate the unlimited prediction vector Ecomplement_norm_raw = Tpos(xwear)·Etotal;

[0195] (2) Performing a clipping operation on each element of Ecomplement_norm_raw yields Ecomplement_norm(x)∈[-1,1] 5 This ensures that the dimensions are consistent with the normalized preference target vector.

[0196] Exemplary examples are described below through specific embodiments.

[0197] The user wears the jewelry on their right wrist, resulting in xwear = 2. Looking up the table, the T_pos vector t = [0.90, 1.00, 0.95, 1.05, 1.05]. After element-wise correction, the unlimited prediction vector E_norm_raw = [0.90 × 0.44, 1.00 × 0.03, 0.95 × (-0.18), 1.05 × (-0.07), 1.05 × 0.02] ≈ [0.40, 0.03, -0.17, -0.07, 0.02]. The limited E_norm = clip(E_norm_raw, -1, 1) = [0.40, 0.03, -0.17, -0.07, 0.02]. In this embodiment, all components are within the range [-1, 1], so truncation is unnecessary. The result is the final label prediction vector, which can be directly used to calculate the objective function Loss.

[0198] The jewelry design objective function Loss constructed in this embodiment consists of four parts: label error term (core), cost constraint, efficiency penalty, and label overflow penalty. Its expression is as follows:

[0199] Loss = ||Edeficient_norm - Esupplemented_norm(x)|| 2 +λ·Cost(x)+μ·Wearability(x;age,gender)+δ·BalancePenalty(x);

[0200] Where ||Edeficient_norm - Ecomplementary_norm(x)|| 2 λ·Cost(x) is the label error term; μ·Wearability(x; age, gender) is the efficiency penalty; δ·BalancePenalty(x) is the label overflow penalty.

[0201] The mathematical expression for the label error term is: ||Emissing_norm - Ecomplementing_norm(x)|| 2 .

[0202] The mathematical expression for the cost constraint is:

[0203] λ·Cost(x)=λ·P(xmat)·Q(x);

[0204] Where P(xmat) is the unit price of the material, in yuan / gram, obtained from the preset market price list; Q(x) is the total mass of the material, in grams; λ is the user's budget sensitivity coefficient, λ>0, the default λ=0.01 yuan, which can be adjusted in the configuration file.

[0205] The mathematical expression for performance penalty is:

[0206] μ·Wearability(x,age,gender)=μ·[w1·ShapePenalty+w2·MassPenalty+w3·SharpPenalty];

[0207] Where μ defaults to 0.05, representing the global penalty coefficient, which is adjustable; w1 = 0.20, w2 = 0.50, w3 = 0.30. These are the default weights, adjustable, with a value range of (0,1).

[0208] ShapePenalty=ShapeBase(xshape)·γ_s;

[0209] ShapeBase(xshape) is a shape enumeration constant table with values ​​from 0 to 1, calibrated by human factors experiments; γ_s∈(0,2] is output by a small MLP, with the activation function Softplus-1, which can amplify or weaken the shape effect.

[0210] MassPenalty=Softplus(xmass-M_max)·0.1;

[0211] M_max=γ_m·BaseMass(age);

[0212] BaseMass segmentation: 5g for children, 8g for women, 12g for men, and 10g for the elderly; γ_m∈(0,2] is output by the same MLP to control weight tolerance.

[0213] SharpPenalty=σ((size_max-S_max) / τ);

[0214] size_max is the measured size of the sharpest part of the item, in mm. It is directly given by the upstream CAD model or physical 3D scan after scanning or measuring the product.

[0215] S_max=γ_p·BaseSize(age);

[0216] BaseSize segmentation: 8mm for children, 12mm for women, 15mm for men, and 10mm for the elderly; γ_p∈(0,2] is output by the same MLP and determines the safety threshold; σ is Sigmoid, and τ=1mm.

[0217] The small MLP structure in this embodiment has 2–3 layers, approximately 100 parameters, takes a continuous vector of “age + gender” as input, and outputs 3 continuous coefficients γ_s, γ_m, γ_p ∈ (0,2], which are continuously differentiable throughout and have stable gradient backpropagation.

