Fabric performance prediction and yarn proportion backstepping method and system

Abstract

The application discloses a fabric performance prediction and yarn proportion reverse deduction method and system, and the method comprises the following steps: a fabric performance prediction and yarn proportion reverse deduction method comprises the following steps: based on a fabric production process, process chain causal gating coding is carried out to obtain a process chain representation vector; based on the process chain representation vector, cross features are extracted, and the cross features are spliced with the process chain representation vector to obtain an enhanced feature vector; based on the enhanced feature vector and a multi-base model, the joint prediction of multiple fabric performance indexes of a target fabric is carried out, and the correlation consistency constraint loss of the prediction loss and the extracted cross features is simultaneously optimized; based on the multiple fabric performance indexes of the target fabric and KAMoE meta-learning, yarn component proportion reverse solving is carried out, and the uncertainty and confidence of the solution are quantitatively evaluated through a posterior covariance. The application forms a complete closed loop of fabric performance forward prediction and reverse yarn proportion reverse deduction, and provides tool support for fabric research and development actual operation links.

Description

Technical Field

[0001] This invention relates to the field of textile fabric research and development technology, specifically to a method and system for predicting fabric performance and inversely estimating yarn proportions. Background Technology

[0002] The research and development of textile fabrics has always relied on the accumulated experience of engineers and repeated sampling and verification. Taking knitted fabrics as an example, a new product needs to go through multiple processes from raw material selection to finished product production, including fiber selection, spinning, weaving, dyeing, and finishing. The parameter selection of each process will affect many functional indicators of the final product, such as weight, bursting strength, pilling resistance, quick-drying properties, and UV resistance. When customers make demands, R&D personnel can usually only provide an initial formula based on experience. After sampling and testing, adjustments are made. This process is repeated, and the development cycle of a new product often takes several weeks or even months. High sampling waste costs and low first-time success rates are common pain points in the industry. The fundamental difficulty in this scenario is that the various properties of the fabric are determined by the raw material ratio and multi-stage process parameters. There is a physical coupling relationship between the various performance indicators, but the amount of historical R&D data actually accumulated by enterprises is extremely small, making it difficult to support sufficient data-driven modeling. In this scenario, the industry's core needs manifest in two interrelated issues: first, accurately predicting multiple finished product performance characteristics given the formula and process parameters; and second, working backward to deduce the recommended yarn fiber composition ratio given the target performance requirements.

[0003] To address performance prediction issues, existing solutions often employ direct feature concatenation or bidirectional fully connected methods to handle multi-stage process parameters. These solutions lack explicit modeling of the unidirectional causal relationship between preceding and subsequent processes, failing to effectively filter the true impact weights of historical processes on the current process. Furthermore, they are highly susceptible to downstream process parameters influencing upstream feature representations, violating the irreversible temporal physical logic of actual industrial production: fiber selection determines spinning conditions, spinning conditions constrain weaving schemes, and weaving results limit the selection of dyeing and finishing processes. Because the feature encoding stage fails to distinguish the causal direction between processes, the feature associations learned by the model contain numerous spurious correlations that violate physical facts. This is particularly prone to overfitting to false cross-process interaction patterns under extremely small sample conditions, leading to a decline in the generalization ability of the prediction results.

[0004] In the joint prediction of multiple performance metrics, existing solutions mostly rely on multi-task hard-shared networks or construct regression chains at the output. In scenarios with extremely small sample sizes, when faced with multiple prediction metrics that have complex physical positive and negative correlations, negative transfer can easily occur due to conflicting optimization directions, or the prediction error can be amplified cascaded in the output chain. Existing solutions passively process the correlation between metrics at the model output, failing to actively utilize known prior physical positive and negative correlations to generate shared cross-driving features during the feature encoding stage. Furthermore, the physical correlation constraints between performance metrics are not included in the loss function during training, resulting in a lack of physical consistency guarantees among multiple metrics in the prediction results.

[0005] For the problem of formula back-calculation, existing solutions typically simplify the back-calculation of yarn component ratios into a direct black-box mapping from target performance parameters to component ratios, or employ stochastic search strategies such as genetic algorithms and response surface methodology for formula optimization. The former fails to utilize the physical mapping relationships established in the forward prediction model, resulting in logically unrelated component predictions; the latter involves enormous computational costs, produces random results, and relies on actual sampling iterations. Yarn component ratio back-calculation is essentially a highly nonlinear, multi-objective constrained inverse problem. Different performance indicators exhibit varying sensitivities to component changes, and the ill-conditioned nature of the gradient field varies depending on the combination of objectives. Existing inverse solution methods generally use fixed solution parameters, lacking an adaptive mechanism to dynamically adjust optimization strategies based on solution difficulty, and also lacking quantification of uncertainty and credibility assessment of the solution results. Furthermore, they fail to guarantee that the solution results satisfy physical constraints such as the sum of component ratios being one and the non-negativity of each component, reducing the usability of the output formula in engineering practice. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method and system for predicting fabric performance and inversely estimating yarn proportions, enabling reliable formula inversion from fabric performance to yarn component proportions, forming a complete closed loop of forward prediction and reverse formula inversion.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] The fabric performance prediction and yarn ratio inverse deduction method of the present invention includes:

[0009] Based on the multi-stage temporal characteristics of fabric production process, process chain causal gating encoding is performed to obtain process chain representation vectors used to characterize the unidirectional causal dependency relationship of each process stage.

[0010] Based on the process chain representation vector and the pre-set prior table of positive and negative correlations between fabric performance indicators, the shared driving features of positive correlation groups and the antagonistic driving features of negative correlation groups are extracted. The shared driving features of positive correlation groups and the antagonistic driving features of negative correlation groups are then concatenated with the process chain representation vector to obtain the enhanced feature vector.

[0011] Based on the integrated network of enhanced feature vector and TabM multibase model, multiple fabric performance indicators of the target fabric are jointly predicted to obtain the fabric performance indicator prediction vector, and the prediction loss and correlation consistency constraint loss are optimized simultaneously.

[0012] Based on the target fabric performance index vector, fixed process parameters, and KAMoE meta-learning, the yarn component ratio is solved inversely, and the uncertainty and confidence of the solution are quantitatively evaluated by the posterior covariance matrix.

[0013] A further improvement of this invention lies in: performing process chain causal gating encoding based on the multi-stage temporal characteristics of the fabric production process, specifically including:

[0014] The fabric production process is divided into multiple process stages. The original features of each process stage are embedded and encoded, and the positional encoding is superimposed on the initial embedding vector of each process stage.

[0015] The initial embedding vectors of each process stage, which have been superimposed with position coding, are subjected to unidirectional causal gating modulation to obtain the causal-modulated embedding vectors of each process stage, as expressed in the following expression:

[0016]

[0017] in: For the first Embedding vectors for each process stage For the first Initial embedding vectors for each process stage For the first Location coding for each process stage For learnable scalar scaling factors, For gated modulation function, For the preceding process stage The initial embedding vector;

[0018] The causal-modulated embedding vectors of each process stage are concatenated in the order of the stages to form the process chain representation vector.

