Method of combinatorial optimization of yarn combination weaving parameters

Abstract

The present application belongs to the technical field of production control, and provides a combination optimization method and system for yarn combination weaving parameters. The method comprises: obtaining sample data containing first parameters such as yarn twist and elasticity and corresponding interaction characteristics, and sample labels such as pattern clarity and fabric flatness. The data is input into an initial neural network, and the loss value is calculated by combining the loss function containing the mean square error and the weighted term reflecting the physical interaction through feature transformation of the input layer and the hidden layer. The random gradient descent algorithm is used to update the parameters and train the model through back propagation when the loss value is greater than the threshold value, until the model loss value meets the requirements and does not overfit in the validation set. Through the mechanism perception layer, the physical constraints such as the influence of yarn twist and elasticity interaction on pattern clarity, fabric density and twist on flatness are embedded to construct effective interaction features, and the precise prediction of silk fabric indicators is realized.

Description

Technical Field

[0001] This invention relates to the field of production control technology, and more specifically, to a method, system, electronic device, and computer program product for optimizing yarn combination weaving parameters. Background Technology

[0002] In the silk fabric production industry, accurately predicting key indicators such as the clarity of fabric patterns and the smoothness of the fabric surface is crucial for improving product quality and optimizing production processes. Traditional prediction methods mostly rely on empirical formulas, which are difficult to fully characterize the nonlinear relationships between complex parameters such as yarn twist, elasticity, and fabric density.

[0003] Therefore, how to achieve high-precision and interpretable prediction of silk fabric indicators to guide the optimization of process parameters in actual production and improve production efficiency and product quality is an urgent problem to be solved in this field. Summary of the Invention

[0004] To address this, the present invention provides a method, system, electronic device, and computer program product for optimizing yarn combination weaving parameters, in order to at least partially solve the above-mentioned technical problems.

[0005] This invention provides a method for optimizing yarn combination weaving parameters, comprising the following steps: obtaining a first parameter, the first parameter including yarn parameters and weaving machine process parameters, wherein the yarn parameters include at least yarn twist, yarn twist direction, yarn fineness, and yarn elasticity, and the weaving machine process parameters include at least machine tension; obtaining a second parameter, the second parameter including yarn parameters and fabric structure parameters; conducting experiments based on the second parameter using a mathematical model to obtain experimental results, the experimental results including the interaction relationships between different yarn parameters; constructing interaction features based on the interaction relationships, wherein the fabric structure parameters include at least fabric interlacing density, weave structure, and structural phase; inputting the first parameter and the interaction features into a pre-trained first neural network model, and predicting fabric indices based on the first neural network model.

[0006] Optionally, the method further includes obtaining experimental results based on the second parameter through a mathematical model, which includes: screening significant factors for the second parameter based on the Plackette-Burman experimental design, determining the optimal range of each significant factor through the steepest ramp experiment method, and obtaining the interaction relationship between the significant factors based on the significant factors and their optimal range through Box-Behnken.

[0007] Optionally, the method further includes the following: the first neural network model structure includes an input layer, a hidden layer, and an output layer; wherein the input layer is used to receive the first parameters and the interaction features; the hidden layer includes a first hidden layer, a second hidden layer, and a mechanism perception layer; the first hidden layer is used to obtain the correlation features between the first parameters and the interaction features; the second hidden layer is used to extract the correlation features; the mechanism perception layer customizes a corresponding activation function based on the experimental results to form the final features; and the output layer is used to output prediction results, the prediction results including fabric pattern clarity and / or fabric smoothness.

[0008] Optionally, the method further includes a pre-training method for the first neural network model, comprising: acquiring multiple sample data features containing first parameters and interaction features, and corresponding sample labels for pattern clarity and fabric smoothness; inputting the multiple sample data features into an initial neural network model, and obtaining multiple sample output parameters, i.e., predicted values ​​for pattern clarity and fabric smoothness, through feature transformation and calculation of the input layer and hidden layer; determining the loss value of the initial prediction model based on the multiple sample output parameters and the multiple sample labels, using a loss function that includes a weighted term reflecting the interaction between twist and elasticity and the synergistic effect of density and twist, wherein the loss value includes at least a basic loss and a mechanism constraint term; if the loss value is greater than a preset loss threshold, continuing to train the initial prediction model using a stochastic gradient descent algorithm, combined with the sample data features and multiple sample labels, and updating the model parameters through backpropagation to obtain a trained model; if the training model loss value corresponding to the trained model is less than or equal to the preset loss threshold, and no overfitting occurs on the validation set, determining the trained model as the final first neural network model.

[0009] This application also provides a combined optimization system for yarn combination weaving parameters, comprising: a first acquisition module for acquiring first parameters, the first parameters including yarn parameters and weaving machine process parameters, wherein the yarn parameters include at least yarn twist, yarn twist direction, yarn fineness, and yarn elasticity, and the weaving machine process parameters include at least machine tension; a second acquisition module for acquiring second parameters, the second parameters including yarn parameters and fabric structure parameters, and obtaining experimental results based on the second parameters through a mathematical model, wherein the experimental results include the interaction relationship between different yarn shape parameters, and constructing interaction features based on the interaction relationship, wherein the fabric structure parameters include at least fabric interlacing density, weave structure, and structural phase; and an output module for inputting the first parameters and the interaction features into a pre-trained first neural network model, and predicting fabric indicators based on the first neural network model.

