A method and system for predicting the quality of spun yarn and optimizing the production formula in a closed loop.

By generating and optimizing the production prescription for colored yarn, and utilizing predictive models and interactive coding techniques, the problems of data dispersion and experience dependence in colored yarn production were solved. This enabled closed-loop optimization and feedback correction of quality prediction and production prescription, thereby improving the stability and efficiency of production.

CN123117932APending Publication Date: 2026-11-17WUHAN TEXTILE UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611608057.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the current colored yarn production process, the raw material ratio and process parameter setting rely on experience, and the data is stored in a scattered manner, making it difficult to achieve closed-loop optimization of quality prediction and production prescription. Moreover, the existing methods fail to accurately characterize the contribution relationship of raw materials and the impact of process conditions.

Method used

By acquiring target quality indicators and a set of available raw materials, candidate production prescriptions are generated. Then, predictive models are used for interactive coding and weighted aggregation to generate target-specific effective contribution weights. Based on differences and deviations, proportional transfer actions are generated, and counterfactual verification and optimization are performed to form an interpretable and executable closed-loop optimization mechanism.

Benefits of technology

It achieves multi-source data fusion based on batch number, improves the availability of quality prediction and the feasibility of prescription optimization, reduces quality gaps and meets production constraints, and forms a continuous feedback correction mechanism.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117932A_ABST
    Figure CN123117932A_ABST
Patent Text Reader

Abstract

This invention relates to a method and system for predicting the quality of colored spun yarn and for closed-loop optimization of production prescriptions. The method includes: acquiring target quality indicators and a set of available raw materials; generating candidate production prescriptions containing initial blending ratios, front-end and downstream process conditions; and inputting these into a prediction model. In the model, raw material characteristics are modulated using front-end process conditions and interactively encoded. After weighted aggregation by target-specific effective contribution weights, the predicted quality indicators are modulated by downstream process conditions. Then, based on quality gaps and contribution deviations, a proportion transfer action is constructed, and recommended prescriptions are iteratively screened through feasible region projection and counterfactual prediction verification. After production, residual correction amounts are generated using prescription similarity, historical sample credibility, and hierarchical backtracking to correct the prediction results of subsequent similar prescriptions. Through these steps, accurate prediction of colored spun yarn quality and closed-loop iterative optimization of production prescriptions are achieved, continuously improving the accuracy of prescription recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of textile technology and production technology, specifically relating to a method and system for predicting the quality of colored spun yarn and optimizing the production formula in a closed loop. Background Technology

[0002] Colored yarn production typically involves the blending of multiple raw materials, the design of proportions for different batches of raw materials, and the coordinated control of multi-stage spinning process parameters. Yarn quality is influenced not only by the fiber properties of the raw materials but also by factors such as the blending ratio, carding or combing process, waste rate, spinning method, draft ratio, twist, yarn count, and spinning process. For colored yarn companies, quality indicators such as strength, evenness (CV), hairiness, and elongation directly affect subsequent weaving efficiency, fabric appearance, and product stability. Therefore, quality prediction and optimization of the production formula before production is crucial for reducing trial spinning costs, shortening process adjustment cycles, and improving order delivery stability.

[0003] In current colored yarn production processes, raw material ratios and process parameter settings still rely heavily on the experience of process engineers, historical similar orders, and small-batch trial spinning results. Although some companies have established Enterprise Resource Planning (ERP) systems, warehouse management systems, laboratory information management systems, and manufacturing execution systems, order, inventory, raw material testing, process execution, and yarn quality testing data are often scattered across different systems. If this data cannot be uniformly linked according to yarn batches or production work orders, it is difficult to form a data foundation that can be used for automatic production prescription generation, quality prediction, and closed-loop optimization.

[0004] In existing data-driven yarn quality prediction methods, a common approach is to convert raw material properties, blending ratios, and process parameters into fixed-length features, and then use regression models, machine learning models, or deep learning models to predict yarn quality. While these methods can characterize the nonlinear relationship between input variables and quality indicators to some extent, they still have the following shortcomings in the context of colored yarn production: First, the types and quantities of raw materials in a production batch may not be fixed, and the raw material set has disordered and variable-length characteristics; second, although the blending ratio contains prior production information, the nominal blending ratio is not equivalent to the actual contribution of each raw material to the final quality; third, the influence of different process stages on quality formation is staged and sequential, and simply inputting all process parameters as global auxiliary variables into the model is difficult to accurately represent the modulating effect of process conditions on the raw material combination relationship.

[0005] While existing data-driven yarn quality prediction methods can output predicted values ​​for quality indicators such as strength, evenness (CV), hairiness, and elongation, most methods only use the prediction results as offline evaluation results and do not further transform the raw material contribution relationships obtained within the model into a basis for adjusting the blending ratio. For colored yarn production, the nominal blending ratio only represents the proportion of each raw material in the formulation. However, the actual contribution of different raw materials to the final quality may vary under different raw material combinations, different upstream process conditions, and different downstream process parameters. Therefore, if the nominal blending ratio is used only as a fixed weight, it is difficult to provide an interpretable technical basis for subsequent formulation adjustments.

[0006] Meanwhile, although enterprise production systems can record order, inventory, raw material inspection, process execution, and yarn inspection data, this data is typically scattered across enterprise resource planning (ERP) systems, warehouse management systems, laboratory information management systems, and manufacturing execution systems. Existing systems are primarily used for data recording, work order flow, and result querying, lacking a closed-loop optimization mechanism that uses yarn batches or production work orders as indexes to connect candidate production prescription generation, quality prediction, raw material effective contribution analysis, prescription ranking, production execution issuance, and yarn inspection feedback correction.

[0007] Furthermore, existing production feedback methods typically involve process engineers adjusting subsequent prescriptions based on experience after yarn inspection. This approach fails to group and statistically analyze predicted residuals according to production dimensions such as raw material batch, raw material combination type, yarn count, color number, process route, and machine. It also fails to use the corrected residuals for subsequent predictions or scoring adjustments for similar candidate prescriptions. Therefore, existing technologies struggle to achieve an interpretable, executable, and sustainable optimization closed loop from quality prediction to production prescription generation and then to production feedback correction. Summary of the Invention

[0008] To address the problems raised in the background art, a method for predicting the quality of spun yarn and optimizing the production formula is provided in a first aspect of the present invention, comprising: Obtain the target quality indicators and the set of available raw materials, and generate at least one candidate production formula that includes the initial blending ratio, front-end process conditions and downstream process conditions; The candidate production prescription is input into a pre-trained prediction model; in the prediction model, the feature representation of each raw material in the available raw material set is modulated using the front-end process conditions and then interactive coding is performed to generate target-specific effective contribution weights. The interactively encoded features are weighted and aggregated based on the target-specific effective contribution weights, and the aggregated features are modulated using the downstream process conditions to output a prediction quality index. Based on the difference between the predicted quality index and the target quality index, and the deviation between the target specific effective contribution weight and the corresponding initial blending ratio, a candidate ratio transfer action is generated for the available raw material set; the counterfactual candidate production prescription adjusted by the candidate ratio transfer action is re-input into the prediction model for counterfactual verification. The validated candidate production prescriptions will be output as recommended production prescriptions for production execution.

[0009] In some embodiments of the present invention, generating target-specific effective contribution weights includes: generating corresponding contribution correction scores for different target quality indicators; and combining the contribution correction scores with the corresponding initial blending ratios so that the generated target-specific effective contribution weights characterize the relative contribution estimates of each raw material in the available raw material set for a specific quality indicator.

[0010] In some embodiments of the present invention, generating candidate proportion transfer actions for the available raw material set includes: determining candidate proportion receiving raw materials and candidate proportion releasing raw materials in the available raw material set according to the differences and deviations; establishing proportion transfer rules between the candidate proportion releasing raw materials and the candidate proportion receiving raw materials to generate an initial proportion transfer action; projecting the temporary blending proportion formed by the initial proportion transfer action onto a preset proportion feasible region to eliminate actions that do not meet production constraints, thereby obtaining the candidate proportion transfer action.

[0011] In some embodiments of the present invention, the output of recommended production prescriptions based on verified candidate production prescriptions includes: The counterfactual quality gain is calculated based on the change in the difference before and after verification; in response to the counterfactual quality gain meeting the preset conditions and the constraint indicators in the target quality indicators not exceeding the risk boundary, the corresponding candidate proportion transfer action is retained, and the next round of candidate production prescriptions is generated based on the retained action for iterative verification until the iteration stop condition is met; based on the degree of improvement of quality gap and the auxiliary score of production constraints, the candidate production prescriptions after the iteration stop are sorted, and the recommended production prescription is output.

[0012] Furthermore, the process of ranking candidate production prescriptions after iteration stops based on the degree of improvement in quality gaps and production constraint auxiliary scores, and outputting the recommended production prescriptions, includes: calculating the quality deviation of each quality indicator based on the numerical deviation between the predicted quality indicator and the allowable range corresponding to the target quality indicator; weighting and summing the changes in the quality deviation before and after iteration adjustment to obtain a quality benefit score representing the degree of improvement; calculating the production constraint auxiliary score of the candidate production prescriptions based on business evaluation dimensions and preset penalty dimensions, wherein the business evaluation dimensions include at least one of cost, inventory, and delivery date matching degree, and the preset penalty dimensions include at least one of quality boundary risk and prescription adjustment range risk; determining the basic judgment priority based on whether the candidate production prescriptions meet all target quality indicators, and combining the calculation results of the quality benefit score and the production constraint auxiliary score, selecting and outputting the recommended production prescriptions from the candidate production prescriptions.

