A method and system for intelligent proportioning of regenerated fiber raw materials based on deep learning technology

Abstract

The present application relates to the technical field of regenerated polyester fiber production, and particularly relates to a method and system for intelligent proportioning of regenerated fiber raw materials based on deep learning technology. The method collects ERP, MES, market, environment and equipment data, constructs a raw material quality instability exclusion matrix, and forms a candidate raw material pool; a deep learning model is used to predict quality, cost, efficiency and confidence, and the target proportioning is solved under the double constraints of price fluctuation risk and prediction confidence; the results are decomposed into work orders and the amount of material under each time window is issued for execution; when the deviation exceeds the threshold, the executed part is frozen, only the remaining batches are locally re-optimized, and the exclusion matrix and model are updated by increasing the weight of the sample. The scheme can improve quality stability, reduce cost and enhance continuous production capacity.

Description

Technical Field

[0001] This invention relates to the field of recycled polyester fiber production technology, and in particular to a method and system for intelligent proportioning of recycled fiber raw materials based on deep learning technology. Background Technology

[0002] In the production of recycled fibers, raw material sources are typically dispersed, exhibiting characteristics such as significant batch-to-batch variations, substantial fluctuations in impurity content, rapid market price changes, and strong environmental disturbances. Currently, most companies rely on manual experience, fixed BOM formulations, or simple cost accounting when blending raw materials. They usually determine an approximate blending ratio based on inventory and purchase prices, then make adjustments afterward based on finished product quality results. While this method is simple to implement, it fails to adequately utilize multi-source information such as historical quality data, real-time market prices, equipment feeding deviations, and environmental changes. This can easily lead to delayed blending decisions, increased product quality fluctuations, and higher raw material costs. Furthermore, if feeding deviations occur during production, adjustments often rely on temporary corrections by operators, making it difficult to balance quality stability, cost control, and continuous production requirements. These problems are particularly pronounced under conditions of increasing raw material types, frequent batch changes, and higher quality standards in recycled fibers.

[0003] In the prior art, Chinese patent application CN109918702A discloses a collaborative multi-objective optimization method for blast furnace batching and operation. This method uses an artificial neural network to predict the batching and operation results and combines this with constraints to solve for multi-objective optimization. This approach illustrates the basic idea of ​​combining data-driven modeling with multi-objective optimization, but it primarily focuses on blast furnace smelting scenarios, emphasizing the collaborative optimization of blast furnace batching and operation parameters. It does not establish a control mechanism for raw material screening, candidate pool formation, and re-optimization of remaining batches, taking into account the characteristics of recycled fiber raw materials, such as raw material quality dispersion, price volatility, and feeding execution deviations.

[0004] In addition, Chinese patent CN113491341B discloses a method for controlling the water addition flow rate of tobacco rehydration based on historical production data modeling. This method uses statistical analysis of current batch effective production data to infer the standard water addition flow rate and performs weighted iterations between the online modeling results of the previous batch and historical values ​​to achieve online self-learning of the model. This approach reflects the idea of ​​updating the model using historical data and current batch data; however, its control object is a single water addition flow rate, focusing on feedforward and feedback moisture control. It does not address issues such as candidate raw material selection, price risk constraints, feed rate freezing, and local re-optimization for remaining batches when multiple raw materials from recycled fiber are used in the blending process.

[0005] Therefore, although existing technologies have provided some insights into data-driven prediction, multi-objective optimization, and online self-learning, they still lack a comprehensive technical solution for intelligent proportioning of recycled fiber raw materials. This solution should be able to first screen out raw materials with unstable quality under the drive of multi-source data, then determine the target proportion under the joint constraints of price risk and prediction reliability, and implement local re-optimization for the unexecuted parts during the production process, thereby achieving closed-loop collaborative control of proportioning decision-making, work order issuance, precise material feeding, and model self-learning updates. Summary of the Invention

[0006] This invention aims to address the problems of complex sources of recycled fiber raw materials, large batch quality fluctuations, rapid market price changes, and difficulty in precise and coordinated control of the production feeding process. It provides an intelligent raw material proportioning method and system based on big data and deep learning. By constructing a raw material quality instability elimination mechanism, a candidate raw material pool screening mechanism, a proportioning optimization mechanism with price risk and prediction confidence constraints, and a local re-optimization closed-loop control mechanism for remaining batches, it achieves intelligent raw material proportioning decisions, precise feeding execution, and adaptive model updates in the recycled fiber production process.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for intelligent proportioning of recycled fiber raw materials based on big data and deep learning includes the following steps: S1, collecting multi-source data from ERP, MES, on-site data acquisition devices, and external data interfaces; S2, cleaning, imputing missing data, time alignment, and feature standardization of the multi-source data, and constructing a raw material quality instability exclusion matrix based on historical quality inspection results; according to the raw material quality instability exclusion matrix, excluding raw materials whose quality instability scores exceed the instability threshold within a preset statistical period to form a candidate raw material pool for the current production batch; S3, inputting the cost characteristics, price fluctuation characteristics, historical BOM formulation characteristics, historical quality characteristics, environmental characteristics, and equipment feeding characteristics corresponding to the candidate raw material pool into a deep learning prediction model, and outputting each candidate proportioning formula. The corresponding quality prediction value, cost per ton of product prediction value, production efficiency prediction value, and prediction confidence level are calculated. S4. Based on the output results, a ratio optimization model with dual constraints is constructed. The ratio of each candidate raw material is used as the decision variable, and the minimum cost per ton of product, the maximum quality prediction value, and the maximum production efficiency prediction value are used as the joint optimization objectives. At the same time, the upper limit of price fluctuation risk and the lower limit of prediction confidence level are used as feasible domain constraints to solve for the target ratio scheme of the current production batch. S5. The target ratio scheme is decomposed into the theoretical total amount of each raw material in the current production batch, and further decomposed into the theoretical amount of each feeding time window. The decomposition results are synchronously written back to the current batch work order and feeding task in the ERP system, and synchronously sent to the belt scale control system.