[0010] The present invention also provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the preceding claims.

[0011] The present invention also provides a computer program product comprising a computer program that can be executed by a processor to implement the method as described in any of the preceding claims.

[0012] The beneficial effects of this invention are as follows: By constructing interactive characteristics such as yarn twist and elasticity, and fabric density and twist, this invention accurately depicts the complex interactive relationships between various parameters. Unlike traditional methods that only consider a single factor or simple linear relationships, this invention can more realistically reflect the actual situation of the synergistic influence of multiple parameters in silk fabric production, significantly improving prediction accuracy.

[0013] By embedding physical constraints at the mechanism perception layer, the constraint model follows physical laws governing the interaction between yarn twist and elasticity on pattern clarity, fabric density, and the effect of yarn twist on smoothness. Compared to purely data-driven models, this avoids predictions that contradict physical principles, making model predictions closer to actual production conditions and improving prediction reliability.

[0014] By incorporating mechanistic constraints, such as gradient direction constraints and interaction effect constraints, into the loss function, the model's learning process is closely linked to physical mechanisms. This clearly explains how factors such as twist and elasticity affect fabric properties according to physical laws, thus improving the model's interpretability.

[0015] Based on accurate predictions, precise process parameter recommendations can be provided for silk fabric production. Production personnel can use the model predictions to optimize parameters such as yarn twist and fabric density in advance, reducing trial-and-error costs and resource waste, improving production efficiency, and lowering production costs. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a response surface diagram showing the effect of the interaction between yarn twist and yarn elasticity on the clarity of silk fabric patterns, as disclosed in the embodiments of the present invention.

[0018] Figure 2This is a response surface diagram of the effect of the interaction between fabric density and yarn twist on the smoothness of silk fabric, as disclosed in the embodiments of the present invention.

[0019] Figure 3 This is a schematic diagram of a method for optimizing yarn combination weaving parameters disclosed in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of a yarn combination weaving parameter combination optimization system disclosed in an embodiment of the present invention. Detailed Implementation

[0021] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0023] like Figure 3 As shown, this embodiment of the invention discloses a method for optimizing the combination of yarn weaving parameters, including the following steps: Step S10, obtaining a first parameter, the first parameter including yarn parameters and weaving machine process parameters, wherein the yarn parameters include at least yarn twist, yarn twist direction, yarn fineness and yarn elasticity, and the weaving machine process parameters include at least the machine tension.

[0024] In some embodiments, the first parameter includes yarn parameters: including twist (affecting yarn rigidity and luster), twist direction (Z / S affects fabric surface texture), fineness (determining the fabric's thinness), and elasticity (affecting pattern clarity and wrinkle resistance). Loom process parameters: primarily the loom tension, affecting the tightness of yarn arrangement and fabric smoothness.

[0025] In the yarn production process, a twist meter is used to measure the twist of the yarn. The twist meter clamps both ends of the yarn, fixing one end while rotating the other, and records the number of twists per unit length (e.g., per 10cm), thus determining the twist value. Different twists change the stiffness and luster of the yarn; higher twists increase yarn stiffness, make the surface smoother, and improve luster.

[0026] Twist direction is usually determined by directly observing the inclination direction of the fibers on the yarn surface. If the fibers incline from bottom to top and from left to right, it is a Z-twist; if the fibers incline from bottom to top and from right to left, it is an S-twist. Different twist directions significantly affect the fabric surface texture. Before weaving, the twist direction of the yarn must be determined according to the fabric texture design requirements. For example, using different twist directions for warp and weft yarns can produce unique fabric texture effects.

[0027] Fineness can be measured indirectly or directly. Indirect measurement methods, for example, utilize yarn linear density (tex) or denier (D), calculated using formulas by measuring the yarn's mass and length. For instance, tex represents the weight per gram of 1000 meters of yarn; denier represents the weight per gram of 9000 meters of yarn. Direct measurement methods use equipment such as microscopes or laser scanners to measure the yarn's diameter or cross-sectional area.

[0028] Yarn elasticity is obtained by tensile testing using a tensile testing machine. The yarn is clamped in the tensile testing machine and stretched to a specified length at a certain speed, then released. The elongation and recovery of the yarn are recorded, and the yarn elasticity is evaluated by calculating indicators such as the elastic recovery rate.

[0029] Loom tension: This is monitored in real time by a tension sensor installed on the loom. The tension sensor is mounted on the path of the warp or weft yarn. When the yarn is under tension, the sensor converts the mechanical force into an electrical signal. After signal processing and conversion, the signal is displayed digitally on the loom control panel. Loom tension directly affects the tightness of the yarn arrangement and the smoothness of the fabric.

[0030] Loom process parameters can also include: Loom speed: the number of weaving cycles completed by the loom per unit time (e.g., beat-ups per minute), affecting production efficiency and fabric quality, and can be obtained through a speed sensor. Shedding time: the moment the heald frames open to form the shed, usually expressed as the spindle angle (e.g., crankshaft position), affecting the stress state during yarn interlacing and fabric style, and can be obtained through a combination of a loom spindle encoder and a heald frame motion position sensor (e.g., a proximity switch). Weft insertion time and tension: Weft insertion time: the start and end times of the weft yarn entering the shed (expressed as the spindle angle), affecting the weft yarn flight stability; weft insertion tension: the tension on the weft yarn during weft insertion, affecting the uniformity of weft yarn arrangement and fabric width. Beat-ups and force: Beat-ups: the number of beat-ups per unit length of fabric, affecting fabric density; beat-ups force: the impact force exerted on the weft yarn by the reed during beat-ups, affecting the tightness of the weft yarn arrangement.