[0013] In some embodiments of the present invention, the method further includes correction based on production execution feedback: obtaining the actual yarn quality inspection results after the recommended production prescription is executed, and calculating the prediction residual between the actual yarn quality inspection results and the predicted quality index; establishing historical residual grouping at multiple levels, and calculating the prescription similarity between the next round of candidate production prescriptions and the historical residual grouping; in response to the prescription similarity reaching a preset threshold, generating a residual correction amount based on the residual statistics and sample confidence factor of the corresponding historical residual grouping to correct the predicted quality index of the next round of candidate production prescriptions.

[0014] Furthermore, the generation of residual correction includes: preferentially selecting fine-grained historical residual groups that meet the similarity threshold and whose sample size meets the preset conditions to calculate the residual correction; in response to the fine-grained historical residual groups not meeting the conditions, sequentially performing backtracking matching to higher-level historical residual groups with fewer dimensions.

[0015] A second aspect of the present invention provides a system for predicting the quality of spun yarn and optimizing the production formula in a closed loop, comprising: The acquisition module is used to acquire the target quality indicators and the set of available raw materials, and generate at least one candidate production formula containing the initial blending ratio, front-end process conditions and downstream process conditions. The modulation module is used to input the candidate production prescription into a pre-trained prediction model; in the prediction model, the feature representation of each raw material in the available raw material set is modulated using the front-end process conditions and then interactive encoding is performed to generate target-specific effective contribution weights; the interactively encoded features are weighted and aggregated based on the target-specific effective contribution weights, and the aggregated features are modulated using the downstream process conditions to output a prediction quality index. The generation module is used to generate candidate ratio transfer actions for the set of available raw materials based on the difference between the predicted quality index and the target quality index, as well as the deviation between the target specific effective contribution weight and the corresponding initial blending ratio. The verification module is used to re-input the counterfactual candidate production prescriptions adjusted by the candidate ratio transfer action into the prediction model for counterfactual verification. The output module is used to output the validated candidate production prescriptions as recommended production prescriptions for production execution.

[0016] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the color-spun yarn quality prediction and production formula closed-loop optimization method provided in the first aspect of the present invention.

[0017] In a fourth aspect, the present invention provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for predicting the quality of spun yarn and optimizing the production formula closed loop provided in the first aspect of the present invention.

[0018] The beneficial effects of this invention are: This invention uses the batch number or production work order number of colored yarn as an index to integrate order, inventory, raw material testing, process execution and yarn quality testing data to form a traceable batch-level production sample, thereby improving the availability of multi-source production data in quality prediction and prescription optimization.

[0019] This invention sets the front-end process conditions before the raw material set interactive encoding and sets the downstream process conditions after the target-specific weighted aggregation, which can express the different modulation effects of process conditions on the raw material state and potential yarn state according to the quality formation stage.

[0020] This invention generates target-specific effective contribution weights for different quality indicators. While retaining the prior nominal blending ratio, it characterizes the relative contribution estimate of each candidate raw material to a specific quality indicator, providing a calculation basis for candidate formulation interpretation and ratio transfer action generation.

[0021] This invention does not directly use the target-specific contribution deviation as the physical proportion adjustment amount. Instead, it generates candidate actions through a proportion transfer diagram and uses proportion feasible region projection and counterfactual quality prediction to verify and screen candidate proportion transfer actions that can reduce the quality gap and meet production constraints, thereby improving the feasibility and stability of prescription adjustment.

[0022] This invention uses quality gap improvement and counterfactual quality benefits as the main evaluation criteria for candidate formulations, and uses cost, inventory, delivery time, process stability and formulation boundary risk as auxiliary ranking factors, so that quality objectives and enterprise production constraints can jointly participate in formulation selection.

[0023] This invention generates residual correction amounts based on the similarity between subsequent candidate production prescriptions and historical residual groupings, the reliability corresponding to the number of historical samples, and multi-level residual groupings. When fine-grained samples are insufficient, hierarchical backtracking is performed to reduce the risk of overcorrection in small sample groupings and erroneous migration of dissimilar production scenarios, thereby forming a closed loop of continuous correction of subsequent prescription prediction and ranking based on production feedback. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the basic process of the method for predicting the quality of colored spun yarn and optimizing the production formula in some embodiments of the present invention. Figure 2 This is a schematic diagram illustrating the specific process of color-spun yarn quality prediction and production formula closed-loop optimization in some embodiments of the present invention. Figure 3 A schematic diagram of the color perception fusion network and the target-specific effective contribution weight generation structure in some embodiments of the present invention; Figure 4 This is a schematic diagram of the production execution, yarn quality data feedback, similarity gating, and reliable hierarchical residual correction process in some embodiments of the present invention; Figure 5 This is a schematic diagram of the basic structure of the colored yarn quality prediction and production formula closed-loop optimization system in some embodiments of the present invention. Figure 6 This is a schematic diagram of the specific structure of the colored yarn quality prediction and production formula closed-loop optimization system in some embodiments of the present invention. Detailed Implementation

[0025] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0026] refer to Figure 1 and Figure 2 The present invention provides a method for predicting the quality of spun yarn and optimizing the production formula in a closed loop, comprising: S100. Obtain the target quality indicators and the set of available raw materials, and generate at least one candidate production formula that includes the initial blending ratio, front-end process conditions and downstream process conditions; S200. Input the candidate production prescription into a pre-trained prediction model; in the prediction model, after modulating the feature representation of each raw material in the available raw material set using the front-end process conditions, perform interactive coding to generate target-specific effective contribution weights; based on the target-specific effective contribution weights, perform weighted aggregation of the interactively encoded features, and use the downstream process conditions to modulate the aggregated features to output a predicted quality index. S300. Based on the difference between the predicted quality index and the target quality index, and the deviation between the target specific effective contribution weight and the corresponding initial blending ratio, generate a candidate ratio transfer action for the available raw material set; S400. The counterfactual candidate production prescription adjusted by the aforementioned candidate ratio transfer action is re-inputted into the prediction model for counterfactual verification; S500. Output the validated candidate production prescription as a recommended production prescription for production execution.

[0027] In step S100 of some embodiments of the present invention, the target quality index and the set of available raw materials are obtained, and at least one candidate production formula including the initial blending ratio, front-end process conditions and downstream process conditions is generated. Specifically, multi-source production data is acquired during the colored yarn production process. This multi-source production data includes order quality targets, raw material inventory information, raw material testing information, historical process parameters, historical production formulas, and yarn quality testing results.

[0028] Using the batch number or production order number of the spun yarn as an associated index, the system aligns the raw material batch number, raw material testing indicators, blending ratio, process parameters, and yarn quality testing results to form batch-level production samples. For each historical production batch, the system establishes a mapping relationship between the raw material set, blending ratio, process parameters, and yarn quality testing results. Through this step, scattered data can be transformed into batch-level production samples that can be used for model training, prescription recall, quality prediction, and feedback correction.

[0029] Based on the order information, the quality objectives and production constraints of the spun yarn are analyzed. The quality objectives include one or more of the following: strength, evenness CV, hairiness, and elongation. The production constraints include one or more of the following: yarn count, cost constraints, delivery time constraints, and process parameter boundaries.

[0030] A pool of available raw materials is constructed based on order quality targets, raw material inventory information, and raw material testing information. Each candidate raw material in the pool includes a raw material batch number, raw material testing attributes, inventory quantity, cost information, and historical quality performance information.

[0031] A set of candidate production prescriptions is generated based on the available raw material candidate pool, order quality targets, and process parameter boundaries. Each candidate production prescription includes a set of candidate raw materials, a corresponding nominal blending ratio, front-end process conditions, and downstream process conditions. The nominal blending ratio satisfies the total ratio constraint and the single raw material ratio boundary, and the front-end process conditions and downstream process conditions satisfy the allowable range of the corresponding equipment and process route. The candidate production prescription includes a set of candidate raw materials, a nominal blending ratio, front-end process conditions, and downstream process conditions; the process conditions include one or more of the following: carding or combing process, cotton waste rate, spinning method, yarn count, draft ratio, twist, ply count, and spinning process.

[0032] Optionally, the generation of candidate production prescriptions may include one or more of the following methods: First, hard constraint filtering is performed based on inventory constraints, cost constraints, and process parameter boundaries to exclude prescriptions with insufficient inventory, excessive costs, or unfeasible equipment.

[0033] Second, recall similar historical production prescriptions that match the current order's quality targets, yarn count, color number, or process route.

[0034] Third, based on similar historical formulations, the blending ratio and process parameters are perturbed to generate new candidate production formulations. The perturbed blending ratio still satisfies the ratio comprehensive constraint, and the perturbed process parameters still satisfy the equipment process boundary constraint.

[0035] In this step, raw material batches can be screened based on order constraints, inventory information, and raw material testing information. Screening criteria include: raw material inventory meeting order requirements, complete raw material testing indicators, raw material attributes within usable range, raw material cost not exceeding the order budget, raw material color or grade meeting the requirements of the yarn-dyed yarn order, and the raw material batch not being marked as abnormal or disabled. This screening process yields a pool of available raw material candidates.

[0036] refer to Figure 3 In step S200 of some embodiments of the present invention, the candidate production prescription is input into a pre-trained prediction model; in the prediction model, the feature representation of each raw material in the available raw material set is modulated using the front-end process conditions and then interactive coding is performed to generate target-specific effective contribution weights; the interactively encoded features are weighted and aggregated based on the target-specific effective contribution weights, and the aggregated features are modulated using the downstream process conditions to output a predicted quality index; The specific effective contribution weights for generating the target include: S201. Generate corresponding contribution correction scores for each of the different target quality indicators; Specifically, the candidate production prescriptions are input into a process-aware fusion network. Before the raw material set is interactively encoded, the process-aware fusion network performs feature-level modulation on the raw material embedding representation of each candidate raw material according to the front-end process conditions, and performs interactive encoding on the modulated unordered, variable-length candidate raw material set to obtain the interactively enhanced representation of each candidate raw material.