[0218] The mathematical expression for the label overflow penalty is:

[0219] δ·BalancePenalty(x)=Σ_{k=1}^5[max(0,|E complement_norm,k(x)|-ζ*|E lack_norm,k|)]2;

[0220] Where δ = 0.05 represents the penalty intensity coefficient, which is adjustable; ζ = 1.2 represents the label overflow tolerance multiplier, which is adjustable.

[0221] Its working principle and mechanism are as follows: For each element of the five-element discrete classification vector dimension k, the following is determined: if |Ecomplement_norm,k| ≤ ζ·|Emissing_norm,k|, it is considered a reasonable prediction, and this term is 0; if |Ecomplement_norm,k| > ζ·|Emissing_norm,k|, it is considered label overflow, and the square of the difference is included in the penalty. ζ = 1.2 is equivalent to allowing ±20% slight overflow; the gradient is piecewise differentiable to the square function via max(0,·), making it compatible with the Adam optimizer.

[0222] For example, when the target preference vector in a certain dimension is +0.50, the system allows a prediction range of [-0.60, +0.60]; if the predicted value is +0.70, then the penalty term is δ·(0.70-0.60). 2 =0.05·0.01=0.0005, which triggers the optimizer to call back.

[0223] Based on the above calculation results, the overall process of the jewelry parameter optimization engine in this embodiment is as follows.

[0224] Input: Preference target vector E_norm, budget ceiling, age, gender, style preference;

[0225] Optimization variables: x = [xmat, xcolor, xshape, xmass, xproc, xcombo, xwear];

[0226] Optimization algorithm: Constrained differentiable gradient descent Adam optimizer + discrete variable hard-straight-through projection;

[0227] Output: Optimal parameter vector x*.

[0228] Discrete variable processing

[0229] xmat, xshape, and xproc are reparameterized using Gumbel-Softmax.

[0230] xmass maintains a manufacturable range of 1–20g through SoftplusClip;

[0231] xcombo uses STE to handle 0 / 1 masks.

[0232] Convergence criterion: Loss is 500 iterations, and the number of iterations is adjustable.

[0233] Having obtained the optimal parameter vector, the customized scheme is generated as follows.

[0234] 1) Template Library

[0235] 3D templates: Each xshape corresponds to a parametric script (OpenSCAD / Grass operator);

[0236] Material, color, quality, and process parameters directly drive script variables;

[0237] Template output: STL / OBJ+PBR material package.

[0238] 2) Parameters

[0239] The list of 3D models is shown in Table 15.

[0240] Table 15

[0241]

[0242] 3) Example of a RESTful API message in the cloud

[0243] POST / api / generate{x}→{stl_url,gcode_url,preview_url};

[0244] Request: POST / api / generate

[0245] Content-Type: application / json

[0246] {

[0247] "x":{

[0248] "xmat":7, / / Aquamarine

[0249] "xcolor":"#0077BE", / / Deep Sea Blue

[0250] "xshape":3, / / water droplet

[0251] "xmass":8.0, / / 8g

[0252] "xproc":2, / / Gilded

[0253] "xcombo":[1,0,0,0,0], / / Only one item is worn.

[0254] "xwear":2 / / right wrist

[0255] }

[0256] }

[0257] Response: 200 OK

[0258] {

[0259] "stl_url":"https: / / factory-cdn.example.com / stl / 20250620-143200-a1b2c3.stl",

[0260] "gcode_url":"https: / / factory-cdn.example.com / gcode / 20250620-143200-a1b2c3.nc",

[0261] "preview_url":"https: / / viewer.example.com / webgl / 20250620-143200-a1b2c3"

[0262] }

[0263] This embodiment also provides the following incremental learning layer and NFT certificate module.

[0264] The incremental learning layer consists of two steps: data collection and online object updating.

[0265] Data collection:

[0266] Within 24 hours of the user wearing the device, a 5-star satisfaction rating and a repeat purchase indicator will be sent.

[0267] If sensor logs exist, record the wearing time and ambient temperature.