[0031] It is understood that the required parameters can be obtained according to the actual production situation. The above parameters are only examples and do not constitute a limitation on this embodiment.

[0032] Step S20: Obtain the second parameter, which includes yarn parameters and fabric structure parameters. Based on the second parameter, conduct experiments using a mathematical model to obtain experimental results. The experimental results include the interaction relationship between different yarn shape parameters. Construct interaction features based on the interaction relationship. The fabric structure parameters include at least fabric interlacing density, weave structure, and structural phase.

[0033] In some embodiments, fabric weave density can be directly measured using a fabric density microscope to count the number of warp and weft yarns per unit length (typically 10 cm). For fabrics with complex weave structures, a local magnification observation method can be used, counting the number of yarns in a specific area under a microscope and then converting the result to standard density units. In practice, multiple measurements need to be taken at different parts of the fabric (e.g., left, center, right) to obtain an average value, in order to eliminate density fluctuations during the weaving process.

[0034] Fabric structure can be determined through disassembly and microscopic observation. Disassembling the fabric allows analysis of the interlacing patterns of the warp and weft yarns to determine whether it is a plain weave, twill weave, satin weave, or other complex structure. For tightly woven fabrics that are difficult to observe directly, a microscope can be used to magnify the distribution of yarn interlacing points, and this information can be combined with weave diagram analysis software for automatic identification to determine the fabric structure type.

[0035] The structural phase can be calculated by measuring parameters such as fabric thickness and density, combined with the linear density of the yarn and the fabric density. The structural phase reflects the crimping state and interlacing tightness of the yarns in the fabric. The calculation formula typically involves the ratio of the actual fabric thickness to the theoretical yarn diameter, as well as a comprehensive analysis of the warp and weft density. In practical applications, a fabric thickness gauge can be used to measure the fabric thickness, and then the linear density of the yarn and the fabric density data can be substituted into the structural phase calculation formula for calculation.

[0036] Preferably, the step of obtaining experimental results by conducting experiments based on the second parameter through a mathematical model includes: screening significant factors for the second parameter based on the Plackette-Burman experimental design, determining the optimal range of each significant factor through the steepest ramp experiment method, and obtaining the interaction relationship between the significant factors based on the significant factors and their optimal range through Box-Behnken.

[0037] Because the appearance and pattern of silk fabrics are influenced by multiple factors, including textile structure and material morphology, the relationships between these factors are complex. This embodiment utilizes an optimized experimental design for systematic research. The Plackett-Burman design method, based on the principle of incomplete equilibrium modules, estimates the primary and secondary effects of different variables using the experimental results obtained with the fewest possible experiments, and then selects the significant influencing factors from these factors.

[0038] This embodiment first uses the Plackett-Burman design method to screen the significant factors affecting the pattern style of silk fabrics. For example, the experiment selects yarn fineness, twist, elasticity, twist direction, fabric weave density, structure, and structural phase (…). X 1, X 2, X 3, X 4, X 5, X 6, X7. Additionally, X 8, X 9, X 10 and X 11 is a blank factor). The Plackette-Burman experimental design is a two-level experimental design method based on a first-order polynomial model as shown in equation (1):

[0039] In the formula, --Predict the response value.

[0040] --Intercept of a first-order polynomial model.

[0041] -- Coefficient of the first term.

[0042] --Independent variable level.

[0043] The numbering of the factors and the level design of the Plackett-Burman design are shown in Table 1.

[0044] Table 1. Plackett-Burman Design: Horizontal Design with Numbered Factors:

[0045]

[0046] Based on the Plackett-Burman design principle, each factor is assigned two levels, a low level (-1) and a high level (+1). According to the Plackett-Burman design requirements, 12 experiments are needed for the 7 factors at 2 levels. The clarity of the appearance pattern and the smoothness of the fabric are used as response values. Through regression analysis, if the significance of each factor is above 95%, then... P If the value is less than 0.05, the factor is considered significant. If it is less than 95%, it is considered to have no significant effect.

[0047] In some embodiments, after screening the factors significantly affecting the clarity and smoothness of fabric appearance patterns using the Plackett-Burman experimental design, the steepest ramp test method is used to determine the optimal level range for each significant factor. Since the response surface fitting equation has high approximation accuracy only within the optimal response region, this region needs to be pre-defined experimentally to ensure the effectiveness of subsequent model construction.

[0048] The steepest ramp experiment systematically advances the levels of each significant factor towards the optimal response region by minimizing the number of trials. In practice, five gradient levels are set for each significant factor, and a reasonable ramp step size is determined based on the magnitude of the factor effect. During the experiment, parameter levels are gradually adjusted along the gradient direction of each factor effect, and response indicators (such as pattern sharpness score and flatness quantification) are continuously monitored. When the response value no longer significantly increases with changes in factor level, this critical point is recorded; this point is the boundary value approaching the optimal response region. This critical value will serve as the central point in subsequent Box-Behnken experimental designs, laying the foundation for constructing an accurate response surface model.