[0037] For different yarn quality indicators, contribution correction scores are generated according to the interaction enhancement representation, and the contribution correction scores are combined with the nominal blending ratio of the corresponding candidate raw materials to obtain the target-specific effective contribution weight, forming a target-specific effective contribution weight matrix.

[0038] S202. The contribution correction score is combined with the corresponding initial blending ratio so that the generated target-specific effective contribution weight represents the relative contribution estimate of each raw material in the available raw material set for a specific quality index.

[0039] Specifically, the interactive enhancement representation is weighted and aggregated according to the target-specific effective contribution weights corresponding to each quality indicator, and then modulated at the feature level according to the downstream process conditions after aggregation to output the predicted quality indicators corresponding to the candidate production prescription. The target-specific effective contribution weights are different from the actual physical feed ratios and are used to characterize the relative contribution estimates of each candidate raw material to a specific quality indicator.

[0040] Optionally, the prediction model (process-aware fusion network) includes a raw material embedding unit, a front-end process condition modulation unit, a raw material set interactive coding unit, a target-specific effective contribution weight generation unit, a target-specific weighted aggregation unit, a downstream process condition injection unit, and a quality prediction unit.

[0041] It is understood that, based on quality indicators with quality gaps, quality indicator priorities, target-specific contribution deviations, and raw material inventory, costs, and proportion adjustment space, candidate proportion receiving raw materials, candidate proportion releasing raw materials, and adjustment priorities are determined. The target-specific contribution deviation is used to generate candidate proportion transfer actions and is not directly used as the final physical blending proportion adjustment result. A proportion transfer diagram is constructed with candidate raw materials as nodes, generating at least one candidate proportion transfer action between candidate proportion releasing raw materials and candidate proportion receiving raw materials. Each candidate proportion transfer action is used to transfer a preset proportion from candidate proportion releasing raw materials to candidate proportion receiving raw materials.

[0042] In one specific embodiment, the process-aware fusion network includes a raw material embedding unit, a front-end process condition modulation unit, a raw material set interactive encoding unit, a target-specific effective contribution weight generation unit, a target-specific weighted aggregation unit, a downstream process condition injection unit, and a quality prediction unit.

[0043] For a candidate production prescription, its input can be represented as: , in, Represents the set of candidate raw materials. This indicates the nominal blending ratio corresponding to each candidate raw material. Indicates the front-end process conditions. This indicates the downstream process conditions.

[0044] Suppose the candidate raw material set includes The first candidate raw material, The detection attribute vector of the candidate raw materials is The nominal blending ratio is ,but: , And satisfy: , The number of candidate raw materials contained in different candidate production formulations may vary, and the order in which the candidate raw materials are arranged in the set is not considered as valid input information.

[0045] The raw material embedding unit maps the detection attributes of each candidate raw material to a unified latent feature space, obtaining the initial raw material embedding representation: , in, This represents the raw material embedding function. The raw material detection attributes include one or more of the following: fiber length, strength, fineness, maturity, short fiber content, impurity content, color index, and other detection attributes that can characterize the quality status of the raw material.

[0046] The front-end process condition modulation unit is located before the raw material set interactive coding unit. The front-end process conditions include one or more of the following: carding process, combing process, carding process, cotton waste rate, and blending method.

[0047] The front-end process condition modulation unit modulates the front-end process conditions. Generate the first scaling parameter and the first offset parameter Furthermore, feature-level modulation is performed on the initial raw material embedding representation of each candidate raw material: , Here, ⊙ represents element-wise multiplication. This indicates the raw material after the front-end process conditions have been adjusted.

[0048] By injecting front-end process conditions before set interaction coding, the coding of subsequent raw material combination relationships can reflect the impact of the current front-end processing conditions on the state of different candidate raw materials.

[0049] The raw material set interactive coding unit performs interactive coding on the modulated unordered, variable-length candidate raw material set to obtain the interactive enhanced representation of each candidate raw material: , in, This represents the interactive coding function for the raw material set. The interactive coding function employs a coding structure capable of handling unordered, variable-length sets, so that the enhanced interactive representation simultaneously includes the properties of the candidate raw materials themselves, as well as the synergistic, complementary, or restrictive relationships between candidate raw materials.

[0050] Regarding the first The yarn quality index, the target-specific effective contribution weight generation unit is based on the first... The interaction enhancement representation of the candidate raw materials generates contribution correction scores: , in, Indicates the first The contribution correction function corresponding to each quality indicator. Indicates the first Candidate raw materials for the first The contribution correction score for each quality indicator.

[0051] The target-specific effective contribution weight generation unit uses the nominal blending ratio as the proportion prior and corrects this proportion prior using the contribution correction score. Candidate raw materials for the first The target-specific effective contribution weight of each quality indicator is expressed as follows: , in, Indicates the target-specific effective contribution weight. This represents a preset constant that is greater than zero.

[0052] For any quality index The target-specific effective contribution weights corresponding to each candidate raw material satisfy the following: , For each of the quality indicators including strength, yarn count (CV), hairiness, and elongation, target-specific effective contribution weights are generated, forming a target-specific effective contribution weight matrix: , in, This indicates the number of quality indicators to be predicted. Different columns in the target-specific effective contribution weight matrix correspond to the relative contribution distribution of each candidate raw material under different quality indicators.

[0053] The target-specific effective contribution weight differs from the actual physical feed ratio. It is used to represent the relative contribution estimate of candidate raw materials to a specific predicted quality index under the current raw material combination, nominal blending ratio, and process conditions.

[0054] The target-specific weighted aggregation unit performs weighted aggregation of the interaction enhancement representations of candidate raw materials based on the corresponding target-specific effective contribution weights for different quality indicators. The target-specific potential yarn state corresponding to each quality indicator is expressed as follows: , Because different quality indicators use different target-specific effective contribution weights, the same candidate production formula can form different target-specific potential yarn states for different quality indicators.

[0055] The downstream process condition injection unit is located after the target-specific weighted polymerization unit. The downstream process conditions include one or more of the following: spinning method, spinning process, yarn count, roving draft ratio, spinning draft ratio, single yarn twist, ply twist, and ply number.

[0056] The downstream process condition injection unit injects according to the downstream process conditions. Generate second scaling parameters Second offset parameter and for the first Modulate the target-specific potential yarn state corresponding to each quality indicator: , in, This indicates the state of the target specific yarn after the injection of downstream process conditions.

[0057] The front-end process conditions and the downstream process conditions are injected separately according to the quality formation stage: the front-end process conditions modulate the raw material representation before the raw material set interactive coding, and the downstream process conditions modulate the potential yarn state after the target-specific weighted polymerization.

[0058] The quality prediction unit sets up corresponding prediction branches for different quality indicators. The predicted values ​​for the quality indicators are expressed as follows: , in, Indicates the first The prediction function corresponding to each quality indicator.

[0059] When the prediction targets include strength, strand CV, hairiness, and elongation, the predicted quality index vector output by the process-aware fusion network is represented as: , The process-aware fusion network is trained under supervision using batch-level production samples. These samples include historical candidate raw material sets, nominal blending ratios, upstream process conditions, downstream process conditions, and corresponding actual yarn quality inspection results. Network parameters are jointly updated based on the loss between predicted quality indicators and actual yarn quality inspection results.

[0060] The process-aware fusion network simultaneously outputs a predicted quality index vector and a target-specific effective contribution weight matrix. For the ... Candidate raw materials and the first The system calculates the target-specific contribution bias for the following quality indicators: , The target specific contribution deviation is used for candidate production formula interpretation, candidate ratio receiving raw material screening and candidate ratio releasing raw material screening, and ratio transfer map construction, but is not directly used as the increase or decrease in the actual blending ratio.

[0061] The system jointly analyzes the target-specific contribution deviation with quality gaps, raw material testing attributes, costs, inventory, and production constraints to generate candidate proportion transfer actions. These candidate proportion transfer actions must be verified by proportion feasible region projection and counterfactual quality prediction before they can be used to generate the next round of candidate production prescriptions.

[0062] Through the above structure, the process-aware fusion network can handle production prescriptions with variable quantities and arrangements of candidate raw materials, and generate target-specific effective contribution weights for different yarn quality indicators, thus establishing a calculable connection between the quality prediction results and subsequent proportional transfer and prescription iteration.

[0063] Therefore, in step S300 of some embodiments of the present invention, a candidate proportion transfer action is generated for the set of available raw materials based on the difference between the predicted quality index and the target quality index, and the deviation between the target specific effective contribution weight and the corresponding initial blending ratio. The step of generating candidate ratio transfers for the set of available raw materials includes: S301. Based on the difference and the deviation, determine the candidate ratio of receiving raw materials and the candidate ratio of releasing raw materials from the set of available raw materials; Specifically, for quality indicators with lower limit requirements, such as strength, the difference between the predicted value and the lower limit constitutes a quality gap when the predicted value is lower than the lower limit; for quality indicators with upper limit requirements, such as strand CV and feathering, the difference between the predicted value and the upper limit constitutes a quality gap when the predicted value is higher than the upper limit; for quality indicators with target ranges, the distance between the predicted value and the nearest target boundary constitutes a quality gap when the predicted value is outside the target range.

[0064] For the Candidate raw materials and the first The system calculates the target-specific contribution bias for the following quality indicators: , in, Indicates the first Candidate raw materials for the first The target-specific effective contribution weight of each quality indicator This indicates its nominal blending ratio.

[0065] The system determines the candidate ratio for receiving raw materials, the candidate ratio for releasing raw materials, and the adjustment priority based on the quality indicators with quality gaps, quality indicator priorities, target-specific contribution deviations, raw material testing attributes, historical quality performance, costs, inventory, and ratio adjustment space.