[0009] Preferably, in step S1, the multi-source data includes at least: raw material purchase unit price, supplier information, inventory and batch work order information obtained from the ERP system; historical BOM formula, historical output and historical quality test results obtained from the MES system; temperature, humidity, dust concentration, real-time feed amount of belt scale and equipment operating status data obtained from the field acquisition device; and real-time unit price and price fluctuation data of raw materials in the market obtained from the external data interface.

[0010] And / or, the raw material quality instability exclusion matrix in step S2 is constructed as follows: The quality pass rate, key performance indicator volatility, and abnormal batch proportion of each raw material in production within a preset statistical period are statistically analyzed to form a raw material quality instability score Z; when Z is greater than a preset instability threshold... When this occurs, the raw material is marked as an excluded raw material and will not participate in the construction of the candidate raw material pool for the current production batch; wherein: Where P is the quality pass rate, W is the volatility of key performance indicators, E is the proportion of abnormal batches, and a, b, and c are weighting coefficients, satisfying: a+b+c=1.

[0011] Preferably, in step S3, the deep learning prediction model includes a time series encoding layer, a feature fusion layer, and a multi-task output layer; the time series encoding layer is used to extract the time-dependent features of the purchase unit price, the real-time market unit price, and price fluctuation data; the feature fusion layer is used to fuse historical BOM formula features, historical quality features, environmental features, and equipment feeding features; and the multi-task output layer simultaneously outputs the quality prediction value, the cost per ton of product prediction value, the production efficiency prediction value, and the prediction confidence level.

[0012] Preferably, in step S3, the predicted cost C per ton of product is calculated according to the following formula: ; in, Let be the mass ratio of the i-th raw material. Let T be the unit price of the i-th raw material, T be the transportation cost of the current batch, U be the converted cost of utilities for the current batch, and Y be the output of the finished product for the current batch.

[0013] And / or, in step S4, the price volatility risk value R is calculated according to the following formula: ; in, Let represent the proportion of the i-th raw material. Let be the price fluctuation coefficient of the i-th raw material within a preset time window; And / or, in step S4, the proportioning optimization model with dual constraints further includes the following constraints: The sum of the proportions of each candidate raw material is 1; The proportion of each candidate raw material shall not be lower than the lower limit of the corresponding process proportion and not higher than the upper limit of the corresponding process proportion; The planned total amount of each candidate raw material shall not exceed the available inventory of the corresponding raw material; Price volatility risk value R shall not exceed the preset risk limit. ; The prediction confidence level G is not lower than the preset confidence lower limit. .

[0014] Preferably, the method further includes: S6. During production execution, the actual feeding amount, cumulative feeding deviation, and online quality inspection results within each feeding time window are obtained in real time. When the deviation between the actual feeding amount and the theoretical feeding amount exceeds the feeding deviation threshold, or the deviation between the online quality inspection result and the quality prediction value exceeds the quality deviation threshold, the actual feeding amount of the executed part in the current production batch is frozen, and a local re-optimization model for the remaining batch is constructed only for the unfinished part. The proportion of the remaining raw materials and the theoretical feeding amount for each subsequent feeding time window are recalculated. Feeding deviation rate D is calculated according to the following formula: ; in, This represents the actual amount of material fed into the current feeding window. This represents the theoretical feed rate for the current feeding time window; S7. Synchronously write back the remaining batch local re-optimization results obtained in step S6 to the remaining batch work orders, feeding tasks and belt scale control settings in the ERP system, and add the actual feeding amount, online quality inspection results and local re-optimization results of the current production batch as incremental samples to the training dataset, and perform variable weight updates on the raw material quality instability exclusion matrix and deep learning prediction model to form a self-learning closed loop for subsequent batches.