[0049] In some embodiments, based on the significant factors and their optimal regions determined by the steepest climb experiment, Box-Behnken response surface methodology is used for in-depth investigation. Box-Behnken design is a mathematical and statistical method based on a three-level partial factorial design. By constructing a quadratic response surface model, it quantifies the influence of each factor and its interaction on the response index. During the experiment, the levels of each significant factor (e.g., yarn twist, fabric interlacing density) are changed systematically, and the response index is measured simultaneously. A mathematical model between the response value and the factor variables is established through multiple quadratic regression analysis.

[0050] The experimental data obtained from the Box-Behnken design were fitted with a quadratic regression to obtain quadratic equations for the first, second, and interaction terms. The optimal values ​​were then determined within the specified range. The coded values ​​and true values ​​of each factor were converted using formula (2):

[0051] In the formula, --The encoded values ​​of each independent variable (-1, 0, +1).

[0052] --The actual value of the independent variable.

[0053] --The actual value of the independent variable at the center point.

[0054] It is the step size of the independent variable.

[0055] The significant factors selected using the Plackett-Burman design were categorized into three levels: low (-1), center (0), and high (+1). The linear relationships and interrelationships among the variables conform to the following second-order polynomial equation:

[0056] In the formula, To predict the response value. , These are the significant factors.

[0057] is the regression coefficient constant.

[0058] , and These are the coefficients of the linear term, the quadratic term, and the interaction term, respectively.

[0059] This embodiment uses Design Expert V8.0.6 software to complete the experimental design and model analysis. The response surface plots for the interaction between yarn twist and yarn elasticity on the clarity of silk fabric patterns, and the effects of fabric density and yarn twist on fabric smoothness, are shown below. Figure 1 and Figure 2 As shown.

[0060] from Figure 1 and Figure 2 It is evident that, considering the interaction between fabric density, yarn twist, and yarn elasticity, the appearance of silk fabrics is synergistically influenced by material properties, yarn morphology, and fabric geometry. Increasing yarn twist improves pattern clarity, while increasing yarn elasticity decreases it. This is primarily because the twist creates a reflective band on the yarn surface, and increasing twist enhances this reflective band. However, increased elasticity causes the reflective band to slip due to yarn elasticity, affecting pattern clarity by influencing its gloss stability.

[0061] In some embodiments, based on the interactions identified in the above experimental results, the identified interactions are transformed into specific interaction features. Based on the above experiments, the effects of twist and elasticity on pattern clarity are as follows: experiments confirm that twist improves clarity by forming reflective bands, while increased elasticity leads to slippage of the reflective bands and a decrease in clarity; the two exhibit an antagonistic relationship. The effects of density and twist on smoothness are as follows: fabric density affects the tightness of yarn interlacing; high-twist yarns are more rigid, and high density can enhance smoothness; low-twist yarns are loose, and high density can easily reduce smoothness due to uneven tension; the two have a synergistic and restrictive relationship. Exemplarily, the specific construction process includes: 1. Data acquisition and preprocessing: collecting experimental data including yarn twist (T), elasticity (E), fabric density (D), and corresponding pattern clarity (Q) and smoothness (S); normalizing the continuous variables (T, E, D) (scaling to the [0,1] interval) to eliminate dimensional differences.

[0062] II. Interactive Feature Construction: 1. Interactive feature construction of yarn twist and elasticity.

[0063] Basic product term: Directly construct the interaction feature T×E, which can linearly reflect the synergistic effect of the two. For example: high T× low E: the product value is small, corresponding to a significant improvement in sharpness (strong and stable reflection band).

[0064] High T × High E: The product value is large, which limits the improvement in sharpness (the reflection band slippage cancels the enhancement effect).

[0065] Nonlinear enhancement terms: To more accurately simulate physical mechanisms, nonlinear features such as (T×E)² or |TE| are introduced.

[0066] (T×E)²: Enhances the negative effects of the combination of high twist and high elasticity, highlighting its suppression of clarity.

[0067] |TE|: Measures the difference between twist and elasticity. The larger the difference, the more significant the competition between the two for sharpness.

[0068] 2. Interactive characteristics of fabric density and yarn twist.

[0069] Product and weighting terms: Construct T×D to represent linear interaction, reflecting the combined effect of density and twist on smoothness (e.g., a high T×high D combination may improve smoothness).

[0070] A T×D×weighting coefficient is introduced, and the weights are adjusted based on the experimental data. For example, if experiments show that twist has a greater impact on smoothness in high-density scenarios, then high-D samples are given higher weights.

[0071] Piecewise function characteristics: Based on experimental conclusions, the piecewise logic is designed as follows: When D > threshold (e.g., high density), construct T×D×1.2 (positive effect of enhancing twist).

[0072] When D < threshold (e.g., low density), construct T×D×0.8 (to weaken the effect of twist).

[0073] Step S30: Input the first parameter and the interaction feature into the pre-trained first neural network model, and predict the fabric index based on the first neural network model.