[0066] S302. Establish a proportional transfer rule between the candidate proportional release material and the candidate proportional receiving material, and generate an initial proportional transfer action; Specifically, the system constructs a proportional transfer graph using candidate raw materials as nodes, and generates candidate proportional transfer actions between candidate proportional release raw materials and candidate proportional receiving raw materials. For the proportional... By the Transfer of candidate raw materials to the first The action of selecting candidate raw materials is represented as: , After the candidate proportion transfer action is executed, the first The proportion of candidate raw materials increased , No. The proportion of candidate raw materials decreased The proportions of other candidate raw materials remain unchanged. The target-specific contribution bias is only used to generate and rank candidate proportion transfer actions and is not directly used as the final blending proportion adjustment result. The candidate proportion transfer actions also need to be verified by proportion feasible region projection and counterfactual prediction.

[0067] S303. Project the temporary blending ratio formed by the initial ratio transfer action onto a preset ratio feasible region to eliminate actions that do not meet production constraints, and obtain the candidate ratio transfer actions.

[0068] Specifically, a temporary blending ratio is generated based on the candidate ratio transfer action, and it is projected onto a ratio feasible domain consisting of one or more of the following: ratio comprehensive constraints, upper and lower limits of single raw material ratios, inventory constraints, raw material combination rules, order constraints, and raw material disabling rules.

[0069] The adjusted blending ratio that satisfies the proportional feasible region constraint is combined with the original candidate raw material set, front-end process conditions, and downstream process conditions to form a counterfactual candidate production prescription. This prescription is then re-input into the process perception fusion network to obtain counterfactual predicted quality indicators.

[0070] In step S400 of some embodiments of the present invention, the counterfactual candidate production prescription adjusted by the candidate ratio transfer action is re-input into the prediction model for counterfactual verification; Specifically, counterfactual quality gains are calculated based on the changes in quality gaps before and after adjustment. When the counterfactual quality gains meet preset conditions, other constrained quality indicators do not exceed risk boundaries, and counterfactual candidate production prescriptions meet production constraints, the corresponding candidate proportion transfer action is retained; otherwise, they are eliminated or adjusted and re-verified.

[0071] The next round of candidate production prescriptions is generated based on the candidate proportion transfer actions verified by counterfactual analysis, and a quality benefit score is calculated based on the reduction in quality gaps for each quality indicator. At the same time, an auxiliary comprehensive score is calculated based on one or more of the following factors: changes in raw material costs, inventory consumption, delivery date matching degree, process stability, prescription boundary risk, and cumulative proportion transfer magnitude.

[0072] Based on the quality benefit score and auxiliary comprehensive score, candidate production prescriptions for the next round of prediction are determined, and the target-specific effective contribution weight matrix is ​​regenerated. The quality gap calculation, contribution deviation analysis, proportion transfer action generation, feasible region projection, and counterfactual verification are then repeated.

[0073] The iteration stops when any of the following conditions are met: 1. Necessary quality indicators meet the order quality objectives; 2. There are no candidate proportion transfer actions that meet the preset profit conditions; 3. The quality improvement of candidate production prescriptions in two consecutive rounds is less than a preset threshold; 4. Reach the maximum number of iterations; 5. Received a manual stop command.

[0074] Optionally, if the counterfactual quality benefits meet the preset conditions, the quality gap of at least one quality indicator to be improved is reduced, other constrained quality indicators have not crossed the risk boundary, and the candidate production prescription meets the constraints of proportion, inventory, cost and process boundaries, the candidate proportion transfer action is retained; otherwise, the action is removed, or the proportion of receiving raw materials, releasing raw materials, or transferring the amount of raw materials is adjusted and then re-verified.

[0075] In one specific embodiment, candidate proportional transfer actions are generated based on the predicted quality indicators and target-specific effective contribution weight matrix output by the process-aware fusion network, combined with the order quality target. The retention of the candidate proportional transfer actions is then determined through proportional feasible region projection and counterfactual quality prediction verification.

[0076] The target-specific effective contribution weight and its deviation from the nominal blending ratio are only used to screen and rank candidate ratio transfer actions, and are not directly used as the increase or decrease in the actual physical blending ratio.

[0077] Let the current candidate production prescription be represented as: , in, Represents the set of candidate raw materials. Indicates the nominal blending ratio. and These represent the upstream process conditions and the downstream process conditions, respectively.

[0078] The process-aware fusion network outputs a predicted quality metric vector: , And the target-specific effective contribution weight matrix: , in, Indicates the quantity of candidate raw materials. Indicates the quantity of quality indicators.

[0079] For the For each quality indicator, the system calculates the normalized quality gap based on the shortest distance between the predicted value and the allowable range of that quality indicator. , in, Indicates the first The order allowable range corresponding to each quality indicator. This represents the shortest distance between the predicted value and the allowed range. Indicates the scale parameter.

[0080] When the predicted value meets the order quality requirements When the predicted value does not meet the order quality requirements, The system identifies quality indicators with a quality gap greater than zero as those that need improvement.

[0081] For the Candidate raw materials and the first For each quality indicator, the target-specific contribution bias is expressed as: , in, Indicates the first Candidate raw materials for the first The target-specific effective contribution weight of each quality indicator This indicates the corresponding nominal blending ratio.

[0082] The system combines the target-specific contribution deviation with the quality indicators to be improved, quality indicator priorities, raw material testing attributes, historical quality performance, ratio upper and lower limits, inventory levels, costs, and raw material combination rules to determine the candidate ratios for receiving raw materials and the candidate ratios for releasing raw materials.

[0083] Candidate proportion receiving raw materials refer to candidate raw materials that can be verified as objects for increasing the proportion under the conditions of meeting the upper limit of the proportion, inventory and raw material combination constraints; candidate proportion releasing raw materials refer to candidate raw materials that can be verified as objects for decreasing the proportion under the conditions of meeting the lower limit of the proportion and raw material combination constraints.

[0084] The system constructs a proportional transfer graph using candidate raw materials as nodes, and establishes directed proportional transfer edges between candidate proportional release raw materials and candidate proportional receiving raw materials. Candidate raw materials to the first Transfer ratio of candidate raw materials The candidate proportion shift action is represented as: ; After executing the candidate ratio transfer action, the temporary blending ratio satisfies: , , The proportions of other candidate raw materials remain unchanged.

[0085] The proportional transfer amount can be selected from a preset set of proportional step sizes, or it can be determined based on the quality gap, proportional adjustment space, and target-specific contribution deviation. The preset proportional step sizes can be 1%, 2%, 3%, or other proportional values ​​that meet the company's batching accuracy requirements.

[0086] The temporary blending ratio formed by the candidate ratio transfer action The system projects it onto the proportionally feasible region: , in, Represents the proportional feasible region. This indicates the candidate blending ratio after projection.

[0087] The feasible region of the proportion is constituted by one or more of the following constraints: 1. The sum of the blending ratios of all candidate raw materials is 1; 2. The blending ratio of each candidate raw material is within the allowable upper and lower limits; 3. The demand for candidate raw materials shall not exceed the corresponding available inventory. 4. The candidate raw material combination meets the requirements for raw material prohibition, mutual exclusion, color, grade, and minimum batch size; 5. The candidate production prescription meets order costs and other production constraints.

[0088] When the temporary blending ratio satisfies all constraints, the projected candidate blending ratio can be the same as the temporary blending ratio; when there is no feasible projection result that satisfies the constraints, the system removes the corresponding candidate ratio transfer action.

[0089] The system combines the projected candidate blending ratios with the original candidate raw material set, upstream process conditions, and downstream process conditions to form counterfactual candidate production prescriptions: , The counterfactual candidate production prescriptions are re-input into the process-aware fusion network to obtain the counterfactual prediction quality index vector: , in, This represents a process-aware fusion network.

[0090] The system recalculates the quality gaps for each quality indicator based on counterfactual predictions. And calculate the counterfactual quality benefits corresponding to the candidate proportion shift action: , in, Indicates the first Priority weights for quality indicators.

[0091] when When the weighted quality gap of the counterfactual candidate production prescription is less than the weighted quality gap of the candidate production prescription before adjustment, it indicates that the counterfactual candidate production prescription has a smaller weighted quality gap than the unadjusted candidate production prescription.

[0092] When a candidate proportion transfer action satisfies the following conditions: the counterfactual quality benefit reaches the preset condition, the quality gap of at least one quality indicator to be improved decreases, other constrained quality indicators do not exceed the quality risk boundary, and the adjusted candidate production prescription meets the constraints of proportion, inventory, cost, and process, the system retains the candidate proportion transfer action.

[0093] When a candidate proportion transfer action does not meet the above conditions, the system removes the action, or reselects a candidate proportion to receive raw materials, release raw materials, or transfer a proportion, and then performs proportion feasible domain projection and counterfactual verification again.

[0094] When multiple candidate proportion transfer actions pass counterfactual verification, the system sorts the corresponding counterfactual candidate production prescriptions according to the counterfactual quality benefits, quality indicator priority, and auxiliary comprehensive score, and then passes them to the candidate prescription iteration and sorting module.

[0095] The following values ​​are used to illustrate the technical process of this embodiment and do not represent the actual production data of a specific company.

[0096] Suppose a candidate production formula contains raw materials A, B, C, and D, and its nominal blending ratio is: , The currently predicted strength, evenness CV, hairiness, and elongation are as follows: , The order quality requirements are: strength not less than 18.00, evenness CV not higher than 12.50, hairiness not higher than 4.00, and elongation between 6.00 and 7.00. Therefore, the current candidate production formulation has issues with insufficient strength and excessive hairiness.

[0097] Based on the target-specific contribution deviation, raw material properties, and proportion adjustment space, the system generates two candidate proportion transfer actions: , as well as: , For candidate proportion shifting action The adjusted candidate blending ratio is as follows: , The re-prediction yielded: , After adjustment, the predicted strength meets the order requirements, the predicted quality gap of the feathers is reduced, and the CV of the yarn and the elongation do not exceed the quality boundary. Therefore, the system retains this candidate ratio transfer action.