[0019] A further improvement of this invention is that the expression for the gated modulation function is:

[0020]

[0021] in: It is an element-wise sigmoid activation function. , , , All are learnable parameters. This is element-wise multiplication.

[0022] A further improvement of the present invention is that the extraction process of shared driving features of positively correlated groups and antagonistic driving features of negatively correlated groups includes:

[0023] Based on the process chain representation vector and the prior table of positive and negative correlations, shared driving features of positively correlated groups and adversarial driving features of negatively correlated pairs are extracted, expressed as follows:

[0024]

[0025]

[0026] in: For the first Positively correlated groups share driving characteristics. For the first A negatively correlated pair of adversarial driving features , , , For learnable parameters, For cross-feature dimensions, Let be a real vector space. This is a process chain characterization vector. It is the ReLU activation function. The tanh activation function;

[0027] All positively correlated groups share the same driving feature, and the negatively correlated groups share the opposing driving feature, which are then concatenated sequentially to form a cross-feature vector:

[0028]

[0029] in: For cross feature vectors, , For the first, Positively correlated groups share driving characteristics. , For the first, A negatively correlated pair of adversarial driving features Let be a real vector space.

[0030] A further improvement of this invention lies in: jointly predicting multiple fabric performance indicators of the target fabric based on an integrated network of enhanced feature vectors and the TabM multi-base model, specifically including:

[0031] Construct a TabM multi-base model ensemble network, where the number of base models is set to be [value missing]. The number of hidden layers is , No. Hidden layer to the first The calculation of each basis model is as follows:

[0032]

[0033] in: , The first , Hidden layer to the first The calculation results of the basic model This is the weight matrix. It is the bias vector;

[0034] Each base model is processed by the first After one hidden layer, a pair is generated through a linear output layer. Prediction vectors for each fabric performance index:

[0035]

[0036] in: The weight matrix of the linear output layer. For the first The output bias vectors of each base model For the first The base model of the first Hidden layer output, For the first Each base model Prediction vectors for each fabric performance index;

[0037] Pick Each base model The arithmetic mean of the prediction vectors for each fabric performance index is the final prediction vector for the fabric performance index. And predict vectors for fabric performance indicators. Apply a differentiable constraint transformation.

[0038] A further improvement of this invention is that the total loss function expression for the synchronous optimization prediction loss and the correlation consistency constraint loss of extracted cross features is:

[0039]

[0040] in: For the total loss function, A set of learnable parameters. For learnable parameters, This refers to the number of historical R&D samples used for model training and validation. For the first The loss weight of each fabric performance index, , These are the weighting coefficients for the relevance constraints. The regularization coefficient is . For the total number of negatively correlated pairs, For the first Negative correlation consistency loss for a negatively correlated pair For the first Positive correlation consistency loss for each positively correlated group The total number of positively correlated groups. , The first The first historical R&D sample Predicted and actual values ​​of various fabric performance indicators This represents the total number of fabric performance indicators.

[0041] A further improvement of this invention lies in: based on the target fabric performance index vector, fixed process parameter conditions, and KAMoE meta-learning, performing inverse solution of yarn component ratio, specifically including:

[0042] A smooth proxy model is constructed using the Spectral-KAN network with the Mish-Sine architecture, and the smooth proxy model is trained.

[0043] The condition number of the Jacobian matrix of historical R&D samples is calculated using a smooth surrogate model. Based on the condition number, a cluster space vector is constructed, and the K-Medoids algorithm is used to perform clustering in the cluster space to extract multiple real historical R&D samples as difficulty prototypes.

[0044] Based on the difficulty prototype and the target fabric performance index vector, a manifold regularized reweighted retrieval is performed to generate an initial coarse solution, global difficulty positioning features, and local manifold statistical features.

[0045] A hybrid expert meta-learner is constructed, taking an initial coarse solution, global difficulty localization features, and local manifold statistical features as inputs. The gating network of the hybrid expert meta-learner calculates the activation weights of each expert based on the global difficulty localization features, and the outputs of all experts are weighted and summed to obtain the original policy vector. ;

[0046] The original policy vector is activated by the activation function. Activation processing is performed to obtain the set of hyperparameters for the solution. ,in, The global damping factor. For preconditional vectors, For robust loss parameters, It is a relaxation factor;

[0047] Based on the Jacobian matrix and the set of hyperparameters, a robust solution for the yarn composition ratio of the target fabric is obtained through a dual correction mechanism. The reliability of the solution is verified to determine the final yarn composition ratio of the target fabric.

[0048] A further improvement of this invention lies in: based on the Jacobian matrix and the set of hyperparameters to be solved, a robust solution for the yarn composition ratio of the target fabric is achieved through a dual correction mechanism, including:

[0049] Construct a robust objective function, expressed as:

[0050]

[0051] in, For the first The update amount of the logarithmic space variable to be obtained in the next iteration. For Jacobian matrices, For the first The residual vector between the predicted fabric performance index vector and the target fabric performance index vector in the next iteration. Huber robust loss function For the first Robust objective function for the next iteration The total number of fabric performance indicators, indicated by the superscript. For transpose, This is the precondition matrix;

[0052] Calculate the double correction matrix, which consists of the iterative reweighting matrix and the preconditioning matrix. ;

[0053] Robust iterative updates within the constraint domain are performed based on the dual correction matrix to obtain the updated logarithmic space yarn composition ratio. The updated logarithmic space yarn composition ratio is then mapped back to the yarn composition ratio using a softmax transformation.

[0054] A further improvement of this invention is that the direction of robust iterative updates is:

[0055]

[0056] in, For the first The proportion of yarn components in the logarithmic space of the next iteration. This is the iterative reweighting matrix.

[0057] The fabric performance prediction and yarn ratio inverse estimation system of the present invention includes:

[0058] The encoding module is used to perform process chain causal gating encoding on the multi-stage temporal characteristics of fabric production process, and obtains process chain representation vectors to represent the unidirectional causal dependencies of each process stage.

[0059] The feature extraction module is used to extract the shared driving features of positively correlated groups and the adversarial driving features of negatively correlated groups, and concatenate the shared driving features of positively correlated groups and the adversarial driving features of negatively correlated groups with the process chain characterization vector to obtain the enhanced feature vector;

[0060] The joint prediction module is used to jointly predict multiple fabric performance indicators of the target fabric based on the enhanced feature vector and the TabM multibase model integrated network, to obtain the fabric performance indicator prediction vector, and simultaneously optimize the prediction loss and the correlation consistency constraint loss.

[0061] The inverse solution module is used to inversely solve the yarn component ratio based on the target fabric performance index vector, fixed process parameters, and KAMoE meta-learning, and to quantitatively evaluate the uncertainty and confidence of the solution results through the posterior covariance matrix.

[0062] The beneficial effects of this invention are: by deploying a one-way causal gating modulation mechanism between each process stage, this invention reduces the risk of reverse interference of subsequent parameters on the feature representation of preceding processes by only allowing the preceding process stage to transmit gated and filtered information to the subsequent process stage, reduces the possibility of spurious correlations introduced due to confusion of causal direction, and keeps the feature representation of the process chain consistent with the irreversible timing of actual production.