[0074] In some embodiments, the prediction process using the first neural network model includes: 1. Input data processing and feature fusion: Data standardization and alignment: The first parameter (yarn twist, elasticity, machine tension, etc.) and the interaction features (e.g., twist × elasticity, density × twist) are normalized according to the same standard used during training (e.g., Z-score standardization or Min-Max scaling) to ensure that the distribution of the input data is consistent with that used during model training. For example, if the twist feature is scaled to the [0,1] interval during training, the same scaling operation needs to be performed on the new input twist value during prediction.

[0075] Feature dimension matching: Ensure that the input feature dimensions are consistent with the design during model training. For example, if the input layer contains 10 nodes during model training (8 original features + 2 interaction features), then feature vectors of the same dimensions must be provided during prediction.

[0076] Missing value handling: If a certain interactive feature (such as twist × elasticity) cannot be directly obtained due to data acquisition limitations, it needs to be generated through interpolation or preset calculation rules. The specific generation process will not be described in this embodiment.

[0077] II. Forward Propagation Process of Neural Network Model: 1. Feature Transformation from Input Layer to Hidden Layer: First Hidden Layer: After performing a linear transformation (weighted summation) on the input features, nonlinearity is introduced through the ReLU activation function: ;in, This is the weight matrix. As a bias vector, the ReLU function ensures that the model can capture the non-linear relationships between features.

[0078] Second hidden layer: Further extracts higher-order correlation features, such as separating the synergistic effect of twist and density from the original features: .

[0079] Physical constraint implementation of the mechanism perception layer: Based on experimental results, a custom activation function is added after the hidden layer to force the model to learn physical laws. ;in, For example, a user-defined function:

[0080] To determine the relationship between twist and sharpness, a piecewise function is used: .

[0081] This simulates the phenomenon where the effect weakens after the twist exceeds a critical value.

[0082] To determine the relationship between elasticity and sharpness, a decay function is used. Increased detail leads to decreased clarity.

[0083] 3. Generation of predicted values ​​for the output layer: Regression tasks (e.g., predicting pattern clarity scores or fabric smoothness scores): .

[0084] The output is a single continuous value, which is directly mapped to the prediction space through a linear activation function.

[0085] Alternatively, a classification task can be used to output a predicted pattern sharpness level or fabric smoothness level, where... .

[0086] The output is converted into a probability distribution for each level using the Softmax function.

[0087] Preferably, the pre-training method of the first neural network model includes: acquiring multiple sample data features containing a first parameter and interaction features, as well as multiple sample labels for pattern clarity and fabric flatness.

[0088] The features of the multiple sample data are input into the initial neural network model. After feature transformation and calculation in the input layer and hidden layer, multiple sample output parameters are obtained, which are the predicted values ​​of pattern clarity and fabric flatness.

[0089] Based on the multiple sample output parameters and the multiple sample labels, a loss function containing a weighted term that includes mean square error, a term reflecting the interaction between twist and elasticity, and the synergistic effect of density and twist is used to determine the loss value of the initial prediction model. The loss value includes at least the basic loss and the mechanism constraint term.

[0090] If the loss value is greater than the preset loss threshold, the initial prediction model is trained again using the stochastic gradient descent algorithm, combined with the features of the sample data and multiple sample labels. The model parameters are then updated through backpropagation to obtain the trained model.

[0091] If the training model loss value corresponding to the training model is less than or equal to the preset loss threshold, and no overfitting occurs on the validation set, the training model is determined as the final first neural network model.

[0092] In some embodiments, exemplarily, the training process of the first neural network model specifically includes: 1. Data preparation and feature engineering: Sample data acquisition: Collect multiple sets of sample data containing first parameters (yarn twist, elasticity, density, machine tension, etc.) and interactive features (e.g., twist × elasticity, density × twist).

[0093] The measured label values ​​for pattern clarity and fabric smoothness for each sample were obtained using professional instruments (such as gloss meters and smoothness testers).

[0094] Data preprocessing: Standardization: Z-score standardization is performed on continuous variables (e.g., twist, elasticity) to eliminate the influence of dimensions. .

[0095] Interactive feature construction: Based on physical mechanisms, nonlinear interaction terms are generated, for example... and piecewise function features (e.g., in high-density scenes) ).

[0096] II. Model Architecture and Forward Propagation: 1. Network Structure Design: Input Layer: Receives the preprocessed first parameter and interaction features (e.g., 10-dimensional input vector).

[0097] Hidden layer: First hidden layer (20 neurons): Extracts basic association features, and the activation function is ReLU.

[0098] The second hidden layer (15 neurons) further refines higher-order features and introduces Dropout (0.2) to prevent overfitting.

[0099] Mechanism-aware layer: Embedding first physical constraints, such as simulating the nonlinear effect of twist on sharpness through an S-shaped function: .

[0100] Here, x represents the independent variable, such as yarn twist. This function is used to describe the relationship between twist and fabric indicators such as pattern clarity. By changing the twist value (the value of x), we can observe its effect on fabric indicators.

[0101] This represents the maximum effect value, that is, the maximum impact that twist can have on fabric indicators such as pattern clarity after reaching a certain level. For example, as twist increases, pattern clarity improves to a certain extent and then no longer improves significantly; the amount of clarity improvement corresponding to this point where it no longer improves is the maximum effect value. .

[0102] K is an adjustable constant that determines the steepness of the function curve. The larger the value of k, the steeper the curve, meaning that the twist affects the fabric properties more drastically when it approaches the threshold; the smaller the value of k, the flatter the curve, and the slower the change in influence.