[0098] For candidate proportion shifting action The adjusted candidate blending ratio is as follows: , The re-prediction yielded: , Although the prediction strongly meets the order requirements, the predicted stem CV exceeds the order limit, and the predicted feathering is not improved. Therefore, the system removes this candidate ratio shift action.

[0099] Based on the above verification, the system retains the counterfactual candidate production prescription formed by transferring 2% of raw material D to raw material A, and uses it to generate the next round of candidate production prescriptions.

[0100] This embodiment uses the target-specific contribution bias to generate candidate proportion transfer actions, and further performs proportion feasible region projection and counterfactual quality prediction verification, so that the blending proportion adjustment is based on the joint verification of quality improvement results and production constraint satisfaction results, avoiding the direct conversion of the internal contribution relationship of the model into unverified physical proportion adjustment.

[0101] In another embodiment, the system uses the candidate production prescription set generated in step S200 as the initial candidate set, and generates new candidate production prescriptions round by round through process-aware fusion network prediction, quality gap and target-specific contribution deviation analysis, proportional transfer action generation, proportional feasible region projection and counterfactual quality verification.

[0102] Let the first The set of candidate production prescriptions is represented as follows: , in, Indicates the iteration round, Indicates the first Number of candidate production prescriptions in the round Indicates the first The first in the wheel One candidate production prescription. The set of candidate production prescriptions generated in step S200 is determined as the initial candidate set. .

[0103] For the For each candidate production prescription in the round, the system inputs it into the process-aware fusion network described in the above embodiments to re-obtain the predicted quality index and the target-specific effective contribution weight matrix. The system calculates the quality gap based on the difference between the predicted quality index and the order quality target, and calculates the target-specific contribution deviation based on the difference between the target-specific effective contribution weight and the nominal blending ratio.

[0104] The target-specific effective contribution weight and target-specific contribution deviation are recalculated in each iteration, instead of directly using the calculation results of the previous round of candidate production prescriptions, in order to reflect the changes in the relative contribution of each candidate raw material to different quality indicators after changes in the blending ratio or process conditions.

[0105] The system constructs a proportion transfer diagram with candidate raw materials as nodes according to the method described in the embodiment, and generates candidate proportion transfer actions based on quality gap, target specific contribution deviation, raw material detection attributes, proportion adjustment space, inventory status and cost information.

[0106] Each candidate ratio transfer action is projected through the ratio feasible region. Candidate ratio transfer actions that cannot meet the ratio comprehensive constraints, single raw material ratio boundaries, inventory constraints, raw material combination rules, or order constraints are eliminated by the system.

[0107] The system combines the candidate blending ratios projected from the feasible region with the corresponding candidate raw material set, front-end process conditions, and downstream process conditions to form counterfactual candidate production prescriptions, and then calls the process-aware fusion network again for quality prediction.

[0108] When the quality gap of a counterfactual candidate production prescription is improved, other constrained quality indicators do not exceed the preset risk boundary, and the counterfactual quality benefits meet the preset conditions, the system retains the counterfactual candidate production prescription; otherwise, the corresponding candidate proportion transfer action is removed.

[0109] By the The candidate production prescriptions generated in each round and verified through counterfactual analysis collectively form a temporary candidate set: , in, Indicates the candidate production prescription Generate a set of candidate production prescriptions that pass counterfactual verification.

[0110] The system merges duplicate production prescriptions in the temporary candidate set. For candidate production prescriptions with the same raw material set, nominal blending ratio, and process conditions, only one is retained; for candidate production prescriptions with prescription differences below a preset threshold, those with higher quality benefit scores or lower prescription boundary risks can be retained.

[0111] According to the hierarchical sorting method described in the embodiment, the system first determines the quality priority of candidate production prescriptions based on the quality compliance status, remaining quality gap, and quality benefit score, and then performs auxiliary sorting based on raw material cost, inventory utilization, delivery date matching, process stability, prescription boundary risk, and cumulative proportion transfer range.

[0112] The system selects a preset number of candidate production prescriptions from the temporary candidate set to form the next round of candidate production prescription set: , in, This indicates that the top-ranked candidates are selected based on a combination of quality-priority evaluation and supplementary comprehensive evaluation. One candidate production prescription.

[0113] In one implementation, to prevent the number of candidate production prescriptions from increasing rapidly with each iteration, the system can limit the number of retention ratio transfer actions corresponding to each candidate production prescription, as well as the total number of candidate production prescriptions that enter the next round of prediction in each round.

[0114] In another implementation, the system can retain candidate production prescriptions that have met the order quality target and whose prescription boundary risk is below a preset threshold in the current round, allowing them to directly enter the iteration end candidate set without having to perform a proportion transfer operation on them.

[0115] For candidate production prescriptions that have not yet met the order quality targets, the system re-executes the following process: .

[0116] In each iteration, the system records the source prescription, proportion of released raw materials, proportion of received raw materials, proportion of transferred raw materials, feasible region projection results, predicted quality indicators before and after adjustment, changes in quality gap, and counterfactual quality benefits of the candidate production prescription, in order to form an iterative path for the candidate production prescription.

[0117] The iterative path is used to explain to process engineers the proportional adjustment process that the recommended production prescription undergoes relative to the initial candidate production prescription, but does not change the requirement that the final production prescription must still be subject to manual review and production constraint verification.

[0118] The system stops iterating on candidate production prescriptions when any of the following conditions are met: 1. At least one candidate production prescription meets all the necessary quality indicators of the order quality objectives, and the prescription boundary risks are within the allowable range; 2. There are no candidate proportional shift actions in the current round that have passed proportional feasible region projection and counterfactual quality verification; 3. The quality improvement of the optimal candidate production prescription in two adjacent rounds is lower than the preset threshold; 4. The cumulative proportion of candidate production prescriptions transferred has reached the preset upper limit; 5. The number of iterations reaches the preset maximum number; 6. Receive a manual stop command from the process engineer.

[0119] When the iteration stops due to meeting the order quality target, the system will output the candidate production prescriptions that meet the quality target and the alternative production prescriptions with similar quality conditions to the prescription review and production execution module.

[0120] When iteration stops due to the absence of an effective proportional transfer action, quality improvement falling below a preset threshold, or reaching the maximum number of iterations, the system outputs the candidate production prescription with the smallest remaining quality gap and production risk within an acceptable range, while marking the unmet quality indicators, remaining quality gap, and reason for stopping.

[0121] Through the above-mentioned multi-round iterative approach, the system does not determine the final blending ratio all at once based on the target-specific contribution deviation. Instead, it recalculates the target-specific effective contribution weight in each round and gradually screens candidate production prescriptions through candidate ratio transfer, production constraint projection, and counterfactual quality prediction. This allows the iterative process to be controlled by both the quality improvement results and actual production constraints.

[0122] It should be noted that counterfactual means that the candidate production prescription has not yet been actually put into production and is only a hypothetical production plan constructed based on proportional transfer actions, used for virtual verification in the prediction model.

[0123] In step S500 of some embodiments of the present invention, the output of recommended production prescriptions based on verified candidate production prescriptions includes: S501. Calculate the counterfactual quality gain based on the change in the difference before and after verification; Specifically, quality priorities are determined based on whether candidate production prescriptions meet order quality targets, remaining quality gaps, the degree of quality improvement relative to the initial prescription, and counterfactual quality benefits. For candidate production prescriptions whose quality priorities meet preset conditions, auxiliary ranking is then performed based on one or more of the following: raw material costs, inventory consumption, delivery date matching degree, process stability, prescription boundary risk, and cumulative proportional transfer range.

[0124] S502. In response to the counterfactual quality benefit meeting the preset conditions and the constraint indicators in the target quality indicators not exceeding the risk boundary, the corresponding candidate proportion transfer action is retained, and the next round of candidate production prescriptions is generated based on the retained action for iterative verification until the iteration stop condition is met.

[0125] When multiple candidate production prescriptions meet the order quality objectives, the candidate production prescription with the higher auxiliary comprehensive score is selected first; when there are no candidate production prescriptions that fully meet the order quality objectives, the candidate production prescription with quality risk within the allowable range is selected based on the priority of quality indicators, the remaining quality gap, and the counterfactual quality benefits.

[0126] S503. Based on the degree of improvement of quality gap and the auxiliary score of production constraints, sort the candidate production prescriptions after the iteration stops, and output the recommended production prescription.

[0127] Furthermore, based on the degree of improvement in quality gaps and production constraint-assisted scoring, the candidate production prescriptions after the iteration stops are ranked, and the recommended production prescriptions are output, including: S5031. Based on the numerical deviation between the predicted quality index and the allowable range corresponding to the target quality index, calculate the quality deviation of each quality index; S5032. Sum the changes in the quality deviation before and after the iterative adjustment by weighting, and obtain a quality benefit score that represents the degree of improvement; Specifically, when multiple candidate production prescriptions meet the order quality objectives, the candidate production prescription with the higher auxiliary comprehensive score is selected first; when there are no candidate production prescriptions that fully meet the order quality objectives, the candidate production prescription with quality risk within the allowable range is selected based on the priority of quality indicators, the remaining quality gap, and counterfactual quality benefits.

[0128] S5033. Calculate the production constraint auxiliary score of the candidate production prescription based on the business evaluation dimension and the preset penalty dimension, wherein the business evaluation dimension includes at least one of cost, inventory and delivery date matching degree, and the preset penalty dimension includes at least one of quality boundary risk and prescription adjustment range risk; Specifically, the output includes a recommended production formula and at least one alternative production formula. The output includes one or more of the following: raw material batch number, recommended blending ratio, recommended process route, recommended process parameters, predicted quality indicators, quality benefit score, auxiliary comprehensive score, explanation of the target-specific contribution, proportion transfer actions retained through counterfactual verification, and constraint verification results.