[0063] This invention constructs positively correlated shared driving features and negatively correlated adversarial driving features at the feature input end based on the known physical correlation between performance indicators. It is also equipped with a joint training framework that integrates a correlation consistency constraint loss function with a multi-base model, so that the prediction results of multiple fabric performance indicators maintain intrinsic consistency in physical coupling relationship, while providing reliable confidence estimation.

[0064] This invention, based on forward prediction, utilizes a smooth surrogate model that maintains a distillation approximation relationship with the forward surrogate model. It calculates gradients and geometric information such as the Jacobian matrix, residual vector, and condition number at the current logarithmic space variables to drive iterative back-calculation. The solution strategy is adaptively adjusted according to the geometric difficulty of the current solution task. The softmax simplex mapping ensures that the yarn component ratios output in each iteration satisfy the physical constraints of summing to 1 and each yarn component being non-negative. The uncertainty of the back-calculation results is quantified by the posterior covariance matrix, and a recommended formula with confidence intervals is output, making the back-calculation results engineering-ready. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0067] Combination Figure 1 As shown, the fabric performance prediction and yarn ratio inverse deduction method in this embodiment includes a forward fabric performance prediction process and a reverse yarn component ratio inverse solution process, specifically including:

[0068] Step 1: Based on the multi-stage temporal characteristics of the fabric production process, perform process chain causal gating encoding to obtain a process chain representation vector that represents the unidirectional causal dependency relationship of each process stage.

[0069] Step 2: Based on the process chain representation vector and the preset prior table of positive and negative correlations between fabric performance indicators, extract the shared driving features of positive correlation groups and the antagonistic driving features of negative correlation groups, and concatenate the shared driving features of positive correlation groups and the antagonistic driving features of negative correlations with the process chain representation vector to obtain the enhanced feature vector.

[0070] Step 3: Based on the integrated network of enhanced feature vector and TabM multibase model, the multiple fabric performance indicators of the target fabric are jointly predicted to obtain the fabric performance indicator prediction vector after physical constraint transformation, and the prediction loss and correlation consistency constraint loss are optimized simultaneously. After training, Steps 1 to 3 together constitute a positive surrogate model, which is used to predict the fabric performance indicators of candidate yarn component ratios and provide distilled soft labels to the smooth surrogate model in Step 4.

[0071] Step 4: Based on the target fabric performance index vector, fixed process parameters, smoothed surrogate model, and KAMoE meta-learning, the yarn component ratio is solved inversely. The uncertainty and confidence of the solution are then quantitatively evaluated using the posterior covariance matrix. The target fabric performance index vector is the target performance vector defined by design requirements, customer needs, or manually set parameters.

[0072] In step 1 of this embodiment, the process chain causal gating encoding arranges the raw material composition characteristics and parameter characteristics of the four process stages—spinning, weaving, dyeing, and finishing—according to the production sequence. Through a one-way causal gating mechanism, selectively filtered conditional influence information from preceding stages is injected into each subsequent stage, generating a process chain representation vector that integrates the causal relationships across the entire chain. This solves the problem of existing methods losing one-way causal dependencies between processes due to simple feature splicing and failing to block reverse information interference. Specific operations include:

[0073] The complete production process of fabric from raw material to finished product is divided into five technological stages, numbered sequentially according to production time. The five stages are fiber selection, spinning process, weaving process, dyeing process, and finishing process, respectively.

[0074] The raw features at each process stage include two categories: categorical features and numerical features. For categorical features, a learnable embedding lookup table is used to obtain categorical embedding vectors; for numerical features, piecewise linear coding is used to obtain numerical coding vectors. Categorical features include fiber type, twist direction, spinning method, etc.; numerical features include fiber ratio, fiber fineness, yarn specifications, and final width, etc. For scalar values ​​within numerical features... Given a set of ordered breakpoints The piecewise linear coding function is:

[0075]

[0076] in, It is a numerical feature. For a pre-defined set of ordered breakpoints, It is a constant. For piecewise linear encoding dimensions, Numerical features The result obtained after segmented linear encoding 3D encoded vector.

[0077] After concatenating all categorical embedding vectors and numerical encoding vectors within the same process stage, a linear transformation is applied to map them to a unified dimension. The expression is as follows:

[0078]

[0079] in, For the first Initial embedding vectors for each process stage For the first A set of categorical original features for each process stage. For the first A set of numerical original features for each process stage. For categorical feature embedding functions, For piecewise linear coding functions of numerical features, This is a vector concatenation operation. and The first The stage mapping weight matrix and bias vector corresponding to each process stage are used as stage embedding parameters. Specifically, the original features of stage one, i.e., the process parameters of the fiber selection stage, include fiber parameters such as fiber type, fiber fineness, fiber strength, fiber elongation, and fiber ratio; the original features of stage two, i.e., the process parameters of the spinning stage, include yarn parameters such as yarn specifications and spinning method; the original features of stage three, i.e., the process parameters of the weaving stage, include weaving parameters such as cam configuration; the original features of stage four, i.e., the process parameters of the dyeing stage, include dyeing parameters such as auxiliary agent concentration; and the original features of stage five, i.e., the process parameters of the finishing stage, include finishing parameters such as finishing temperature and finishing width.

[0080] Learnable positional encoding vectors are assigned to each of the five process stages. This is used to provide the model with explicit identification information about the position of each stage in the production process, where, It is a real number vector space. Position encoding is added element-wise to the initial embedding vector of the corresponding process stage.

[0081] The initial embedding vectors of each process stage, which have been superimposed with position coding, are subjected to unidirectional causal gating modulation to obtain the causal-modulated embedding vectors of each process stage, as expressed in the following expression:

[0082]

[0083] in: For the first Embedding vectors for each process stage For the first Location coding for each process stage For learnable scalar scaling factors, For gated modulation function, For the preceding process stage The initial embedding vector, for the first process stage, i.e. There are no preceding process stages, therefore .

[0084] The expression for the gated modulation function is:

[0085]

[0086] in: The element-wise sigmoid activation function has an output range of . As a gate vector, , These are learnable parameters, used as information vectors. This is element-wise multiplication.

[0087] One-way causal constraints are achieved through summation upper bounds. To achieve this, information from subsequent stages must be strictly prohibited from flowing back into the representation of preceding processes. Selective gating is implemented using a sigmoid function; when the gating value approaches zero, the information flow of the corresponding preceding process is closed, and when it approaches one, it is allowed. The selection of information dimensions can be learned through linear transformation. This is achieved by extracting information dimensions useful for the current process from the initial embedding vectors of the preceding process stages. The gate vector and the information vector are then multiplied element-wise, and then scaled by a learnable scalar scaling factor. By scaling and injecting the current process stage characterization, fine-tuning of the influence intensity of previous process stages is achieved.

[0088] This embodiment divides the five process stages into three stage groups according to their functional attributes: raw material yarn group. Weaving and Manufacturing Group Dyeing and Finishing Group For the current stage of technology Compared with the preceding process stage For preceding gates belonging to the same stage group, the gate parameters within that stage group are reused; for the current process stage... Compared with the preceding process stage Cross-group gating belonging to different stage groups requires independent gating parameters set according to the inter-group transfer relationship. Therefore, the five process stages require pairwise unidirectional connections. The group gating parameters are categorized into five groups: in-group raw yarn gating, in-group dyeing and finishing gating, raw yarn group to weaving group gating, weaving group to dyeing and finishing group gating, and raw yarn group to dyeing and finishing group gating. These gating parameters include... and .