[0103] The threshold is a crucial dividing point. When the twist (x) reaches this threshold, the effect of twist on fabric properties (such as pattern clarity) begins to change significantly. For example, before the threshold, an increase in twist may rapidly improve pattern clarity, but after exceeding the threshold, the rate of improvement slows down.

[0104] Embed a second physical constraint, for example, by simulating the nonlinear effect of elasticity on sharpness through an exponential decay function: Where a and b are parameters obtained by fitting experimental data. As the elasticity x increases, the function value (f(x) (representing the degree to which the sharpness is affected) gradually decreases, and the rate of decrease is fast at first and then slows down.

[0105] For example, a specific embedding method could be to input the original twist and elasticity features into a custom function module in the mechanistic perception layer of the neural network to obtain feature values ​​after nonlinear transformation. These transformed feature values, along with other features (such as fabric density), are then input into subsequent neuron calculations, allowing the model to adhere to the physical constraints of the nonlinear influence of elasticity, twist, and fabric density on clarity and smoothness when calculating predicted values ​​for pattern sharpness and fabric smoothness.

[0106] Output layer: Dual output nodes, predicting pattern sharpness and flatness respectively, using a linear activation function.

[0107] 2. Forward Propagation Calculation: Input features are transformed through each layer to generate predicted values: .

[0108] Here, x represents the input features, which are the original data input into the neural network model. These include primary parameters such as yarn twist, yarn elasticity, and fabric density, as well as processed and constructed feature data such as the interaction features between yarn twist and yarn elasticity, and the interaction features between fabric density and yarn twist.

[0109] This represents the computational operations of the hidden layer. It performs feature transformation and extraction on the input features, and through weighted connections and activation function operations between neurons, it mines the complex relationships and potential information in the input features, transforming the original input features into a more abstract and representative feature representation.

[0110] This refers to the computational operations of the mechanism perception layer. The main function of this layer is to embed physical constraints. It addresses the physical mechanisms such as the influence of the interaction between yarn twist and yarn elasticity on the clarity of silk fabric patterns, and the influence of fabric density and yarn twist on the smoothness of fabric appearance. Through specific functions, it further processes the features output by the hidden layer, enabling the model to perform calculations in accordance with physical laws.

[0111] The output layer performs computational operations. Based on the features output by the mechanism perception layer, the output layer generates predicted values ​​through linear transformations and other operations. In this embodiment, It is usually a prediction result of fabric indicators such as pattern clarity and fabric smoothness.

[0112] III. Loss Function Design and Optimization: 1. Construction of Composite Loss Function: Basic Loss: Mean Squared Error (MSE) is used to measure the deviation between the predicted value and the true label. ;in, , For predicted values, , For the true value, The sample size is a natural number greater than 0.

[0113] Mechanism constraint: Gradient direction constraint: Forced twist has a positive gradient with respect to sharpness, while elasticity has a negative gradient with respect to sharpness. ;in, , These are weighting coefficients used to adjust the relative importance of twist and elasticity to the sharpness gradient constraint. The larger their values, the stronger the influence of the corresponding constraint on the overall mechanistic constraint term. For example, if... A larger value indicates that the model training process will more strictly ensure that the gradient of twist with sharpness is positive.

[0114] Indicates the predicted value of pattern sharpness The partial derivative of yarn twist reflects the rate of change of the predicted pattern sharpness as the yarn twist changes. Physically, it represents the degree and direction of the influence of twist on sharpness.

[0115] It is the predicted value of pattern sharpness. The partial derivative with respect to yarn elasticity represents how the predicted value of pattern sharpness changes when the yarn elasticity changes, reflecting the influence of elasticity on sharpness.

[0116] Indicates taking 0 and The maximum value in the formula is used to include the gradient of twist with sharpness in the loss calculation when the gradient is negative (which does not conform to physical expectations, as twist should have a positive effect on sharpness). This prompts the model to adjust its parameters to make the gradient positive. If the gradient itself is positive, this term is 0 and does not generate additional loss.

[0117] Indicates taking 0 and The maximum value in the loss is calculated as follows: When the gradient of elasticity with respect to sharpness is positive (which does not conform to physical expectations, as elasticity should negatively affect sharpness physically), it is included in the loss and pushes the model to correct its parameters so that the gradient is negative; if the gradient itself is negative, this term is 0 and does not increase the loss.

[0118] Interaction effect constraint: Enhancing the influence of interaction terms such as twist × elasticity and density × twist: ;in, This represents a weighting coefficient used to control the importance of the twist and elasticity interaction effect constraint in the overall mechanistic constraint term. The larger the value, the stronger the constraint on this interaction effect.

[0119] Indicates the predicted value of pattern sharpness The partial derivative of the twist and elasticity interaction term (twist × elasticity) measures the degree and direction of the influence of the twist and elasticity interaction on the predicted value of pattern sharpness.

[0120] The target value, pre-set based on physical experiments or professional knowledge, represents the ideal rate of change of the effect of twist and elasticity interaction on pattern sharpness. By calculating the absolute value of the difference between the partial derivative and the target value, the model learns the relationship between the effect of twist and elasticity interaction on sharpness that conforms to actual physical conditions.