[0129] S5034. Based on whether the candidate production prescriptions meet all target quality indicators, determine the basic judgment priority, and combine the calculation results of the quality benefit score and the production constraint auxiliary score to screen and output the recommended production prescriptions from the candidate production prescriptions.

[0130] Specifically, process engineers perform operations such as confirming, modifying, rejecting, or regenerating recommended production prescriptions. When process engineers modify raw material batch numbers, blending ratios, or process parameters, the system re-enters the modified production prescription into the process-aware fusion network and re-executes quality prediction and constraint verification.

[0131] The production prescription that passes quality prediction and production constraint verification is determined as the final production prescription and sent to the manufacturing execution system to generate production work orders or process parameter sheets.

[0132] Record the differences in prescriptions before and after manual review, the review results, and the reasons for the review. Use this information as update data for subsequent similar prescription recall, candidate ratio transfer generation, and candidate production prescription ranking.

[0133] In one specific embodiment, the auxiliary comprehensive score is not used to replace counterfactual quality verification. Candidate production prescriptions that fail counterfactual quality verification, violate proportional feasible region constraints, or pose unacceptable quality risks are not included in the recommended results due to high cost, inventory, or delivery time scores.

[0134] Let the first The candidate production formulation was in the first The normalized quality gap under each quality indicator is The weighted residual quality gap of the candidate production prescription is then expressed as: ,in, Indicates the quantity of quality indicators. Indicates the first Priority weights for quality indicators.

[0135] When all the necessary quality indicators of a candidate production prescription meet the order quality target, the quality gap of each necessary quality indicator is zero, and the system marks the candidate production prescription as a quality-compliant candidate production prescription.

[0136] When the candidate production prescription does not fully meet the order quality target, the weighted residual quality gap The smaller the value, the smaller the overall deviation between the candidate production prescription and the order quality target.

[0137] For the first generation generated from the pre-adjustment candidate production prescription The quality benefit score of each candidate production prescription is expressed as follows: , in, This indicates that the candidate production formulation before adjustment was in the [number]th [year]. The quality gap under each quality indicator. A higher quality benefit score indicates a greater improvement in the weighted quality gap of the candidate production prescription compared to the original candidate production prescription.

[0138] The system determines the quality priority of candidate production prescriptions in the following order: First, prioritize candidate production prescriptions that meet all necessary quality indicators to achieve the order quality objectives. Second, among the candidate production prescriptions that all meet the order quality objectives, priority should be given to the candidate production prescriptions with a larger safety margin between the predicted quality indicators and the order quality boundary. Third, among the candidate production prescriptions that do not fully meet the order quality targets, priority should be given to the candidate production prescriptions with smaller weighted residual quality gaps. Fourth, when the weighted residual quality gap is the same or within a preset range, the candidate production prescription with the higher quality benefit score is given priority.

[0139] For candidate production prescriptions with the same quality priority or at the same preset quality level, the system further calculates an auxiliary comprehensive score. The auxiliary comprehensive score can be expressed as: , in, Indicates the first The first candidate production prescription Positive auxiliary scoring, Indicates the first Risk or penalty indicators, and This indicates the corresponding weight.

[0140] The positive auxiliary scoring includes one or more of the following: raw material cost scoring, inventory utilization scoring, delivery date matching scoring, and process stability scoring; the risk or penalty indicators include one or more of the following: prescription boundary risk, quality boundary risk, and cumulative proportion transfer range.

[0141] Raw material cost scores are determined based on the relationship between the overall raw material cost of candidate production prescriptions and the upper limit of order costs. Under the premise of meeting order cost constraints, the lower the overall raw material cost, the higher the corresponding raw material cost score.

[0142] Inventory utilization scores are determined based on the available inventory of candidate raw materials, inventory age, inventory turnover priority, and material input requirements. Provided quality requirements are met, the inventory utilization score can be increased for raw materials with sufficient inventory or those requiring priority consumption.

[0143] Delivery date matching score is determined based on raw material availability time, machine availability time, estimated production cycle, and order delivery deadline. Candidate production formulas that can meet order delivery requirements receive a higher delivery date matching score.

[0144] The process stability score is determined based on the degree of deviation between the candidate process parameters and the historical stable process range. The closer the candidate process parameters are to the historical stable production range, the higher the corresponding process stability score.

[0145] Recipe boundary risk is determined based on one or more of the following: the distance between the predicted quality indicator and the order quality boundary; the distance between the candidate blending ratio and its upper and lower limits; and the distance between the candidate process parameters and the equipment process boundary. The closer the candidate production recipe is to the quality boundary, ratio boundary, or equipment process boundary, the higher the recipe boundary risk.

[0146] The cumulative percentage shift indicates the overall adjustment degree of the current candidate production prescription relative to the initial candidate production prescription. When the cumulative percentage shift is large, the system can increase the corresponding risk penalty to reduce the production switchover risk caused by excessive prescription changes.

[0147] The system sorts candidate production prescriptions based on quality priority and auxiliary comprehensive score, and the sorting order is as follows: .

[0148] When multiple candidate production prescriptions meet the order quality objectives, the system prioritizes the candidate production prescriptions that meet the preset requirements for quality and safety margin and have a higher auxiliary comprehensive score.

[0149] When there are no candidate production prescriptions that fully meet the order quality objectives, the system can select candidate production prescriptions with a small weighted remaining quality gap, a high quality benefit score, and quality risks within the allowable range. However, it should also output the unmet quality indicators, remaining quality gaps, and corresponding risk warnings, and submit them to process personnel for review.

[0150] After sorting, the system will determine the candidate production prescription with the highest ranking and that meets the production constraints as the recommended production prescription, and select at least one other candidate production prescription as an alternative production prescription.

[0151] Recommended and alternative production prescriptions include one or more of the following: raw material batch number, nominal blending ratio, upstream process conditions, downstream process conditions, predicted quality indicators, remaining quality gap, quality benefit score, auxiliary comprehensive score, proportion transfer actions retained after counterfactual verification, and production constraint verification results.

[0152] By using the above-mentioned hierarchical sorting method, this embodiment places quality compliance and quality improvement before cost, inventory, delivery time and process stability evaluation, avoiding the use of a single comprehensive score to offset quality gaps with cost or inventory advantages, and enabling the selection of a recommended production prescription that better meets actual production requirements among candidate production prescriptions with the same or similar quality conditions.

[0153] After the candidate blending ratio is determined, the candidate process parameters are further perturbed to generate more candidate production formulas that meet production constraints.

[0154] The set of candidate process parameters can be represented as: , Among them, the candidate process parameters include one or more of the following: cotton waste rate, roving draft ratio, yarn draft ratio, single yarn twist, ply twist, yarn count, spinning method, and spinning process.

[0155] The system generates one or more continuous process parameters by perturbing them based on equipment process boundaries, historically stable process areas, and order requirements. For example, the system can increase or decrease the cotton waste rate based on historically similar prescriptions by a preset step size, generating multiple candidate values ​​within the allowable ranges of roving draft ratio, yarn draft ratio, single yarn twist, or ply twist. For categorical process parameters, the system can limit the selectable categories based on order requirements and historically similar prescriptions.

[0156] For the generated candidate production prescriptions, the system executes the following rejection rules: First, eliminate candidate production prescriptions that exceed the equipment's process boundaries; Second, eliminate candidate production prescriptions that do not meet inventory constraints or cost constraints; Third, candidate production prescriptions whose process parameters deviate from the historical stable process range by more than a preset threshold are eliminated; Fourth, candidate production prescriptions whose quality prediction results clearly do not meet the order quality targets are eliminated.

[0157] Through the above methods, the system can expand the candidate prescription search space while ensuring feasibility, so that the recommended production prescription not only includes raw material ratio optimization, but also process parameter co-optimization.

[0158] As can be understood, the above steps complete the final evaluation and ranking of the candidate production prescription set at the end of the iteration, and output recommended production prescriptions and alternative production prescriptions. The output includes one or more of the following: recommended raw material batch number, recommended blending ratio, recommended process conditions, predicted quality indicators, quality benefit score, auxiliary comprehensive score, explanation of target specific contribution, retained proportion transfer action, counterfactual verification results, and constraint verification results.

[0159] Recommended production prescriptions are manually reviewed before being sent to the Manufacturing Execution System (MES). When process engineers modify raw material batch numbers, blending ratios, or process parameters, the modified production prescriptions are re-tested for quality prediction and constraint verification. The differences between the prescriptions before and after manual review, the reasons for the review, and the review results are recorded as update data for subsequent similar prescription recall, ratio transfer generation, and candidate production prescription ranking.

[0160] refer to Figure 4 In some embodiments of the present invention, corrections based on production execution feedback are also included: S601. Obtain the actual yarn quality test results after the recommended production prescription is executed, and calculate the prediction residual between the actual yarn quality test results and the predicted quality index; Specifically, for the first The quality indicators and their predicted residuals are expressed as follows: , in, This indicates the actual yarn quality test results. This represents the predicted quality indicators before production execution. Historical residuals are grouped into different granular groups based on one or more of the following: raw material batch, raw material combination type, nominal blending ratio distribution, yarn count, color number, spinning method, front-end process conditions, downstream process parameters, process route, and machine. The actual yarn quality inspection results after the final production formula are executed are used to generate predicted residuals based on the differences between the actual yarn quality inspection results and the predicted quality indicators before production execution.

[0161] S602. Establish historical residual grouping at multiple levels, and calculate the prescription similarity between the next round of candidate production prescriptions and the historical residual groupings; Specifically, historical residuals of different fine granularities are grouped according to one or more of the following: raw material combination, nominal blending ratio, yarn count, color number, spinning method, front-end process conditions, downstream process conditions, process route, and machine.