[0089] The linear transformation matrix in each gating group is decomposed into a low-rank form:

[0090]

[0091] in, , , , All are low-rank decomposition matrices. , is the rank parameter of the low-rank decomposition. , indicating rank Matrix product approximation The gating matrix reduces the number of learnable parameters and lowers the risk of overfitting under very small sample conditions. The superscript T stands for transpose.

[0092] The causal-modulated embedding vectors of each process stage are concatenated in stage order to form a process chain representation vector:

[0093] .

[0094] In step 2 of this embodiment, based on the process chain representation vector and a pre-defined prior table of positive and negative physical correlations between fabric performance indicators, shared driving features of positively correlated groups and antagonistic driving features of negatively correlated groups (i.e., cross features) are extracted. These extracted cross features are then concatenated with the process chain representation vector from step 1 to form an enhanced feature vector. This addresses the problem of existing methods passively processing indicator correlations only at the output end, leading to negative transfer and error cascading in small sample sizes. Specifically, this includes:

[0095] definition The values ​​of each fabric performance index are Construct a prior correlation matrix :

[0096] .

[0097] Based on the prior correlation matrix and the pre-defined prior table of positive and negative correlations between fabric performance indicators, the positively correlated fabric performance indicators are divided into... positively correlated groups negative correlation Divided into One negative correlation pair .

[0098] Extract the shared driving features of the positively correlated groups, for the first group positively correlated groups Define a shared-driven feature extractor to represent vectors using the process chain. For input:

[0099]

[0100] in: , For the first Learnable parameters of the shared-drive feature extractor corresponding to each positively correlated group. For cross-feature dimensions, , Let be a real vector space. It is the ReLU activation function. For the first Positively correlated groups share driving characteristics. This method is used to extract shared information from the process chain representation vector that can jointly drive the coordinated changes in positively correlated fabric performance indicators within a group. This embodiment employs the ReLU activation function to ensure that the shared driving features of positively correlated groups maintain non-negative expression, and combines this with positive correlation consistency loss to ensure that the shared driving features of positively correlated groups tend to capture information about the coordinated changes in fabric performance indicators within the group during training.

[0101] Extracting negatively correlated adversarial driving features, for the first One negative correlation pair , Define an adversarial-driven feature extractor, which also uses the process chain to represent vectors. For input:

[0102]

[0103] in: , For learnable parameters, The tanh activation function is used. For the first A negatively correlated pair of adversarial driving features The goal is to capture information that causes the inverse relationship between two fabric performance indicators in a negatively correlated pair. This embodiment uses the tanh activation function to make the output range [missing information]. .

[0104] All positively correlated groups share the same driving feature, and the negatively correlated groups share the opposing driving feature, which are then concatenated sequentially to form a cross-feature vector:

[0105]

[0106] in: For cross feature vectors, , For the first, Positively correlated groups share driving characteristics. , For the first, A negatively correlated pair of adversarial driving features Let be a real vector space.

[0107] Combine the cross feature vector with the process chain representation vector from step 1. The features are concatenated to form an enhanced feature vector. :

[0108]

[0109] in, Let be a real vector space.

[0110] Simply constructing a network structure with cross-features is insufficient to guarantee that the shared-driven feature extractor and the adversarial-driven feature extractor actually capture the corresponding positive and negative correlation patterns. Therefore, this embodiment designs a correlation consistency constraint loss function, which is optimized synchronously with the prediction loss in step 3. The gradient is shared through the positive correlation group driving features. and negative correlation with adversarial driving features Backpropagation is performed to the cross-feature extraction parameters respectively. , , , and scalar projection in correlation consistency constraints , Simultaneously, the gradient is further characterized as a vector via the process chain. Backpropagation is performed to the stage embedding parameters, position encoding vectors, and gating parameters from step 1, achieving end-to-end joint optimization from steps 1 to 3. For the... For each positively correlated group, calculate the scalar projection of the shared driving feature of the positively correlated groups. Let scalar projection The Pearson correlation coefficient between the predicted values ​​of each fabric performance index within the group and the historical R&D sample set used for model training should be as large as possible. The expression for the positive correlation consistency loss is:

[0111]

[0112] in: For fabric performance indicators The predicted value, For the first Positive correlation consistency loss for each positively correlated group Calculate the Pearson correlation coefficient.

[0113] For the first One negative correlation pair , , For fabric performance indicators and The value of is used to calculate the scalar projection of the negatively correlated adversarial driving features. Let scalar projection The correlation coefficients of the predicted values ​​of the two fabric performance indicators in a negatively correlated pair are in the opposite direction, i.e., the expression for the negative correlation consistency loss is:

[0114]

[0115] in: For fabric performance indicators The predicted value, For the first Negative correlation consistency loss for each negatively correlated pair.

[0116] Step 3 involves inputting the enhanced feature vectors into the TabM multi-base model ensemble network for joint prediction of multiple fabric performance indicators. During training, the prediction loss and correlation consistency constraint loss are simultaneously optimized to obtain the predicted values ​​and confidence intervals for multiple fabric performance indicators. Specifically, this includes:

[0117] Set the number of base models to The number of hidden layers is The enhanced feature vectors are input into the TabM multi-base model ensemble network, the first... Hidden layer to the first The calculation of each basis model is as follows:

[0118]

[0119] in: , The first , Hidden layer to the first The calculation results of each basic model, when hour, In this embodiment, all base models share the same enhanced feature vector input. The weight matrix is ​​the weight matrix. exist The base models are fully shared. The bias vector, the bias vector Each base model has its own independent bias vector. This design achieves this through the diversity of bias vectors without increasing the number of parameters. The differentiation of each base model in the embedding space. The shared weight matrix ensures that all base models learn the same feature interaction pattern, while having independent bias vectors allows each base model to have different decision boundary offsets in the feature space, thereby achieving variance reduction in the ensemble output.

[0120] Each base model goes through the first After one hidden layer, a pair is generated through a linear output layer. Prediction vectors for each fabric performance index:

[0121]

[0122] in: The prediction vector generated for the linear output layer. Let be a real vector space. The weight matrix of the linear output layer. For the first The output bias vectors of each base model For the first The base model of the first Hidden layer output, For the first Each base model A prediction vector for each fabric performance index.

[0123] Prediction vector Pick The arithmetic mean of the basic models:

[0124]

[0125] For fabric performance indicators with physical limits, a differentiable constraint transformation is applied. For non-negative fabric performance indicators, such as weight per unit area and bursting strength, Softplus activation is preferably applied.

[0126]

[0127]

[0128] For fabric performance indicators with upper and lower bounds, apply tanh scaling:

[0129]

[0130] in: , The first The maximum and minimum physical limits of each fabric performance index. For the predicted first A linear prediction value without physical constraint transformation. For the first Each predicted value of fabric performance index is obtained through non-negativity constraint or upper and lower bound constraint transformation, resulting in a fabric performance index prediction vector. include Predicted values ​​for fabric performance indicators.