[0121] Total loss function: ;in, This represents the weighting coefficient of L2 regularization, used to weigh the importance of the L2 regularization term in the total loss function. The larger the value, the stronger the constraint on the model parameters, and the greater the effect of preventing overfitting, but it may also lead to underfitting of the model. The smaller the value, the weaker the constraint on the parameters, and the model may be at risk of overfitting.

[0122] The L2 norm, also called the Euclidean norm, represents the L2 norm of all parameters W in the model. The formula for calculating it is: ,in, These are the individual parameter values ​​in the model parameter matrix W. By penalizing the model parameters, excessively large parameter values ​​are limited, making the model smoother, reducing model complexity, and thus preventing overfitting.

[0123] 2. Algorithm and parameter update optimization: Using the Adam optimizer, for example, the initial learning rate is set to 0.001 and the batch size is 32.

[0124] When the loss value exceeds a preset threshold (e.g., 0.5), the weights are updated via backpropagation. ,in, This represents the set of weight parameters in the neural network model at the t-th training iteration. The weight parameters determine the strength of the connections between neurons in the neural network. During each training iteration, the weight parameters are adjusted based on feedback from the loss function.

[0125] This is the set of weight parameters updated during the (t+1)th training iteration. It is obtained by adjusting the weight parameters at iteration t, based on the gradient of the loss function and the set learning rate.

[0126] The learning rate, in this embodiment, is initially set to 0.001. The learning rate determines the step size for updating the weight parameters in each training iteration. An excessively large learning rate may cause the model to skip the optimal solution during training, preventing convergence; an excessively small learning rate will slow down the model training process, requiring more iterations to converge.

[0127] The loss function L represents the loss function with respect to the weight parameters. The gradient is a vector whose direction indicates the direction in which the loss function grows the fastest, and whose magnitude reflects the rate of change of the loss function at that point. During backpropagation, the gradient is calculated to determine the direction and magnitude in which the weight parameters should be adjusted, so that the loss function can be optimized in the direction of decreasing loss.

[0128] IV. Training Process and Convergence Control: Phased Training Strategy: Initial Phase: Focus on optimizing the basic loss (MSE) to quickly converge to a feasible solution.

[0129] Later stage: Gradually increase the weight of the mechanism constraint term (e.g.) (Incrementing linearly from 0.1 to 1.0) to ensure the model conforms to physical laws.

[0130] Early stopping and validation: Evaluate model performance on the validation set every 10 epochs of training: Calculate the validation loss and monitor for signs of overfitting (e.g., training loss decreases but validation loss increases). Use SHAP values ​​to analyze feature importance and ensure that interactive features (e.g., twist × elasticity) contribute to predictions as expected.

[0131] The model is finalized as follows: when the training loss is ≤0.3 and the validation set loss fluctuation is less than 5%, the current model is saved as the final version. Cross-validation (e.g., 5-fold cross-validation) is used to ensure the model's generalization ability.

[0132] See Figure 4As shown in the figure, the yarn combination weaving parameter combination optimization system 200 provided in this application embodiment includes: a first acquisition module, which acquires a first parameter, the first parameter including yarn parameters and weaving machine process parameters, wherein the yarn parameters include at least yarn twist, yarn twist direction, yarn fineness and yarn elasticity, and the weaving machine process parameters include at least machine tension.

[0133] The second acquisition module acquires second parameters, which include yarn parameters and fabric structure parameters. Based on the second parameters, experiments are conducted using a mathematical model to obtain experimental results. The experimental results include the interaction relationships between different yarn shape parameters. Based on the interaction relationships, interactive features are constructed. The fabric structure parameters include at least fabric interlacing density, weave structure, and structural phase.

[0134] The output module is used to input the first parameter and the interaction feature into a pre-trained first neural network model, and predict fabric indicators based on the first neural network model.

[0135] This invention also provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the foregoing embodiments.

[0136] This invention also provides a computer program product comprising a computer program that can be executed by a processor to implement the method as described in any of the foregoing embodiments.

[0137] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0138] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of combinatorial optimization of yarn combination weaving parameters, characterized in that, The method comprises the following steps: obtaining first parameters, the first parameters comprising yarn parameters and loom process parameters, wherein the yarn parameters at least include yarn twist, yarn twist direction, yarn fineness and yarn elasticity, and the loom process parameters at least include on-machine tension; obtaining second parameters, the second parameters comprising yarn parameters and fabric structure parameters, obtaining experimental results through mathematical modeling based on the second parameters, the experimental results comprising interaction relationships between different yarn parameters, and constructing interaction features based on the interaction relationships; inputting the first parameters and the interaction features into a pre-trained first neural network model, and predicting fabric indexes based on the first neural network model; wherein the first neural network model structure comprises an input layer, a hidden layer and an output layer; the input layer is used to receive the first parameters and the interaction features, the hidden layer comprises a first hidden layer, a second hidden layer and a mechanism perception layer, the first hidden layer is used to obtain associated features between the first parameters and the interaction features, the second hidden layer is used to refine the associated features, the mechanism perception layer defines corresponding activation functions based on the experimental results to form final features; the output layer is used to output prediction results, and the prediction results comprise fabric pattern clarity and / or fabric flatness; the mechanism perception layer defines corresponding activation functions based on the experimental results, specifically, a self-defined activation function is added after the hidden layer, wherein a segmented function is used to simulate the phenomenon that the effect is weakened when the twist exceeds a critical value for the relationship between the twist and the clarity, and a decay function is used to depict the decline of the clarity caused by the increase of the elasticity for the relationship between the elasticity and the clarity.