[0015] As a preferred option, the method for establishing the partial re-optimization model for the remaining batch in step S6 is as follows: the cumulative actual amount of each raw material in the executed part of the current production batch is kept unchanged, the remaining raw material demand corresponding to the remaining planned output of the current production batch is taken as the optimization object, and only the theoretical amount of material input for the subsequent unexecuted material input time window is recalculated, so as to avoid the overall cancellation and rearrangement of the executed work orders. And / or, the weighted update in step S7 includes: assigning higher sample weights to the current batch of incremental samples than to historical samples, so as to improve the responsiveness of the raw material quality instability exclusion matrix and the deep learning prediction model to the latest changes in raw material quality, the latest changes in market price, and the latest changes in equipment status; its training loss function L is expressed as: ; in, Let j be the sample weight of the j-th sample. Let be the single-sample loss value corresponding to the j-th sample, and m be the total number of samples participating in this iteration of training.

[0016] Secondly, the present invention also provides an intelligent proportioning system for regenerated fiber raw materials for implementing the method, comprising: The data acquisition module is used to collect multi-source data from ERP systems, MES systems, field acquisition devices, and external data interfaces; The data preprocessing module is used to clean, fill in missing data, align time and standardize features of the multi-source data, and to construct a raw material quality instability exclusion matrix and a candidate raw material pool. The deep learning prediction module is used to output the quality prediction value, cost per ton of product prediction value, production efficiency prediction value and prediction confidence value for each candidate ratio scheme based on the candidate raw material pool and the processed multi-source feature data. The proportioning optimization module is used to solve the target proportioning scheme based on the predicted quality value, the predicted cost per ton of product, the predicted production efficiency value, the prediction confidence level, and the constraints. The work order integration module is used to decompose the target proportioning scheme and synchronously write it back to the current batch of work orders and material feeding tasks in the ERP system; The feeding control module is used to control the belt scale to feed materials according to the theoretical feeding amount and to collect the actual feeding amount; The local re-optimization module is used to freeze the actual feed amount of the executed part when the deviation exceeds the threshold, and only recalculates the remaining raw material ratio and subsequent theoretical feed amount for the part that has not yet been completed in the current production batch. The model update module is used to update the raw material quality instability exclusion matrix and the deep learning prediction model by taking the actual feed amount, online quality detection results and local re-optimization results as incremental samples.

[0017] Preferably, the data acquisition module includes an ERP interface unit, a MES interface unit, an environmental acquisition unit, a belt scale acquisition unit, and a market data interface unit; wherein, the ERP interface unit is used to acquire purchase unit price, supplier information, inventory quantity, and batch work order information; the MES interface unit is used to acquire historical BOM formula, historical output, and historical quality test results; the environmental acquisition unit is used to acquire temperature, humidity, and dust concentration; the belt scale acquisition unit is used to acquire real-time feed rate; and the market data interface unit is used to acquire real-time raw material market unit price and price fluctuation data. And / or, a consistent control relationship is established between the work order integration module and the material feeding control module for the remaining execution segments of the same batch. The local re-optimization results of the remaining batch output by the local re-optimization module include: the target total material feeding in the remaining batch work order, the theoretical material feeding amount of the subsequent material feeding time window, and the belt scale control setting value. The above three are updated synchronously using the same batch identifier or the same version identifier. And / or, the model update module is connected to both the data preprocessing module and the deep learning prediction module, and is used to: update the raw material quality instability exclusion matrix based on the online quality inspection results of the current batch; update the deep learning prediction model parameters based on the actual feed amount of the current batch, the online quality inspection results, and the local re-optimization results; and use the updated raw material quality instability exclusion matrix and deep learning prediction model parameters for the next batch of candidate raw material pool screening and target ratio scheme solution.

[0018] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.

[0019] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0020] Compared with existing technologies, this invention first establishes a raw material quality instability exclusion matrix by jointly collecting and processing multi-source data such as procurement, inventory, historical BOM, historical quality, market prices, environmental parameters, and equipment feeding status. This pre-screens raw materials with large quality fluctuations, reducing the probability of unstable raw materials entering the formulation scheme from the source and improving the reliability of the candidate raw material pool. Second, this invention uses a deep learning model to jointly evaluate the quality, cost, efficiency, and prediction confidence of candidate formulation schemes. During optimization, it simultaneously introduces upper limits for price fluctuation risk and lower limits for prediction confidence, ensuring that the resulting formulation scheme not only balances low cost and high quality but also possesses stronger risk control capabilities and predictive reliability. Third, during the production execution phase, when feeding deviations or quality deviations exceed limits, this invention does not recalculate the entire batch of work orders, but... This invention freezes the already executed portion and performs local re-optimization only on the incomplete portion, thereby reducing disturbances to the already executed production process, avoiding the production interruption risk caused by batch cancellation or large-scale rescheduling, and improving the stability and scheduling efficiency of continuous production. In addition, the invention synchronously writes the local re-optimization results back to the ERP work order, material feeding task, and belt scale control settings, and uses actual material feeding data, online quality inspection results, and re-optimization results as incremental samples to perform weighted updates on the raw material quality instability elimination matrix and deep learning model, forming a self-learning closed loop for subsequent batches. This enables the system to continuously adapt to changes in raw material quality, market price, and equipment operating conditions, ultimately achieving comprehensive technical effects such as improved accuracy of recycled fiber raw material proportioning, enhanced product quality stability, reduced raw material costs, increased production efficiency, and improved intelligent production management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0022] Figure 2 This is a flowchart of the method of the present invention;

[0023] Figure 3 Schematic diagram of constructing the raw material quality instability exclusion matrix and forming the candidate raw material pool;

[0024] Figure 4 This is a schematic diagram of the structure of a deep learning prediction model;

[0025] Figure 5 This is a schematic diagram of dual-constraint ratio optimization and work order decomposition.