[0131] The total training loss of the positive surrogate model is a weighted combination of the relevance consistency constraint loss from step 2, the prediction loss from step 3, and the weight decay regularization loss, achieving end-to-end joint optimization of steps 1 to 3. The expression is:

[0132]

[0133] in: For the total loss function, The set of learnable parameters specifically includes: a learnable embedding lookup table for obtaining categorical embedding vectors, relevant parameters of the piecewise linear coding function, and a stage mapping weight matrix. Bias vector Location coding Gating parameters , and the corresponding low-rank decomposition parameters and learnable scalar scaling factors Learnable parameters , , and Scalar projection in correlation consistency constraints and Weight matrix Bias vector The weight matrix of the linear output layer and the output bias vectors of each base model; For learnable parameters, This refers to the number of historical R&D samples used for model training and validation. For the first The loss weight of each fabric performance index, , These are the weighting coefficients for the relevance constraints. The regularization coefficient is . For the total number of negatively correlated pairs, For the first Negative correlation consistency loss for a negatively correlated pair For the first Positive correlation consistency loss for each positively correlated group The total number of positively correlated groups. For the first The first historical R&D sample Predicted values ​​of each fabric performance index For the first The first historical R&D sample The true values ​​of each fabric performance index. This embodiment uses the Adam optimizer for gradient calculation and parameter update, with the gradient passed through the total loss function. Backpropagation is performed to all learnable parameters from steps 1 to 3 to achieve global end-to-end optimization.

[0134] Given the small sample size of the dataset, this embodiment employs leave-one-out cross-validation for model performance evaluation. The dataset consists of N historical R&D samples, each including yarn composition ratios, process parameters for five process stages, and a vector of measured fabric performance indicators. Each time, the [number of samples] is selected... One historical R&D sample is used as the validation sample, and the remaining N-1 historical R&D samples are used as the training samples. The total loss function is used. Train a positive surrogate model containing all model structures from steps 1 to 3, and make predictions on validation samples, recording the predicted values. Traverse all After reviewing historical R&D samples, evaluation indicators for each fabric performance index were calculated:

[0135] Evaluation of the first The coefficient of determination for each fabric performance index, determined by leave-one-out cross-validation, is expressed as follows:

[0136]

[0137] in:

[0138] in: For the first Leave-one-out cross-validation coefficient of determination for each fabric performance index When using leave-one-out cross-validation, the first The first historical R&D sample Predicted values ​​for each fabric performance index for The first historical R&D sample The average of the true values ​​of each fabric performance index.

[0139] N complete models generated by leave-one-out cross-validation are used to test new samples. Perform complete forward propagation predictions from step 1 to step 3 respectively, and obtain N predicted values. Calculate the predicted mean and standard deviation .in, This represents the model input, which consists of the candidate yarn component ratios of the target fabric and fixed process parameters; during the inverse solution process, the candidate yarn component ratios obtained in each iteration are... When combined with fixed process parameters, corresponding new sample inputs can be generated. The predicted standard deviation for the new sample. The confidence interval used to construct an approximate 95% positive prediction of fabric performance indicators is used in step 4 as the confidence assessment and preconditioning vector for inverse solution. The generated reference information, where, for the first For each fabric performance index, a confidence interval with approximately 95% positive prediction is constructed: .

[0140] In step 4, based on the forward surrogate model constructed in steps 1 to 3, the reverse solution from the target fabric performance parameters to the optimal yarn composition ratio is realized. The forward surrogate model generates a fabric performance index prediction vector under the conditions of candidate yarn composition ratios and fixed process parameters, and forms the basis for prediction uncertainty and calibration; the fabric performance index prediction vector is used for the reverse solution in step 4. Step 4 uses the target fabric performance index vector and fixed process parameter conditions as constraints, and the yarn composition ratio as the variable to be solved. A smooth surrogate model is used to continuously and differentiably predict the fabric performance corresponding to the candidate yarn composition ratios. The yarn composition ratio is iteratively updated based on the deviation between the predicted performance and the target performance, finally obtaining the recommended yarn composition ratio. The input to step 4 is the target fabric performance index vector. And fixed process parameter conditions; fixed process parameter conditions refer to process parameters that have been determined or preset in this reverse engineering task and are not used as variables to be determined, including spinning method, yarn specifications, weaving parameters, dyeing parameters, finishing temperature, finishing width, etc. Step 4 outputs the optimal yarn composition ratio and confidence interval. Specifically, it includes:

[0141] To reduce the nonlinearity between the yarn composition ratio and the fabric performance index vector, the yarn composition ratio... Perform logarithmic transformation operation:

[0142]

[0143] in: It is a constant, and its value is... Used to prevent mathematical singularities. The proportion of yarn components in logarithmic space.

[0144] A smooth surrogate model is established using the Mish-Sine architecture of Spectral-KAN, which is to establish a mapping function from the yarn component ratio in logarithmic space to the fabric performance index vector.

[0145] The expression for the smooth surrogate model is:

[0146]

[0147] in, This is the fabric performance index prediction vector output by the smooth surrogate model. For mapping functions, The input feature vector is composed of the yarn component ratios in logarithmic space and fixed process parameters. The number of spectral expansion layers in the smooth surrogate model. For the spectral expansion layer index, The number of spectral basis functions per layer. Index of the basis functions for each spectral layer. This represents the learnable output weight vector corresponding to the Mish component. This represents the learnable output weight vector corresponding to the Sine spectral components. and Learnable parameters in the Mish component, superscript For transpose, For frequency parameters, For phase parameters, For smooth activation function, It is a sine function.

[0148] The smooth proxy model learns the measured performance labels simultaneously. And the soft labels output by the positive proxy model Among them, the measured performance label is the vector of measured fabric performance indicators from historical R&D samples, and the soft label is the predicted vector of fabric performance indicators. The joint training loss of the smooth surrogate model is:

[0149]

[0150] in, The weights of the measured performance label supervision items, The weighting coefficients for the smoothing regularization term, This is the regularization term for gradient smoothness.

[0151] Using the trained smooth surrogate model, implicit physical geometric properties are extracted to quantify the difficulty of inverse solving and to generate an initial coarse solution. This embodiment trains the smooth surrogate model using historical R&D samples, which include yarn composition ratios, process parameters for five process stages, and measured fabric performance index vectors. The forward input includes yarn composition ratios and fixed process parameter conditions, along with corresponding measured performance labels. And the soft labels output by the positive proxy model As output labels. For the first... A historical R&D sample, with the yarn component ratio set as follows: The fabric performance index vector is Fixed process parameters are .Will After logarithmic transformation And the proportion of yarn components in logarithmic space and The smooth surrogate model, trained by common input, is used to calculate the Jacobian matrix through automatic differentiation. .according to Calculate the condition number Used to characterize historical R&D samples The degree of ill-conditioning in the inverse solution of the vicinity. Constructing the cluster space vector. ,in To represent the difficulty of the solution, the dimensions of the cluster space vector are standardized to ensure that performance metrics and solution difficulty have a balanced influence on clustering. The K-Medoids algorithm is used to cluster the data in the cluster space. Euclidean distance is used as the distance metric between samples during clustering. A K-Medoids initialization strategy is used to select four real historical R&D samples as initial centroids. Clustering iterations are completed according to a set termination condition. These four real historical R&D samples are extracted as difficulty prototypes. Based on the median value of the solution difficulty within each cluster, the clusters corresponding to the four difficulty prototypes are sequentially labeled as easy, medium, hard, and extremely hard solution conditions.