2. A method of combined optimization of a yarn combination weaving parameter according to claim 1, characterized in that: The experimental results obtained through mathematical modeling based on the second parameters comprise the following steps: screening significant factors of the second parameters based on Plackette-Burman experimental design, determining the optimal range of each significant factor through the steepest ascent experiment method, and obtaining the interaction relationships between the significant factors through Box-Behnken based on the significant factors and the optimal range.

3. A method of combined optimization of a yarn combination weaving parameter according to claim 1, characterized in that: The pre-training method of the first neural network model comprises: acquiring a plurality of sample data features containing first parameters and interaction features, and corresponding sample labels of pattern clarity and fabric flatness; inputting the plurality of sample data features into an initial neural network model, performing feature transformation and calculation through an input layer and a hidden layer, and obtaining a plurality of sample output parameters, that is, predicted values of the pattern clarity and the fabric flatness; determining a loss value of the initial prediction model by using a loss function containing a mean square error and a weighted term reflecting the interaction between twist and elasticity and the synergistic effect of density and twist, according to the plurality of sample output parameters and the plurality of sample labels, wherein the loss value at least includes a basic loss and a mechanism constraint term; in the case that the loss value is greater than a preset loss threshold, the initial prediction model is continuously trained by using a stochastic gradient descent algorithm in combination with sample data features and a plurality of sample labels, the model parameters are updated through back propagation, and a trained model is obtained; in the case that the trained model loss value corresponding to the trained model is less than or equal to the preset loss threshold and overfitting does not occur on the validation set, the trained model is determined as the final first neural network model.

4. A system for combined optimization of yarn assembly weaving parameters, characterized in that, Comprise: The first acquisition module acquires first parameters, the first parameters comprising yarn parameters and loom process parameters, wherein the yarn parameters at least include yarn twist, yarn twist direction, yarn fineness and yarn elasticity, and the loom process parameters at least include on-machine tension; the second acquisition module acquires second parameters, the second parameters comprising yarn parameters and fabric structure parameters, and experimental results are obtained through mathematical modeling based on the second parameters, the experimental results comprising interaction relationships between different yarn parameters, and an interaction feature is constructed based on the interaction relationships, wherein the fabric structure parameters at least include fabric interweaving density, weave structure and structure phase; The output module is used for inputting the first parameters and the interaction feature into a pre-trained first neural network model, and predicting fabric indexes based on the first neural network model; wherein the first neural network model structure comprises an input layer, a hidden layer and an output layer; wherein the input layer is used for receiving the first parameters and the interaction feature, the hidden layer comprises a first hidden layer, a second hidden layer and a mechanism perception layer, the first hidden layer is used for acquiring associated features between the first parameters and the interaction feature, the second hidden layer is used for refining the associated features, and the mechanism perception layer forms final features based on self-defined corresponding activation functions of the experimental results; The output layer is used for outputting predicted results, the predicted results comprising fabric pattern clarity and / or fabric flatness; the mechanism perception layer self-defines corresponding activation functions based on the experimental results, specifically, a self-defined activation function is added after the hidden layer, wherein a segmented function is used to simulate the phenomenon that the effect is weakened after the twist exceeds a critical value for the relationship between the twist and the clarity, and an attenuation function is used to depict the decline of the clarity caused by the increase of the elasticity for the relationship between the elasticity and the clarity.

5. A system for combined optimization of yarn assembly construction parameters according to claim 4, characterized in that, Comprise: The experiment result is obtained through a mathematical model based on the second parameters, including: screening significant factors based on Plackette-Burman experimental design on the second parameters, determining the optimal range of each significant factor through the steepest ascent experiment method, and obtaining the interaction relationship between the significant factors based on the significant factors and their optimal range through Box-Behnken.

6. A system for combined optimization of yarn assembly construction parameters according to claim 4, characterized in that, Including: The pre-training module obtains a plurality of sample data features containing first parameters and interaction features, and a plurality of sample labels corresponding to pattern clarity and fabric flatness; The plurality of sample data features are input into an initial neural network model, and after feature transformation and calculation of the input layer and the hidden layer, a plurality of sample output parameters, i.e. predicted values of pattern clarity and fabric flatness, are obtained; according to the plurality of sample output parameters and the plurality of sample labels, a loss function containing a mean square error, a weighted term reflecting the interaction between twist and elasticity, and the synergistic effect of density and twist is used to determine the loss value of the initial prediction model, and the loss value at least includes a basic loss and a mechanism constraint term; in the case that the loss value is greater than a preset loss threshold, the initial prediction model is continuously trained by using a stochastic gradient descent algorithm combined with sample data features and a plurality of sample labels, model parameters are updated through back propagation to obtain a training model; in the case that the training model loss value corresponding to the training model is less than or equal to the preset loss threshold, and overfitting does not occur on the validation set, the training model is determined as the final first neural network model.

7. An electronic device, comprising: The electronic device comprises at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, the computer program being implemented when executed by the processor to implement the method of any one of claims 1-3.

8. A computer program product, characterized by: The computer program product comprises a computer program executable by a processor to implement the method of any one of claims 1-3.

Images