[0026] Figure 6 A schematic diagram of the local re-optimization control process for the remaining batches;

[0027] Figure 7 A diagram illustrating the self-learning closed loop for excluding matrix and prediction model weighted updates;

[0028] Figure 8 A schematic diagram illustrating the changes in the raw material pool before and after screening the candidate raw material pool;

[0029] Figure 9 This is a schematic diagram of the change curve of the feeding deviation rate before and after the local re-optimization trigger in Example 1;

[0030] Figure 10 This is a schematic diagram comparing the cost per ton of product, quality pass rate, and batch fluctuation rate between the example and the comparative examples. Detailed Implementation

[0031] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and claims. This embodiment is illustrated using a recycled polyester staple fiber production scenario as an example; however, the present invention is not limited to recycled polyester staple fiber, but is also applicable to recycled polyamide fiber, recycled viscose fiber, recycled composite fiber, and other recycled fiber production scenarios requiring multi-raw material synergistic proportioning and online feeding control. Without departing from the concept of the present invention, any conventional substitutions, parameter adjustments, equipment adaptations, and equivalent modifications made by those skilled in the art based on this embodiment should fall within the protection scope of the present invention.

[0032] I. Terminology Explanation

[0033] 1. Regenerated fiber raw materials This refers to all input materials used to form the finished recycled fiber product, including main raw materials, auxiliary raw materials, and auxiliaries. Main raw materials can be recycled polyester bottle chips, recycled chips, recycled melt feedstock, and recycled scraps; auxiliary raw materials can be recycled staple fibers of different grades and recycled spinning waste fibers; auxiliaries can be matting agents, masterbatches, antioxidants, antistatic agents, lubricants, stabilizers, and other process additives.

[0034] 2. Historical BOM Recipes This refers to the formula structure data that has been executed and formed production records in historical production batches, including at least the planned input amount, actual input amount, proportion ratio, total batch output, target product model, and process version number of each raw material.

[0035] 3. Raw material quality instability elimination matrix It refers to an evaluation matrix constructed based on indicators such as the quality pass rate, key performance index volatility, abnormal batch ratio, supplier stability, and environmental sensitivity of each raw material within a preset statistical period. It is used to determine whether a raw material is suitable to enter the candidate raw material pool for the current batch.

[0036] 4. Candidate Raw Material Pool This refers to the set of raw materials that, after being screened using a raw material quality instability exclusion matrix, are allowed to participate in the current batch proportioning optimization solution. The raw materials in the candidate raw material pool possess quality stability, inventory availability, process compatibility, and risk acceptability.

[0037] 5. Prediction Confidence The confidence level refers to the reliability of the prediction results of a predictive model on the quality, cost, and efficiency of a candidate formulation. Prediction confidence can be directly output by the model, or it can be calculated based on the variance of multiple sampling results, the degree of divergence in the ensemble model, or uncertainty estimation results.

[0038] 6. Feeding time window This refers to dividing the material feeding process of a complete production batch into several consecutive time periods, each time period corresponding to a set of theoretical and actual material feeding quantities. In actual engineering, the material feeding time window can be set to 30 seconds, 60 seconds, 120 seconds, 180 seconds, or 300 seconds.

[0039] 7. Remaining batches undergo partial re-optimization This refers to a situation in batch production where, when the material input of an executed portion is irreversible and the current material input deviation or quality deviation exceeds a threshold, only the subsequent proportions and theoretical material input quantities for the unexecuted portion are recalculated, without canceling or rearranging the entire batch of work orders.

[0040] 8. Change of Power Update This refers to assigning different weights to samples of different time periods, confidence levels, and anomaly levels during iterative updates of the training dataset, so that the model can adapt more quickly to changes in the latest raw material quality, market prices, and equipment operating conditions.

[0041] II. System Structure

[0042] like Figure 1 As shown, the intelligent proportioning system for recycled fiber raw materials of the present invention includes at least a data acquisition module, a data preprocessing module, a deep learning prediction module, a proportioning optimization module, a work order integration module, a feeding control module, a local re-optimization module, and a model update module. Each module can be deployed on the same server or distributed across enterprise intranet servers, edge computing nodes, and industrial controllers.

[0043] 1. Data Acquisition Module

[0044] The data acquisition module is used to collect all source data required for the current batch and historical batches. Preferably, this module includes: 1) ERP interface unit, used to obtain purchase unit price, supplier information, inventory quantity, batch work order, planned output, order priority, and delivery deadline; 2) MES interface unit, used to acquire historical BOM formulas, historical feeding records, historical quality test results, equipment start-up and shutdown records, and process settings; 3) On-site environmental data acquisition unit, used to acquire workshop temperature, humidity, dust concentration, and wind speed; 4) Belt scale data acquisition unit, used to acquire instantaneous flow rate, cumulative feed amount, feed amount per window, and equipment status; 5) Market data interface unit, used to obtain real-time unit price of raw materials, price series of the past 7 to 30 days, price volatility coefficient and supply risk indicators.