[0268] Online update objects:

[0269] Material - Five-element discrete classification vector residual Δ_mat; color mapping matrix W_col; ergonomic penalty small MLP weights γ_s, γ_m, γ_p.

[0270] The NFT certificate module includes contract standards, minting trigger conditions, and copyright protection processes.

[0271] Contract Standards:

[0272] Ethereum ERC-721, compatible with Polygon low-gas chain;

[0273] Metadata JSON field illustration:

[0274] {"traits":x*,"model_hash":SHA3-256(STL),"image":IPFS_CID,"timestamp":UnixTime}.

[0275] Minting trigger conditions:

[0276] The factory confirms production completion and then calls mint().

[0277] Failure rollback: If a production exception occurs, the NFT will be automatically burned.

[0278] Copyright protection:

[0279] Once the hash is on-chain, anyone can use the on-chain hash to verify the integrity of the STL.

[0280] Infringement detection script: Compares on-chain hash with public model and automatically sends DMCA notification.

[0281] The following embodiment illustrates the implementation process of the method of the present invention and proves its effectiveness through a specific case.

[0282] 1) Scene setting

[0283] User preference target vector (original): Emissing = [Metal: 0, Wood: +1.2, Water: -0.8, Fire: 0, Earth: 0]; Constraints: Age 18, Gender: Female, Wearing position: Left wrist, Single piece.

[0284] Normalization (adaptive scaling layer):

[0285] Edeficient_norm = clip(Edeficient / max(|Edeficient|,ε),-1,1)

[0286] =clip([0,1.2,-0.8,0,0] / 1.2,-1,1)

[0287] = [0,1,-0.67,0,0];

[0288] 2) The optimizer outputs the optimal parameter x as shown in Table 16.

[0289] Table 16

[0290] parameter value Meaning (Technical Solution Mapping) xmat 7 Aquamarine (material number, see C1) xcolor #0077BE Deep Sea Blue (RGB→Lab→RBF→E_col) xshape 1 Round beads (shape number, see C3) xmass 8g Weight (1–20g, see C4) xproc 1 Micro-inlaid silver wire (process number, see C5) xcombo [1,0,0,0,0] Wearing a single piece (see C6) xwear 1 Left wrist (wearing position number, see C7)

[0291] 3) Calculate Ecomplement_norm(x) term by term.

[0292] 3.1) Output vector E_mat of material culture label

[0293] BaseTable(Aquamarine) = [-0.10, -0.05, +0.75, -0.05, -0.05];

[0294] Δ_mat(aquamarine) = [+0.02,-0.01,+0.04,0,-0.01];

[0295] E_mat=[-0.08,-0.06,+0.79,-0.05,-0.06].

[0296] 3.2) Color culture label output vector E_col

[0297] Lab(#0077BE)≈(42,-9,-46);

[0298] RBF activation φ = [0.00, 0.00, 0.14, 0.00, 0.00] T ;

[0299] W_col·φ=[0.00,0.00,+0.13,0.00,0.00];

[0300] E_col=[0.00,0.00,+0.13,0.00,0.00].

[0301] 3.3) Shape culture label output vector E_shp

[0302] The characteristics of a round bead are f = [f1 = 1.00, f2 = 1.00, f3 = 0.00, f4 = 0.10], while those of a sphere are normalized.

[0303] W_shp·f=[+0.71,+0.65,+0.08,-0.01,+0.17];

[0304] E_shp=[+0.71,+0.65,+0.08,-0.01,+0.17].

[0305] 3.4) Output vector E_proc for craft culture label

[0306] Process number 1 corresponds to micro-inlaid silver wire to obtain [+0.30,0,+0.10,0,0]. Silver belongs to metal, and metal generates water.

[0307] 3.5) Mass magnification factor K_mass

[0308] K_mass=β·SoftplusClip(xmass)=0.05×8=0.40;

[0309] 3.6) Single-item prediction vector E 1

[0310] E 1 =K_mass 1 ·(E_mat 1 +E_col1+E_shp 1 +E_proc 1 )

[0311] =0.40×[0.93,0.59,1.10,-0.06,0.11]

[0312] =[0.372,0.236,0.440,-0.024,0.044]≈[0.37,0.24,0.44,-0.02,0.04].