[0152] For the target fabric, the target fabric performance index vector is: First, an initial neighborhood is selected from historical R&D samples based on performance distance. Then, the historical condition number of this initial neighborhood is used to weight and estimate the condition number of the target back-introduction task. ,based on Constructing target cluster space vectors The objective back-calculation task involves deriving the corresponding yarn component proportions from the target fabric performance index vector. The estimation condition number for this objective back-calculation task... The calculation process includes:

[0153] Standardize the fabric performance index vectors of historical R&D samples and the target fabric performance index vectors:

[0154]

[0155]

[0156] in, This is the standardized vector of target fabric performance indicators. For the standardized first A vector of fabric performance indicators from historical R&D samples. For performance index standardization transformation;

[0157] Based on the standardized target performance index vector, an initial neighborhood is selected according to the performance distance:

[0158]

[0159] in, For the initial neighborhood, To sort the evaluation indicators from largest to smallest, select the top... A historical research and development sample.

[0160] Calculate performance similarity weights for historical R&D samples in the initial neighborhood. :

[0161]

[0162] in, This is the performance distance scale parameter.

[0163] The expression for calculating the estimated condition number of the target inverse task is as follows:

[0164]

[0165]

[0166] in, It is a constant.

[0167] The manifold regularized reweighted retrieval process includes:

[0168] Calculate the first Reference weights for each historical R&D sample:

[0169]

[0170] The normalized reference weights are:

[0171]

[0172] in, For the first The reference weight of each historical R&D sample The normalized reference weights, To estimate the condition number of the task by working backward from the objective, The geometric difficulty distance scale parameter. This represents the weighting coefficient for the geometric difficulty term. In this embodiment, the measured fabric performance index vector is missing from historical R&D samples. At that time, the predicted values ​​of the fabric performance index vector obtained by the positive surrogate model or the smooth surrogate model are completed.

[0173] The historical R&D samples were sorted from largest to smallest according to their normalized reference weights, and the top [number] samples were selected. Each historical R&D sample constitutes a neighborhood historical R&D sample set. The reference weights corresponding to the selected historical R&D samples are then normalized again:

[0174]

[0175] Initial coarse solution for:

[0176]

[0177]

[0178]

[0179] in, Features are used to locate global difficulty and represent the target clustering space vector. The distance vectors to the four difficulty prototypes are used to macroscopically determine the difficulty range to which the target-based task belongs. , , and These are cluster space vectors representing four difficulty prototypes: easy, medium, hard, and extremely hard. Local manifold statistical characteristics. Includes a collection of historical R&D samples from the surrounding area. Mean of internal condition number Variance of yarn component ratio and neighboring historical R&D sample collection variance of internal condition number .

[0180] A physics-guided hybrid expert meta-learner is constructed, taking an initial coarse solution, global difficulty localization features, and local manifold statistical features as inputs. Specifically, the initial coarse solution... The initial coarse solution converted to logarithmic space is used as the initial iteration point in the IRLS-GN solution process. Then, the global difficulty localization features of the initial coarse solution in logarithmic space are concatenated with the local manifold statistical features to obtain the input of the hybrid expert meta-learner. In the hybrid expert meta-learner, each expert is a Spectral-KAN network, expressed as:

[0181]

[0182] in, To solve for the policy vector, For expert serial numbers, For the first One Spectral-KAN network, , For learnable weight matrix, This is a learnable frequency parameter.

[0183] The gating network of the hybrid expert meta-learner employs the Softmax function and is based on global difficulty localization features derived from the objective-based task inversion. The activation weight for each expert is calculated using the following expression:

[0184]

[0185] For the The training task formed by a set of historical R&D samples is represented as:

[0186]

[0187] For the target reverse engineering task, it is represented as:

[0188]

[0189] The temperature coefficient is constrained to a positive value using the Softplus function, and is expressed as follows:

[0190]

[0191] in, For the first The activation weight of each expert For the number of experts, To determine the global difficulty characteristics of a task by working backward from the target. , The first , The learnable scale parameter corresponding to each expert , The first , The learnable bias parameters corresponding to each expert The initial temperature coefficient, This is a temperature coefficient used to control the smoothness of the gating distribution. , and The KAMoE meta-learner is obtained through joint learning via historical back-reasoning tasks during training. For the... The training task formed from historical R&D samples, and the corresponding global difficulty localization features are denoted as... .

[0192] Each expert outputs an inactive solution strategy vector. , Including the global damping factor Precondition vector Robust loss parameters and relaxation factor The corresponding original parameter components. Expert activation weights obtained from the gating network. The original policy vector is obtained by weighted summation of all expert outputs. :

[0193]

[0194]

[0195] in: , , and All are original policy vectors The components are used to generate the global damping factor, precondition vector, robust loss parameter, and relaxation factor, respectively, with the superscript T indicating transpose.

[0196] For the original policy vector By applying corresponding activation functions to different components in the solution, a set of hyperparameters can be obtained. .in, Activation is performed using the Softplus function to ensure a positive value. Non-negative preconditioning vectors are obtained by activation using the Softplus function, and these vectors are used to construct the preconditioning matrix. , The dimension and the number of yarn components to be determined Consistent Mapping to a preset robust parameter range via the Sigmoid function [ , ], The Sigmoid function maps the data to the (0,1] interval to scale the update step size; the expression is:

[0197]

[0198]

[0199]

[0200]

[0201] .

[0202] An improved iterative reweighted least squares-Gauss-Newton (IRLS-GN) solution is performed. The solution process utilizes the Jacobian matrix provided by the smooth surrogate model, combined with the adaptive solution hyperparameters output by the KAMoE meta-learner. Robust solution for yarn component ratios is achieved through a dual correction mechanism consisting of an iterative reweighting matrix and a preconditioning matrix, along with softmax physical constraint projection. Specifically, this includes:

[0203] To construct a robust objective function and reduce the impact of abnormal fabric performance indicators and local prediction errors on the solution results, a robust objective function is constructed in the first step. In the next iteration, a robust objective function is constructed, expressed as:

[0204]

[0205] in, For the first The update amount of the logarithmic space variable to be obtained in the next iteration. Let be the Jacobian matrix, and let be the result of the trained smooth surrogate model in the th case. Yarn component proportions in the logarithmic space of the next iteration The value is obtained through automatic differential calculation. For the first The residual vector between the predicted fabric performance index vector and the target fabric performance index vector in the next iteration. For the first Robust objective function for the next iteration The Huber robust loss function is defined by the robust loss parameters. Control, the expression is: .

[0206] Reverse solver The fabric performance index prediction vector in the next iteration satisfies:

[0207]

[0208] in, For the first The fabric performance index prediction vector obtained from the next iteration. For the first The proportion of yarn components in the logarithmic space of the next iteration. This is to establish fixed process parameters.