[0045] 2. Data Preprocessing Module The data preprocessing module is used to handle missing values, identify outliers, align time, standardize features, construct a raw material quality instability exclusion matrix, and generate a candidate raw material pool. This module can connect to databases, data warehouses, time-series databases, and message queues.

[0046] 3. Deep Learning Prediction Module The deep learning prediction module is used to predict the quality, cost, efficiency, and confidence of candidate matching schemes. This module preferably includes a time-series encoding network, a static feature embedding network, a feature fusion layer, a multi-task output layer, and an uncertainty estimation unit.

[0047] 4. Proportion Optimization Module The proportioning optimization module is used to solve for the proportions of raw materials within the candidate raw material pool to obtain a target proportioning scheme that meets the requirements of quality, cost, efficiency, price risk, and confidence level. This module can employ multi-objective evolutionary algorithms, constrained nonlinear programming, or multi-stage optimization algorithms.

[0048] 5. Work Order Integration Module The work order integration module is used to convert the target proportioning scheme into the theoretical total amount of material to be fed and the theoretical amount of material to be fed in each time window in the current batch of work orders, and write it back to the ERP system and send it to the material feeding control module.

[0049] 6. Feeding control module The feeding control module connects to a belt scale, PLC, SCADA, or industrial control gateway to execute the theoretical feeding amount for each time window, monitor the actual feeding amount, and provide feedback on deviations.

[0050] 7. Local Re-optimization Module The local re-optimization module is used to freeze the executed portion when an error occurs, and only solves the remaining unexecuted portion, updating subsequent window control variables. This module is one of the important innovative modules of this invention.

[0051] 8. Model Update Module The model update module is used to write the newly added data of this batch into the training set and to perform variable weight updates on the raw material quality instability exclusion matrix and the deep learning prediction model, forming a reusable self-learning closed loop for the next batch.

[0052] III. Overall Technical Route for Implementing the Method of the Invention

[0053] like Figure 2 As shown, the method of the present invention is carried out according to the following technical route: Multi-source data acquisition → data cleaning and time alignment → construction of raw material quality instability exclusion matrix → formation of candidate raw material pool → deep learning prediction → dual-constraint ratio optimization → work order decomposition → material feeding execution → deviation perception → local re-optimization of remaining batches → data write-back → variable weight update. The following section provides a more in-depth and structured disclosure of the above steps.

[0054] (a) Step S1: Multi-source data acquisition

[0055] The purpose of step S1 is to establish a raw data foundation that can be used for subsequent screening, prediction and optimization.

[0056] 1. Data source and field definition

[0057] (1) ERP data Preferably, at least the following fields should be collected: raw material number; raw material category; raw material purchase price; current available inventory; supplier number; batch work order number; planned output; order priority; scheduled delivery time.

[0058] (2) MES data Ideally, at least the following fields should be collected: historical BOM formula; planned raw material input; actual raw material input; output; finished product grade; quality inspection value; equipment operating status; and production interruption records.

[0059] (3) On-site environment and equipment data Preferably, at least the following fields should be collected: temperature; humidity; dust concentration; wind speed; instantaneous flow rate of belt scale; cumulative feed amount of belt scale; motor current; control frequency; valve opening degree.

[0060] (4) Market data Ideally, at least the following fields should be collected: current market price of raw materials; recent... Price series at specific time points; price volatility; regional logistics costs; supply risk indicators. Among these, Indicates the length of the price time window, such as 7 days, 15 days, or 30 days.

[0061] 2. Acquisition frequency and synchronization strategy ERP and MES data can be synchronized in batches or incrementally every 5 or 15 minutes; environmental and equipment data are preferably sampled at 1 second, 5 seconds, or 10 seconds; market data can be synchronized by hour or day.

[0062] All data ultimately needs to be mapped to a unified timeline. Ideally, aggregation should be based on the batch start time and performed according to the feeding time window. For example, if the feeding time window is 60 seconds, the environmental data within each window can be calculated by taking the average, maximum, minimum, and standard deviation to form window-level features.

[0063] (II) Step S2: Construction of raw material quality instability exclusion matrix and candidate raw material pool

[0064] Step S2 is not about simply removing substandard raw materials, but about constructing a quantifiable, updatable pre-screening mechanism for raw material quality stability that can be used as a constraint for subsequent optimization. This mechanism transforms subsequent optimization from searching the entire raw material space to searching within a pool of stable candidate raw materials, thereby reducing the dimensionality of the solution and improving the executability and robustness of the formulation scheme.

[0065] 1. Data cleaning and feature preparation

[0066] (1) Handling missing values

[0067] For ERP structured data such as purchase price and inventory, if the missing rate is less than 5%, it can be supplemented using the mean or median of the nearest batches of the same raw material; for quality inspection values, if the missing values ​​are a small number of single points, they can be supplemented using the historical mean of the same product model and the same supplier; if the missing rate is higher than 20%, the batch sample is marked as a low confidence sample; for environmental data, linear interpolation, forward filling or moving average filling can be used.