[0162] S603. In response to the prescription similarity reaching a preset threshold, a residual correction amount is generated based on the residual statistics of the corresponding historical residual group and the sample confidence factor to correct the prediction quality index of the next round of candidate production prescriptions.

[0163] Specifically, for subsequent candidate production prescriptions, the prescription similarity between them and each historical residual group is calculated. Only when the prescription similarity reaches a preset threshold does the corresponding historical residual group participate in residual correction; the residual correction amount is generated based on the historical residual statistics, the confidence factor corresponding to the number of historical samples, and the prescription similarity.

[0164] Calculate the grouping confidence factor based on the number of valid historical samples included in the historical residual grouping: , in, Indicates the first The number of valid historical samples contained in each historical residual group This indicates the preset confidence level shrinkage parameter. The fewer the historical samples, the greater the degree of shrinkage in the residual correction.

[0165] The system generates residual correction values ​​based on historical residual statistics, prescription similarity, and grouping confidence factors, and uses these residual correction values ​​to correct the predicted quality indicators of subsequent similar candidate production prescriptions.

[0166] Preferably, the generated residual correction amount includes: S6031. Prioritize selecting fine-grained historical residual groups that meet the similarity threshold and whose sample size meets the preset conditions to calculate the residual correction amount; Specifically, when the fine-grained historical residual grouping does not meet the preset sample quantity or similarity conditions, it regresses to the higher-level residual grouping composed of fewer production dimensions; when none of the preset conditions are met, no grouping residual correction is performed, or a global residual correction amount is used.

[0167] S6032. In response to the condition that the fine-grained historical residual grouping does not meet the requirements, backtrack to the higher-level historical residual grouping with fewer dimensions.

[0168] Specifically, the residual correction amount is used to correct the predicted quality indicators, quality benefit scores, or auxiliary comprehensive scores of subsequent similar candidate production prescriptions; when the predicted residuals of the same historical residual group continue to show the same direction and exceed the preset update conditions, the parameter calibration, incremental update, or retraining of the process-aware fusion network is triggered.

[0169] In one specific embodiment, after the recommended production prescription is reviewed and approved by process engineers, the prescription review and production execution distribution module sends the final production prescription to the manufacturing execution system. The final production prescription includes at least the raw material batch number, nominal blending ratio, upstream process conditions, downstream process conditions, production machine, target yarn count, and corresponding prescription version identifier.

[0170] The manufacturing execution system generates a production work order or process parameter sheet based on the final production prescription, and associates the prescription version identifier with the order number, production work order number, and yarn batch number to enable traceability of the production prescription, production execution records, and yarn quality inspection results.

[0171] During the final production formulation execution process, the manufacturing execution system records one or more of the following: the actual batch number of raw materials used, the actual amount of materials used, the actual blending ratio, the actual process parameters, the production machine, the production shift, the production time, and the abnormal handling records.

[0172] When there are discrepancies between the actual execution data and the planned data in the final production prescription, the system retains both the planned prescription data and the actual execution data. The planned prescription data is used to record the prescription decision results upon approval, while the actual execution data is used for subsequent prediction residual calculations and production feedback analysis.

[0173] After production is completed, the system obtains the actual yarn quality test results for the corresponding production batch from the laboratory information management system or yarn quality testing equipment. The actual yarn quality test results include one or more of the following: actual strength, actual evenness (CV), actual hairiness, and actual elongation.

[0174] The system uses the production order number or yarn batch number as the associated index to align the following data: 1. The final manufacturing prescription and its version; 2. Predicted quality indicators before production execution; 3. The actual batch numbers of raw materials used and the actual blending ratio; 4. Actual process parameters and machine information; 5. Actual yarn quality inspection results; 6. Records of production anomalies and human intervention.

[0175] After data alignment is completed, a production feedback sample is formed. This production feedback sample is used to represent the predicted results, execution status, and actual quality results of the final production prescription under actual production conditions.

[0176] In one implementation, a production feedback sample can be represented as: , in, Indicates the first The planned production formula for each production batch Indicates the actual prescription being executed. This indicates the predicted quality indicators before production execution. This indicates the actual yarn quality test results. It displays production environment information such as machine, shift, production time, and abnormal records.

[0177] When the planned production prescription differs from the actual prescription, the system uses the actual prescription as the primary basis for subsequent prediction residual calculation and similar prescription matching, and retains the difference between the planned prescription and the actual prescription as a production deviation record.

[0178] The system performs integrity and consistency checks on production feedback samples. These checks include one or more of the following: batch correlation verification, quality indicator unit verification, actual blending ratio comprehensive constraint verification, process parameter effective range verification, and outlier identification.

[0179] For production feedback samples that lack key raw material batch numbers, actual blending ratios, main process parameters, or actual quality test results, the system marks them as incomplete samples and does not directly use them for residual correction; for samples with production or testing anomalies, the system records the corresponding anomaly identifiers and decides whether to include them in subsequent residual statistics based on preset rules.

[0180] The validated production feedback samples are sent to the similarity gating and reliable hierarchical residual correction module, which is used to establish historical residual grouping, calculate prescription similarity, generate reliable residual correction amount, and determine whether to trigger model parameter update.

[0181] This embodiment only illustrates the data formation process from the final production prescription's review and issuance, actual execution, to the feedback of yarn quality results. Prescription similarity calculation, credibility reduction, residual correction amount generation, and hierarchical rollback methods are further explained in specific embodiments.

[0182] refer to Figure 5 In a second aspect, the present invention provides a colored yarn quality prediction and production prescription closed-loop optimization system 1, comprising: an acquisition module 11, configured to acquire a target quality index and a set of available raw materials, and generate at least one candidate production prescription including an initial blending ratio, front-end process conditions, and downstream process conditions; a modulation module 12, configured to input the candidate production prescription into a pre-trained prediction model; in the prediction model, using the front-end process conditions to modulate the feature representation of each raw material in the set of available raw materials and then performing interactive encoding to generate a target-specific effective contribution weight; based on the target-specific effective contribution weight, performing weighted aggregation of the interactively encoded features, and using the downstream process conditions to modulate the aggregated features, and outputting a predicted quality index; a generation module 13, configured to generate a candidate ratio transfer action for the set of available raw materials based on the difference between the predicted quality index and the target quality index, and the deviation between the target-specific effective contribution weight and the corresponding initial blending ratio; a verification module 14, configured to re-input the counterfactual candidate production prescription adjusted by the candidate ratio transfer action into the prediction model for counterfactual verification; and an output module 15, configured to output the verified candidate production prescription as a recommended production prescription for production execution.

[0183] refer to Figure 6 This embodiment provides a closed-loop optimization system for colored yarn quality prediction and production prescription based on a process-aware fusion network, including a multi-source data access module 10, a batch-level data fusion module 20, an order quality target parsing module 30, a raw material candidate pool construction module 40, a candidate production prescription generation module 50, a process-aware fusion network prediction module 60, a quality gap and contribution deviation joint analysis module 70, a proportion transfer graph construction module 80, a proportion feasible region projection module 90, a counterfactual prescription verification module 100, a candidate prescription iteration and ranking module 110, a prescription review and production execution issuance module 120, and a similarity gating and reliable hierarchical residual correction module 130.

[0184] The multi-source data access module 10 connects to one or more of the enterprise resource planning system, warehouse management system, laboratory information management system, and manufacturing execution system to acquire order, inventory, raw material testing, historical prescriptions, process execution, and yarn quality testing data. The batch-level data fusion module 20 uses the yarn batch number or production work order number as an index to associate the multi-source data, forming batch-level production samples that can be used for model training, historical prescription retrieval, quality prediction, and residual correction.

[0185] The order quality target parsing module 30 parses one or more quality targets from the order information, including strength, evenness CV, hairiness, and elongation, and determines the corresponding quality direction and production constraints. The raw material candidate pool construction module 40 filters candidate raw materials based on the quality targets, raw material inventory, raw material testing attributes, cost, and raw material availability. The candidate production prescription generation module 50 generates candidate production prescriptions, including a set of candidate raw materials, nominal blending ratio, front-end process conditions, and downstream process conditions, based on the available raw material candidate pool, historical similar prescriptions, and production constraints.

[0186] The process-aware fusion network prediction module 60 receives candidate production prescriptions, embeds the detection attributes of each candidate raw material into a representation, injects front-end process conditions before the raw material set interaction encoding, performs interaction encoding on unordered and variable-length candidate raw material sets, and generates target-specific effective contribution weights for different quality indicators. These target-specific effective contribution weights are used to aggregate the interaction enhancement representations of the candidate raw materials, and after aggregation, are injected into the downstream process conditions, outputting the predicted quality indicators and the target-specific effective contribution weight matrix.

[0187] The joint analysis module 70 for quality gaps and contribution deviations receives predicted quality indicators, order quality targets, nominal blending ratios, and target-specific effective contribution weight matrices. It calculates the quality gap for each quality indicator and the target-specific contribution deviation for each candidate raw material, and determines the candidate ratio for receiving raw materials, releasing raw materials, and adjustment priorities. The target-specific contribution deviation is used to generate candidate ratio transfer actions and is not directly used as the final physical blending ratio adjustment result.

[0188] The proportional transfer graph construction module 80 uses candidate raw materials as nodes and establishes directed proportional transfer edges between candidate proportional release raw materials and candidate proportional receiving raw materials. Each directed proportional transfer edge corresponds to a candidate proportional transfer action, which is used to transfer the preset proportion from one candidate raw material to another candidate raw material.

[0189] The proportional feasible region projection module 90 performs constraint verification on the temporary blending ratios formed by candidate proportional transfer actions and projects them to the proportional feasible region that satisfies the comprehensive proportional constraints, single raw material ratio boundaries, inventory constraints, raw material combination rules, and order constraints. Candidate proportional transfer actions for which no feasible projection results exist are eliminated.