[0209] No. The fabric performance index prediction vector for the next iteration is calculated by the trained smooth surrogate model:

[0210]

[0211]

[0212]

[0213] in: For the smooth surrogate model in the first The output of the nth iteration Predicted values ​​for each fabric performance index For the first The iteration of the ... The predicted value of the fabric performance index is the same as the first fabric performance index prediction value. Target fabric performance indicators The residuals between them.

[0214] Calculate the double correction matrix, which consists of the iterative reweighting matrix and the preconditioning matrix. The expression is:

[0215]

[0216]

[0217]

[0218]

[0219] in: Let H be the first derivative of the Huber robust loss function with respect to the residual variable. Let be the second derivative of the Huber robust loss function with respect to the residual variable. , , These are the first and second results obtained when the Huber robust loss function is converted to a locally weighted quadratic form at the current residual. , Equivalent weights of each fabric performance index This is the iterative reweighting matrix. The iterative reweighting matrix and the preconditioning matrix acting on the logarithmic space of the yarn components... The solution process is corrected by modifying the residual space of fabric performance indicators and the variable space of yarn composition respectively.

[0220] Solving for the IRLS-GN update direction based on the iterative reweighting matrix and the preconditioning matrix:

[0221] .

[0222] Construct candidate logarithmic spatial variables, candidate yarn component ratios, candidate fabric performance index prediction vectors, and candidate residuals:

[0223]

[0224]

[0225]

[0226]

[0227] in, For candidate logarithmic space variables, The proportion of candidate yarn components. This is the prediction vector for the performance indicators of candidate fabrics. For candidate residual vectors, , , and By the The iteration is based on trial calculations of candidate update steps.

[0228] When a candidate update is accepted, let:

[0229]

[0230]

[0231]

[0232] in, For the first The proportion of yarn components in the logarithmic space of the next iteration. For the first The proportion of yarn components in each layer. For the first The residual vector between the predicted fabric performance index vector and the target fabric performance index vector in the next iteration.

[0233] Provided that the validation conditions of the positive surrogate model are met, the acceptance condition for candidate updates is either a decrease in the candidate residual vector or a decrease in the robust objective function:

[0234]

[0235] Or it satisfies:

[0236]

[0237] in, , To preset the descent threshold, This is the upper limit of the residual norm. , The objective function is robust.

[0238] The validation conditions for the positive surrogate model are expressed as follows:

[0239]

[0240] in, To verify the deviation threshold of the positive surrogate model, this embodiment... Using the 95th quantile of the predicted residual norm for leave-one-out cross-validation samples:

[0241]

[0242] in, It is the 95th percentile operator for the sample set.

[0243] If a candidate update fails the descent test, i.e. does not meet the acceptance criteria for a candidate update, then a backtracking linear search is triggered, let:

[0244] ;

[0245] And reconstruct the candidate logarithmic space variables, candidate yarn component ratios, candidate fabric performance index prediction vectors, and candidate residuals; if If the error fails to pass even after falling below the preset lower limit or reaching the maximum number of backtracking attempts, then:

[0246] ;

[0247] And based on the updated Resolve .in, The relaxation factor is the backtracking decay coefficient. This is the global damping factor amplification coefficient. If the damping adjustment attempts fail after reaching the upper limit, the automatic solution will terminate and a mark indicating that manual verification is required will be output.

[0248] After convergence, the solution is based on the Jacobian matrix at the final time step. and iterative reweighting matrix Calculate the posterior covariance matrix The expression is:

[0249]

[0250] The optimal formula is output as follows:

[0251]

[0252] The confidence interval for the optimal formulation is expressed as:

[0253] .

[0254] The criteria for determining that there is no physically feasible solution can be summarized as follows:

[0255]

[0256] in, The maximum number of conditions is preset. The condition number is the optimal formulation. This is the prediction vector for the fabric performance indicators of the optimal formulation.

[0257] This invention provides tool support directly for practical operational aspects of fabric research and development. In the formulation evaluation stage, after determining the yarn formulation (i.e., the yarn component ratios and process parameters), researchers can obtain predicted values ​​and confidence intervals for multiple finished product performance indicators (i.e., fabric performance indicators) within seconds using the forward proxy model comprised of steps 1 to 3. This allows for assessment of the likelihood of the yarn formulation meeting target requirements before actual sampling, eliminating obviously infeasible formulations on the computer and reducing unnecessary sampling. The forward proxy model of this invention covers the entire parameter input chain from fiber selection to finishing processes, making it suitable for formulation evaluation scenarios under existing production conditions. In the formulation design stage, when customers specify the target fabric's performance indicators, researchers can use the reverse engineering operation of this invention—that is, inputting the target fabric's performance indicators—to obtain recommended yarn component ratios and confidence intervals. R&D personnel can use this information to determine the reliability of recommended formulations. For recommended yarn component ratio schemes with high confidence, sampling verification is arranged after combining given fixed process parameters and manual review. Schemes with low confidence are marked as requiring manual review, avoiding blind investment in sampling resources. Regarding data utilization, the modeling method of this invention is designed for the realistic conditions of small historical R&D samples in enterprises. It eliminates spurious correlations through causal coding of the process chain, constructs cross-features through prior physical correlations between indicators, and suppresses overfitting through multi-base model ensemble networks. This allows the positive surrogate model to achieve usable predictive accuracy even with small samples, without requiring enterprises to invest heavily in accumulating data through experiments in the early stages. The above positive and negative closed loops can be packaged into a fabric R&D auxiliary decision-making tool for direct use by enterprise R&D personnel in their daily work.

[0258] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0259] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting fabric performance and inversely estimating yarn ratio, characterized in that, include: Based on the multi-stage temporal characteristics of fabric production process, process chain causal gating encoding is performed to obtain process chain representation vectors used to characterize the unidirectional causal dependency relationship of each process stage. Based on the process chain representation vector and the pre-set prior table of positive and negative correlations between fabric performance indicators, the shared driving features of positive correlation groups and the antagonistic driving features of negative correlation groups are extracted. The shared driving features of positive correlation groups and the antagonistic driving features of negative correlation groups are then concatenated with the process chain representation vector to obtain the enhanced feature vector. Based on the integrated network of enhanced feature vector and TabM multibase model, multiple fabric performance indicators of the target fabric are jointly predicted to obtain the fabric performance indicator prediction vector, and the prediction loss and correlation consistency constraint loss are optimized simultaneously. Based on the target fabric performance index vector, fixed process parameter conditions, smooth surrogate model and KAMoE meta-learning, the yarn component ratio is solved in reverse, and the uncertainty and confidence of the solution result are quantitatively evaluated by the posterior covariance matrix.

2. The fabric performance prediction and yarn ratio inverse estimation method according to claim 1, characterized in that, Based on the multi-stage temporal characteristics of fabric production processes, process chain causal gating coding is performed, specifically including: The fabric production process is divided into multiple process stages. The original features of each process stage are embedded and encoded, and the positional encoding is superimposed on the initial embedding vector of each process stage. The initial embedding vectors of each process stage, which have been superimposed with position coding, are subjected to unidirectional causal gating modulation to obtain the causal-modulated embedding vectors of each process stage, as expressed in the following expression: ； in: For the first Embedding vectors for each process stage For the first Initial embedding vectors for each process stage For the first Location coding for each process stage For learnable scalar scaling factors, For gated modulation function, For the preceding process stage The initial embedding vector; The causal-modulated embedding vectors of each process stage are concatenated in the order of the stages to form the process chain representation vector.