[0068] (2) Outlier handling

[0069] The unit price, material input quantity, and quality indicators are initially screened using the three-standard-deviation method or box plot method; values ​​that are obviously due to equipment failure or input errors are removed; genuine abnormal production batches are not removed, but are retained as abnormal samples for use in calculating the proportion of abnormal batches.

[0070] (3) Standardization

[0071] Z-score standardization is applied to numerical features: ; in, These are the original eigenvalues. This is the mean of the feature in the training set. is the standard deviation of this feature in the training set.

[0072] 2. Quality Instability Index System

[0073] For each raw material, the following indicators are extracted within the preset statistical period: 1) Quality pass rate P: This represents the proportion of finished products in a batch produced using this raw material that meet the set quality standards; 2) Key performance indicator volatility W: This can be calculated separately for key indicators such as breaking strength, elongation, color difference, and impurity content, and then synthesized into a comprehensive volatility. The preferred method is: ;in, For the standard deviation of key quality indicators, This represents the average of key quality indicators.

[0074] 3) Percentage of abnormal batches E This indicates the percentage of batches produced using this raw material that were deemed abnormal. If necessary, the following can also be included: supplier switching frequency; environmental sensitivity indicators; raw material impurity fluctuation levels; and raw material moisture content fluctuation levels.

[0075] 3. Calculation of Raw Material Quality Instability Score The raw material quality instability score Z is calculated using the following formula: ; Where P is the quality pass rate, W is the volatility of key performance indicators, E is the proportion of abnormal batches, and a, b, and c are weighting coefficients, satisfying: a+b+c=1.

[0076] Weighting method: If the company emphasizes finished product quality, a=0.4, b=0.4, c=0.2 can be selected; if the company emphasizes abnormal risk control, a=0.3, b=0.3, c=0.4 can be selected.

[0077] 4. Exclusion Matrix Construction

[0078] like Figure 3 As shown, a raw material quality instability exclusion matrix is ​​constructed with raw material number as the row and evaluation index as the column. Typical fields of the matrix include: raw material number; supplier number; quality pass rate; volatility; percentage of abnormal batches; instability score; exclusion flag; and last update time. Let the instability threshold be... ,when When, the raw material is marked as excluded; when At that time, the raw material is allowed to enter the candidate raw material pool. The value can be selected based on the company's quality management level, such as 0.25, 0.30, or 0.35.

[0079] 5. Rules for the formation of candidate raw material pools

[0080] For a raw material to be included in the current batch of candidate raw material pool, it is preferred to meet all of the following conditions: 1) Not marked as excluded raw material in the exclusion matrix; 2) Current inventory meets minimum material input requirements; 3) The current supply status is available; 4) Compatible with the current product model and current process version; 5) Current market prices and price volatility levels do not exceed the risk range set by the company; 6) The raw materials are not under quality control lockout.

[0081] By employing this pre-screening method, this invention excludes unstable, scarce, and high-risk raw materials from the optimization process, creating a more refined, controllable, and executable pool of candidate raw materials. This step directly addresses the problem in existing technologies where all raw materials are included in the model, resulting in mathematically feasible optimization results but high engineering risks.

[0082] 6. Examples of Project Implementation

[0083] On a production line with an annual output of tens of thousands of tons of recycled polyester staple fiber, there were originally nine types of usable raw materials for a certain batch. After calculation using an exclusion matrix, two types were excluded due to recent large fluctuations in color difference and a high proportion of abnormal batches, and one type was removed from the candidate pool due to insufficient inventory. Ultimately, only six types were retained for subsequent calculations.

[0084] (III) Step S3: Deep Learning Prediction Model

[0085] Step S3 is used to jointly predict the quality, cost, efficiency, and confidence of different candidate formulation schemes based on the candidate raw material pool. To meet the feasibility requirements, the model structure, training method, parameter selection, and dataset usage are disclosed in a structured manner below.

[0086] 1. Sample Definition and Dataset Construction

[0087] Each training sample consists of input features and target labels.

[0088] (1) Input features

[0089] The input features include six categories: 1) Cost-related characteristics Current purchase unit price, real-time market unit price, transportation costs, and converted costs of public works.

[0090] 2) Price series characteristics recent Price series, price volatility coefficient, and trend slope at each point in time.

[0091] 3) Formula-related characteristics

[0092] The percentage of each raw material in the historical BOM, the percentage of the target candidate ratio, and the formulas of similar products.

[0093] 4) Historical quality characteristics

[0094] Historical strength, elongation, color difference value, impurity content, and pass rate.

[0095] 5) Environmental characteristics

[0096] Window-level characteristics such as temperature, humidity, dust concentration, and wind speed.

[0097] 6) Equipment characteristics

[0098] Belt scale flow rate, motor frequency, current, valve opening degree, and fault indicators.