[0190] The counterfactual prescription verification module 100 combines the projected candidate blending ratios with the candidate raw material set and process conditions to form a counterfactual candidate production prescription, and then calls the process-aware fusion network prediction module 60 to re-predict the quality. Based on the changes in the quality gap before and after adjustment, the risk boundaries of other quality indicators, and the results of meeting production constraints, the counterfactual prescription verification module 100 determines whether the candidate ratio transfer action is retained.

[0191] The candidate prescription iteration and ranking module 110 generates the next round of candidate production prescriptions based on the candidate proportion transfer actions that have passed counterfactual verification, calculates the quality benefit score based on the reduction in quality gap, and calculates an auxiliary comprehensive score based on one or more of the following: cost, inventory, delivery time, process stability, prescription boundary risk, and cumulative proportion transfer range.

[0192] When the candidate production prescription has not yet met the iteration stop condition, the candidate prescription iteration and sorting module 110 resends the next round of candidate production prescriptions to the process-aware fusion network prediction module 60, and repeats the generation of target-specific effective contribution weights, contribution deviation analysis, proportional transfer, feasible region projection and counterfactual verification. The iteration stops when the order quality target is met, there is no proportional transfer action that reaches the preset benefit condition, the improvement of two adjacent rounds is lower than the preset threshold, the maximum number of iteration rounds is reached, or a manual stop instruction is received.

[0193] The prescription review and production execution module 120 performs a final evaluation of the candidate production prescriptions at the end of the iteration, outputs recommended production prescriptions and alternative production prescriptions, and displays one or more of the following: predicted quality indicators, explanation of target-specific contributions, proportional transfer actions, counterfactual quality benefits, and constraint verification results.

[0194] When process engineers modify the recommended production prescription, the modified production prescription is resent to the process perception fusion network prediction module 60 for quality prediction and constraint verification; the final production prescription that has been approved is issued to the manufacturing execution system to generate production work orders or process parameter sheets.

[0195] After the final production prescription is executed, the similarity gating and reliable hierarchical residual correction module 130 obtains the actual yarn quality inspection results and calculates the predicted residual between the actual inspection results and the predicted results before production execution. This module establishes multi-level historical residual grouping according to one or more of the following: raw material combination, nominal blending ratio, yarn count, color number, process conditions, process route, and machine.

[0196] For subsequent candidate production prescriptions, the similarity gating and credibility hierarchical residual correction module 130 calculates the prescription similarity between it and the historical residual group, and generates the residual correction amount by combining the historical residual statistics and the credibility factor corresponding to the number of historical samples. When the fine-grained historical residual group does not meet the sample number or similarity conditions, it backtracks to a higher-level historical residual group.

[0197] The residual correction amount is fed back to the process-aware fusion network prediction module 60 or the candidate prescription iteration and sorting module 110 to correct the predicted quality indicators, quality benefit scores or auxiliary comprehensive scores of subsequent similar candidate production prescriptions, thereby forming an inner loop of candidate prescription prediction and iteration, and an outer loop of production execution feedback.

[0198] The above modules can be set up separately in servers, industrial computers, edge computing devices or production management terminals, or two or more modules can be integrated into the same processing device.

[0199] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of the first aspect of the present invention.

[0200] Electronic devices may include processing units (such as central processing units, graphics processing units, etc.) that can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). RAM also stores various programs and data required for the operation of the electronic device. The processing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0201] The following devices can typically be connected to the I / O interface: input devices such as touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices such as liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices such as hard disks; and communication devices. Communication devices allow electronic devices to communicate wirelessly or wiredly with other devices to exchange data.

[0202] The above description is only a preferred 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 of color spun yarn quality prediction and production recipe closed loop optimization, characterized in that, include: Obtain the target quality indicators and the set of available raw materials, and generate at least one candidate production formula that includes the initial blending ratio, front-end process conditions and downstream process conditions; The candidate production prescription is input into a pre-trained prediction model; in the prediction model, the feature representation of each raw material in the available raw material set is modulated using the front-end process conditions and then interactive coding is performed to generate target-specific effective contribution weights; the interactively coded features are weighted and aggregated based on the target-specific effective contribution weights, and the aggregated features are modulated using the downstream process conditions to output the predicted quality index. Based on the difference between the predicted quality index and the target quality index, and the deviation between the target specific effective contribution weight and the corresponding initial blending ratio, a candidate ratio transfer action is generated for the available raw material set. The counterfactual candidate production prescriptions adjusted by the aforementioned candidate ratio transfer action are re-input into the prediction model for counterfactual verification. The validated candidate production prescriptions will be output as recommended production prescriptions for production execution.

2. The spun color yarn quality prediction and production recipe closed loop optimization method according to claim 1, characterized in that, The generated target-specific effective contribution weights include: Generate corresponding contribution correction scores for each of the different target quality indicators; The contribution correction score is combined with the corresponding initial blending ratio so that the generated target-specific effective contribution weight represents the relative contribution estimate of each raw material in the available raw material set for a specific quality index.

3. The spun color yarn quality prediction and production recipe closed loop optimization method according to claim 1, wherein, The process of generating candidate ratio shifts for the available raw material set includes: Based on the differences and deviations, candidate proportions of receiving raw materials and candidate proportions of releasing raw materials are determined from the set of available raw materials. Establish a proportional transfer rule between the candidate proportional release material and the candidate proportional receiving material, and generate an initial proportional transfer action; The temporary blending ratio formed by the initial ratio transfer action is projected onto a preset ratio feasible region to eliminate actions that do not meet production constraints, thereby obtaining the candidate ratio transfer actions.

4. The spun color yarn quality prediction and production recipe closed loop optimization method of claim 1, wherein, The recommended production prescriptions output based on the validated candidate production prescriptions include: Calculate the counterfactual quality gain based on the change in the difference before and after verification; In response to the counterfactual quality benefit meeting the preset conditions and the constraint indicators in the target quality indicators not exceeding the risk boundary, the corresponding candidate proportion transfer action is retained, and the next round of candidate production prescriptions is generated based on the retained action for iterative verification until the iteration stop condition is met. Based on the degree of improvement in quality gaps and production constraint auxiliary scoring, the candidate production prescriptions after the iteration stops are ranked, and the recommended production prescription is output.

5. The spun color yarn quality prediction and production recipe closed loop optimization method according to claim 4, characterized in that, The process of ranking candidate production prescriptions after iteration stops based on the degree of improvement in quality gaps and production constraints, and outputting recommended production prescriptions, includes: Based on the numerical deviation between the predicted quality index and the allowable range corresponding to the target quality index, the quality deviation of each quality index is calculated. The changes in the quality deviation before and after the iterative adjustment are weighted and summed to obtain a quality gain score that represents the degree of improvement. Based on business evaluation dimensions and preset penalty dimensions, the production constraint auxiliary score of the candidate production prescription is calculated, wherein the business evaluation dimensions include at least one of cost, inventory and delivery date matching degree, and the preset penalty dimensions include at least one of quality boundary risk and prescription adjustment range risk; Based on whether the candidate production prescriptions meet all target quality indicators, the basic judgment priority is determined, and the recommended production prescriptions are selected and output from the candidate production prescriptions by combining the calculation results of the quality benefit score and the production constraint auxiliary score.

6. The spun color yarn quality prediction and production recipe closed loop optimization method of claim 1, wherein, It also includes corrections based on production execution feedback: Obtain the actual yarn quality test results after the recommended production prescription is implemented, and calculate the prediction residual between the actual yarn quality test results and the predicted quality index; Establish historical residual groupings at multiple levels, and calculate the prescription similarity between the next round of candidate production prescriptions and the historical residual groupings; In response to the prescription similarity reaching a preset threshold, a residual correction amount is generated based on the residual statistics of the corresponding historical residual group and the sample confidence factor to correct the prediction quality index of the next round of candidate production prescriptions.

7. The spun color yarn quality prediction and production recipe closed loop optimization method according to claim 6, characterized in that, The generated residual correction amount includes: The residual correction amount is calculated by preferentially selecting fine-grained historical residual groups that meet the similarity threshold and the sample size meets the preset conditions; In response to the failure of fine-grained historical residual grouping to meet the conditions, backtracking matching is performed sequentially to higher-level historical residual groupings with fewer dimensions.

8. A colored spun yarn quality prediction and production recipe closed loop optimization system, characterized in that, include: The acquisition module is used to acquire the target quality indicators and the set of available raw materials, and generate at least one candidate production formula containing the initial blending ratio, front-end process conditions and downstream process conditions. The modulation module is used to input the candidate production prescription into a pre-trained prediction model; in the prediction model, the feature representation of each raw material in the available raw material set is modulated using the front-end process conditions and then interactive encoding is performed to generate target-specific effective contribution weights; the interactively encoded features are weighted and aggregated based on the target-specific effective contribution weights, and the aggregated features are modulated using the downstream process conditions to output a prediction quality index. The generation module is used to generate candidate ratio transfer actions for the set of available raw materials based on the difference between the predicted quality index and the target quality index, as well as the deviation between the target specific effective contribution weight and the corresponding initial blending ratio. The verification module is used to re-input the counterfactual candidate production prescriptions adjusted by the candidate ratio transfer action into the prediction model for counterfactual verification. The output module is used to output the validated candidate production prescriptions as recommended production prescriptions for production execution.

9. An electronic device comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the color-spun yarn quality prediction and production formula closed-loop optimization method as described in any one of claims 1 to 7.

10. A computer readable medium having stored thereon a computer program, wherein, When the computer program is executed by the processor, it implements the method for predicting the quality of colored spun yarn and optimizing the production formula as described in any one of claims 1 to 7.