3. The fabric performance prediction and yarn ratio inverse estimation method according to claim 2, characterized in that, The expression for the gated modulation function is: ； in: It is an element-wise sigmoid activation function. , , , All are learnable parameters. This is element-wise multiplication.

4. The fabric performance prediction and yarn ratio inverse estimation method according to claim 1, characterized in that, The extraction process for shared driving features of positively correlated groups and adversarial driving features of negatively correlated pairs includes: Based on a prior table of positive and negative correlations between process chain representation vectors and fabric performance indicators, shared driving features of positively correlated groups and antagonistic driving features of negatively correlated pairs are extracted, expressed as follows: ； ； in: For the first Positively correlated groups share driving characteristics. For the first A negatively correlated pair of adversarial driving features , , , For learnable parameters, For cross-feature dimensions, Let be a real vector space. This is a process chain characterization vector. It is the ReLU activation function. The tanh activation function; All positively correlated groups share the same driving feature, and the negatively correlated groups share the opposing driving feature, which are then concatenated sequentially to form a cross-feature vector: ； in: For cross feature vectors, , For the first, Positively correlated groups share driving characteristics. , For the first, A negatively correlated pair of adversarial driving features Let be a real vector space.

5. The fabric performance prediction and yarn ratio inverse estimation method according to claim 1, characterized in that, Based on the ensemble network of enhanced feature vectors and TabM multi-base model, multiple fabric performance indicators of the target fabric are jointly predicted, including: Construct a TabM multi-base model ensemble network, where the number of base models is set to be [value missing]. The number of hidden layers is , No. Hidden layer for the first The calculation of each basic model is as follows: ； in: , The first , Hidden layer for the first The calculation results of the basic model This is the weight matrix. It is the bias vector; Each base model is processed by the first After one hidden layer, a pair is generated through a linear output layer. Prediction vectors for each fabric performance index: ； in: The weight matrix of the linear output layer. For the first The output bias vectors of each base model For the first The base model of the first Hidden layer output, For the first Each base model Prediction vectors for each fabric performance index; Pick Each base model The arithmetic mean of the prediction vectors for each fabric performance index is the final prediction vector for the fabric performance index. And predict vectors for fabric performance indicators. Apply differentiable constraint transformations.

6. The fabric performance prediction and yarn ratio inverse estimation method according to claim 1, characterized in that, The total loss function expression for simultaneously optimizing the prediction loss and the correlation consistency constraint loss of extracted cross features is: ； in: For the total loss function, A set of learnable parameters. For learnable parameters, This refers to the number of historical R&D samples used for model training and validation. For the first The loss weight of each fabric performance index, , These are the weighting coefficients for the relevance constraints. The regularization coefficient is . For the total number of negatively correlated pairs, For the first Negative correlation consistency loss for a negatively correlated pair For the first Positive correlation consistency loss for each positively correlated group The total number of positively correlated groups. , The first The first historical R&D sample Predicted and actual values ​​of various fabric performance indicators This represents the total number of fabric performance indicators.

7. The fabric performance prediction and yarn ratio inverse estimation method according to claim 1, characterized in that, Based on the target fabric performance index vector, fixed process parameter conditions, smooth surrogate model, and KAMoE meta-learning, the yarn component ratio is solved inversely, specifically including: A smooth proxy model is constructed using the Spectral-KAN network with the Mish-Sine architecture, and the smooth proxy model is trained. The condition number of the Jacobian matrix of historical R&D samples is calculated using a smooth surrogate model. Based on the condition number, a cluster space vector is constructed, and the K-Medoids algorithm is used to perform clustering in the cluster space to extract multiple real historical R&D samples as difficulty prototypes. Based on the difficulty prototype and target fabric performance indicators, a manifold regularized reweighted retrieval is performed to generate an initial coarse solution, global difficulty location features, and local manifold statistical features. A hybrid expert meta-learner is constructed, taking an initial coarse solution, global difficulty localization features, and local manifold statistical features as inputs. The gating network of the hybrid expert meta-learner calculates the activation weights of each expert based on the global difficulty localization features, and performs a weighted summation of the outputs of all experts to obtain the original policy vector. ; The original policy vector is activated by the activation function. Activation processing is performed to obtain the set of hyperparameters for the solution. ,in, The global damping factor. For preconditional vectors, For robust loss parameters, It is a relaxation factor; Based on the Jacobian matrix and the set of hyperparameters, a robust solution for the yarn composition ratio of the target fabric is obtained through a dual correction mechanism. The reliability of the solution is verified to determine the final yarn composition ratio of the target fabric.

8. The fabric performance prediction and yarn ratio inverse estimation method according to claim 7, characterized in that, Based on the Jacobian matrix and the set of hyperparameters, a robust solution for the yarn composition ratio of the target fabric is achieved through a dual correction mechanism, including: Construct a robust objective function, expressed as: ； in, For the first The update amount of the logarithmic space variable to be obtained in the next iteration. For Jacobian matrices, For the first The residual vector between the predicted fabric performance index vector and the target fabric performance index vector in the next iteration. Huber robust loss function For the first The robust objective function for the next iteration, superscript For transpose, The total number of fabric performance indicators, indicated by the superscript. For transpose, This is the precondition matrix; Calculate the double correction matrix, which consists of the iterative reweighting matrix and the preconditioning matrix. ; Robust iterative updates within the constraint domain are performed based on the dual correction matrix to obtain the updated logarithmic space yarn composition ratio. The updated logarithmic space yarn composition ratio is then mapped back to the yarn composition ratio using a softmax transformation.

9. The fabric performance prediction and yarn ratio inverse estimation method according to claim 8, characterized in that, The direction of robust iterative updates is: ； in, For the first The proportion of yarn components in the logarithmic space of the next iteration. This is the iterative reweighting matrix.

10. A fabric performance prediction and yarn ratio inverse estimation system, characterized in that, include: The encoding module is used to perform process chain causal gating encoding on the multi-stage temporal characteristics of fabric production process, and obtains process chain representation vectors to represent the unidirectional causal dependencies of each process stage. The feature extraction module is used to extract the shared driving features of positively correlated groups and the adversarial driving features of negatively correlated groups, and concatenate the shared driving features of positively correlated groups and the adversarial driving features of negatively correlated groups with the process chain characterization vector to obtain the enhanced feature vector; The joint prediction module is used to jointly predict multiple fabric performance indicators of the target fabric based on the enhanced feature vector and the TabM multibase model integrated network, to obtain the fabric performance indicator prediction vector, and simultaneously optimize the prediction loss and the correlation consistency constraint loss. The inverse solution module is used to inversely solve the yarn component ratio based on the target fabric performance index vector, fixed process parameters, and KAMoE meta-learning, and to quantitatively evaluate the uncertainty and confidence of the solution results through the posterior covariance matrix.

Images