[0099] (2) Target label

[0100] The labels should include at least: the true value of quality (Q); the true value of cost per ton (C); and the true value of production efficiency (E). If confidence level monitoring is required, the following labels may also be added: prediction error level label; batch stability label; and posterior confidence label.

[0101] (3) Dataset partitioning

[0102] Historical data from the most recent 6 to 24 months is preferred.

[0103] The dataset is split into batches to avoid data from the same batch appearing in both the training and validation sets. The preferred ratio is: Training set: 70%; Validation set: 15%; Test set: 15%.

[0104] 2. Model Structure

[0105] like Figure 4 As shown, preferred prediction models include: (1) Time series coding layer This is used to process price series, environmental time series, and device time series data. A two-layer LSTM structure is preferred. Number of hidden units in the first layer: 64; Number of hidden units in the second layer: 32; Dropout: 0.2 to 0.4.

[0106] If computing resources are limited, a single-layer GRU can also be used.

[0107] (2) Static feature embedding layer This is used to handle static features such as proportions, inventory, product models, and supplier categories. A two-layer fully connected network is preferred, with 64 and 32 hidden units per layer, and ReLU activation function is used.

[0108] (3) Feature fusion layer The time series encoding result and the static embedding result are concatenated and then fed into the fusion layer. A two-layer fully connected network is preferred, with 128 and 64 hidden units, respectively.

[0109] (4) Multi-task output layer Set four output heads: Quality prediction head, output ; Cost forecasting head, output ; Efficiency prediction head, output ; The prediction confidence output header outputs G.

[0110] 3. Loss Function Design

[0111] The preferred expression for the total loss function is: ; in, , , , These represent losses in quality, cost, efficiency, and confidence, respectively. This represents the task weight.

[0112] For tasks requiring quality, cost, and efficiency, mean squared error is preferred. ; ; ; Parameter definition: N: number of samples; The predicted value of the i-th sample; : The true value of the i-th sample.

[0113] If confidence is achieved using directly supervised learning, then Mean squared error or cross-entropy can be used; if the confidence level is derived using uncertainty, then... This can be omitted.

[0114] 4. Training methods

[0115] Preferred optimizer: Adam or AdamW; Initial learning rate: 10 -3 Or 5×10 -4 ; Batch size: 32, 64, or 128; Number of training rounds: 50 to 300; Early stopping strategy: Stop if the validation set loss does not decrease for 15 consecutive epochs; Regularization: Dropout + L2 regularization.

[0116] 5. Confidence generation method

[0117] If multiple Dropout samplings are used, then K forward computations are performed on the same input to calculate the output variance. The confidence level is defined as: ; in, G represents the standard deviation of multiple predictions. A larger G indicates higher reliability.

[0118] 6. Key Implementation Points

[0119] This invention does not require a fixed model framework, but it does require: 1) At least three predictions should be output simultaneously: quality, cost, and efficiency; 2) Able to provide confidence level or equivalent credibility index; 3) Training data must come from real historical batch data; 4) The model input should cover four categories of information: formula, price, environment, and equipment.

[0120] (iv) Step S4: Dual-constraint ratio optimization

[0121] Step S4 is another core step with high inventiveness in this invention. Unlike optimization that only considers cost or only considers quality, this invention introduces an upper limit for price fluctuation risk and a lower limit for prediction confidence in the multi-objective solution, forming a feasible region with dual constraints.

[0122] 1. Decision variables

[0123] Suppose there are n types of raw materials in the candidate raw material pool, and the decision variables are represented as follows: ; Among them, w1, w2...w n The proportion of the first to nth candidate raw materials.

[0124] 2. Multi-objective function

[0125] The synthesis objective function can be defined as follows: ; in, This is the projected cost per ton. This is a predicted quality value. This is a predicted value for production efficiency. The target weight.

[0126] In practical engineering, NSGA-II can be directly used for multi-objective non-dominated ranking, without having to hard-weight multiple objectives into a single objective value.

[0127] 3. Constraints

[0128] (1) Sum-up constraint .

[0129] (2) Upper and lower limits of process constraints .

[0130] (3) Inventory constraints

[0131] in, This represents the total material requirement for the current batch. Let represent the inventory of the i-th type of raw material.

[0132] (4) Price fluctuation risk constraints ; Where R is the price volatility risk value. Let be the price fluctuation coefficient of the i-th raw material. This represents the upper limit of risk.

[0133] (5) Prediction confidence constraints ; Where G is the prediction confidence level of the scheme. This represents the lower confidence level.

[0134] 4. Target Solution Selection

[0135] If NSGA-II is used, a set of Pareto solutions can be obtained. Enterprises can select the final solution based on preference rules, for example: If current market prices fluctuate greatly, prioritize the option with low R and high G. If higher quality requirements are currently needed for the product, then the higher quality option should be selected first. The solution; If the project is scheduled to be completed quickly, then high-quality materials should be prioritized. The solution.

[0136] The core of this step is not simply using algorithms to find the formula, but rather writing the upper limit of price fluctuation risk and the lower limit of prediction confidence into the feasible region, so that the formula is not only mathematically optimal, but also executable in terms of supply chain and model credibility.