[0313] 3.7) Total prediction vector E_total

[0314]

[0315] When worn as a single piece, xcombo = [1,0,0,0,0], i = 1, α = 0.8;

[0316] E_total = 0.8 1 ×[0.37,0.24,0.44,-0.02,0.04]

[0317] =[0.296,0.192,0.352,-0.016,0.032]≈[0.30,0.19,0.35,-0.02,0.03].

[0318] 3.8) Wearing position correction T_pos

[0319] The left wrist corresponds to xwear=1), diag([1.10,1.00,1.05,0.95,0.95]);

[0320] T_pos=[1.10,1.00,1.05,0.95,0.95];

[0321] 3.9) The final label prediction vector Ecomplement_norm(x)

[0322] E_norm_raw = T_pos·E_total

[0323] = [1.10×0.296, 1.00×0.192, 1.05×0.352, 0.95×(-0.016), 0.95×0.032]

[0324] =[0.3256,0.192,0.3696,-0.0152,0.0304]≈[0.33,0.19,0.37,-0.02,0.03].

[0325] Ecomplement_norm = clip(Ecomplement_norm_raw, -1, 1). All values ​​are within [-1, 1], so no truncation is needed. Therefore, after clipping, it equals [0.33, 0.19, 0.37, -0.02, 0.03].

[0326] 4) Calculation of the objective function Loss

[0327] 4.1) Label error item

[0328] ||E_norm_E_norm||2

[0329] =||[0,1,-0.67,0,0]-[0.33,0.19,0.37,-0.02,0.03]||2

[0330] The squared L2 norm of [-0.33, 0.81, -1.04, 0.02, -0.03]

[0331] =(-0.33) 2 +0.81 2 +(-1.04) 2 +0.02 2 +(-0.03)2

[0332] =0.1089+0.6561+1.0816+0.0004+0.0009≈1.85.

[0333] 4.2) Cost constraints

[0334] Cost(x) = P(Aquamarine) × Q(x) = 150 yuan / g × 8g = 1200 yuan

[0335] λ·Cost(x)=0.01×1200=12.0.

[0336] 4.3) Work efficiency penalty

[0337] The round bead shape is comfortable, and ShapePenalty = 0.1;

[0338] 8g≤female

[0339] Without sharp edges, SharpPenalty = 0;

[0340] Wearability(x)=0.05×(0.2×0.1+0.5×0+0.3×0)=0.001.

[0341] 4.4) Tag Overflow Penalty

[0342]

[0343] 4.5) Jewelry Design Objective Function Loss

[0344] Loss≈1.85+12.0+0.001+0≈13.85.

[0345] It should be noted that further optimization can reduce the loss, such as by adjusting the quality or process. This example only demonstrates the calculation process.

[0346] 5) The results are interpreted as shown in Table 17, including its aligned SHAP decomposition.

[0347] Table 17

[0348] Sources of contribution (including α = 0.8) Predicted value Percentage (Water Label Dimension) Material +0.79×0.40×1.05×0.8＝+0.265 71.6％ color +0.13×0.40×1.05×0.8＝+0.044 11.9％ shape +0.08×0.40×1.05×0.8＝+0.027 7.3％ process +0.10×0.40×1.05×0.8＝+0.034 9.2％ total +0.37 100％

[0349] As can be seen, the design scheme generated in this embodiment provides approximately 55% of the predicted values ​​for water labels.

[0350] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations in the above embodiments are described; only preferred embodiments of the present invention are illustrated, and their descriptions are relatively specific and detailed. However, this should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered to fall within the scope of this specification. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A personalized jewelry generative design method based on cultural tag vectors, characterized in that: Includes the following steps: Step S1: Construct a design parameter vector including material, color, shape, quality, process, combination, and wearing position; Step S2: Construct a quantitative mapping function between material, color, shape, quality, craftsmanship, and cultural label vectors, and obtain the final cultural label prediction vector by combining and correcting the data based on the wearing position; Step S3: The difference between the set five-element discrete classification vector preference target vector and the final cultural label prediction vector is used as the label error term, and a differentiable objective function is constructed by combining cost constraints, efficiency penalties and label overflow penalties; Step S4: Optimize the objective function to obtain the optimal design parameter vector, and generate the corresponding digital model design scheme for the jewelry based on the optimal design parameter vector.

2. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for the quantization mapping function between material and cultural tag vector in step S2 is: E_mat=BaseTable(i,xmat)+Δ_mat(xmat); In the formula, E_mat represents the material culture label output vector; BaseTable(i,xmat) represents the material-five-element discrete classification culture label base mapping matrix, i∈{metal,wood,water,fire,earth}, xmat represents the material type; Δ_mat represents the training residual vector corresponding to the material.

3. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for the quantization mapping function between color and cultural tag vector in step S2 is: E_col = W_col·RBF(Lab); In the formula, E_col represents the color culture label output vector; W_col represents the color-five-element discrete classification vector basis mapping matrix; Lab represents the three-dimensional vector in the CIE-Lab space; and RBF represents the radial basis function mapping centered on white, cyan, black, red, and yellow.

4. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for the quantization mapping function between shape and cultural tag vector in step S2 is: E_shp = W_shp·f; In the formula, E_shp represents the shape culture label output vector; W_shp represents the shape-pentavariate discrete classification vector base mapping matrix; f represents the geometric feature vector containing roundness or sphericity, aspect ratio or length-to-diameter ratio, sharpness, and mean curvature.

5. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for the quantization mapping function between quality and cultural tag vectors in step S2 is: K_mass=β·SoftplusClip(xmass); In the formula, K_mass represents the quality amplification factor, i.e., the quality label weighting factor; β represents the scaling factor; and xmass represents the quality.

6. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The quantitative mapping relationship between process and cultural label vectors in step S2 is obtained by looking up the process-five-element discrete classification vector basic mapping matrix.

7. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for modification through combination in step S2 is: AND i =K_mass i ·(E_mat i +E_col i +E_shp i +E_proc i ); In the formula, E total Represents the predicted vector after combination correction; N represents the number of accessories; αi represents the attenuation coefficient of the i-th accessory; xcombo[i] represents the binary mask variable of the i-th accessory, 0 indicates no superposition, 1 indicates superposition; E i E_mat represents the cultural tag prediction vector for the i-th accessory. i E_col i E_shp i and E_proc i These are the output vectors for the material culture label, color culture label, shape culture label, and craftsmanship culture label of the i-th ornament, respectively, where K_mass is the output vector. i Let be the mass magnification factor for the i-th piece of jewelry.

8. The personalized jewelry generative design method based on cultural tag vectors according to claim 7, characterized in that: The expression for obtaining the final cultural tag prediction vector by correcting the wearing position in step S2 is as follows: E complement_norm(x)=clip(T_pos(xwear)·E_total,-1,1); In the formula, E_norm(x) represents the final label prediction vector; clip represents the element-wise limiting function; and T_pos(xwear) represents the five-element discrete classification mapping vector obtained by querying the wear position correction-five-element discrete classification vector base mapping matrix based on the wearing position xwear.

9. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The expression for the objective function Loss in step S3 is: Loss=||E lack_norm-E complement_norm(x)||2+λ·Cost(x)+μ·Wearability(x;age,gender)+δ·BalancePenalty(x); In the formula, Emissing_norm represents the preference target vector; Ecomplementing_norm(x) represents the final label prediction vector; λ·Cost(x) represents the cost constraint, and λ represents the user budget sensitivity coefficient; μ·Wearability(x; age, gender) represents the efficiency penalty, and μ represents the global penalty coefficient; δ·BalancePenalty(x) represents the label overflow penalty, and δ represents the penalty intensity coefficient.

10. The personalized jewelry generative design method based on cultural tag vectors according to claim 1, characterized in that: The jewelry design scheme generated in step S4 includes a 3D model file in STL or OBJ format.

Images