[0137] (V) Step S5: Work order breakdown and theoretical material input generation

[0138] Step S5 is responsible for turning the optimization results into control variables that the device can execute.

[0139] Let the total amount of material fed in the current batch be... Then the theoretical total amount of raw material to be fed is: ; Parameter definition: The theoretical total amount of the i-th raw material to be fed; : The proportion of the i-th raw material; Total batch feed amount.

[0140] If the batch is divided into T time windows, then the theoretical feed amount of the i-th raw material in the t-th window is: ; in, Let be the allocation coefficient for the t-th time window, satisfying: ; Parameter definition: The theoretical feed amount of the i-th raw material in the t-th window; : Material allocation coefficient for the t-th window.

[0141] The work order integration module will and The data is simultaneously written into the ERP work order, material feeding task sheet, and belt scale control setting sheet, and assigned the same version number.

[0142] (vi) Step S6: Local re-optimization of the remaining batches

[0143] When existing systems encounter errors, they often either simply correct the current traffic or reschedule the entire batch. This invention proposes freezing the already executed portion and performing local re-optimization only on the remaining batches.

[0144] 1. Deviation detection logic

[0145] For any given time window, calculate the feeding deviation rate: ; in, This represents the actual amount of material fed into the machine. This represents the theoretical feed rate.

[0146] like This indicates that the current window's material feeding deviation exceeds the limit. Among them, The feeding deviation threshold is preferably set to 0.03, 0.05, or 0.08.

[0147] Meanwhile, online quality inspection values Compared with model predictions Comparison, when: When the time is exceeded, it indicates that the quality deviation exceeds the limit. This is the quality deviation threshold.

[0148] 2. Freeze the executed portion

[0149] Let the total amount of material fed in the current batch be... The cumulative amount of material fed into the already executed portion is The remaining amount of material to be fed is: ; This invention requires that the actual material input quantity, cumulative material input records, and corresponding work order execution traces of the executed portion be frozen, and rollback or reallocation is not allowed.

[0150] 3. Remaining batch local re-optimization model

[0151] Let the proportion of the remaining new raw materials be... Then the optimization model is only built for the remaining part: ; And satisfy: ; in, This represents the remaining available inventory of the i-th raw material at the current moment.

[0152] Its objective function can be the same as step S4, or a deviation recovery objective can be added, for example: ; in, This indicates the degree to which the remaining part compensates for the preceding deviation.

[0153] 4. Update the theoretical feed rate for subsequent windows.

[0154] If the remaining part has If there are 10 windows, then the theoretical feed amount of the i-th raw material in the t-th remaining window is: ; in, Assign coefficients to the remaining window, satisfying: ;

[0155] 5. Work order consistency write-back

[0156] The results of local re-optimization include: the target total amount of material to be fed in the remaining batches of work orders; the theoretical amount of material to be fed in the subsequent time window; and the control setpoint of the belt scale.

[0157] The above three must be updated synchronously using the same batch identifier and the same version number to avoid inconsistencies between ERP, work orders, and equipment control quantities.

[0158] (vii) Step S7: Variable weight update and self-learning closed loop

[0159] Step S7 is used to ensure that the system continues to learn and adapt to the latest state.

[0160] 1. Incremental Sample Generation

[0161] After each batch is completed, an incremental sample is automatically generated, which includes: the actual feed quantity sequence; online quality inspection results; final quality inspection results; local re-optimization records; price change records; and environmental and equipment status records.

[0162] 2. Update the raw material quality instability elimination matrix

[0163] For each raw material, recalculate: pass rate P; volatility W; percentage of abnormal batches E; instability score Z. If a raw material has recently experienced an abnormal increase, its Z value will rise, and it may be excluded from subsequent batches.

[0164] 3. Contingency Training

[0165] The total loss function is: ; in, Let j be the weight of the j-th sample. is the single-sample loss value, and m is the number of training samples in this round.

[0166] The preferred weighting strategy is: Samples from the last 7 days: weight 2.0; Samples from the last 30 days: weight 1.5; Other historical samples: weight 1.0.

[0167] 4. Update cycle Minor update: To be performed after each batch is completed; Updated offline once a day; Major update: Full training once a week or once a month.

[0168] IV. Specific Application Examples

[0169] (a) Test subjects, raw materials and equipment conditions

[0170] 1. Test subjects

[0171] The subject of this test is recycled polyester staple fiber, with a target product specification of 1.4D×38mm. The requirements for the target product include: Fracture strength: not less than 4.80 cN / dtex; Elongation at break: 28%–38%; Color difference value: not higher than 1.20; Impurity content: not higher than 0.20%; Oil uniformity deviation: not higher than 5%.

[0172] 2. Raw material composition

[0173] There are a total of 8 raw materials available, denoted as A1 to A8, as follows: A1: Primary recycled bottle flakes, sourced from supplier S1; A2: Primary recycled bottle flakes, sourced from supplier S2; A3: Secondary recycled bottle flakes, sourced from supplier S3; A4: Recycled polyester scraps, sourced from supplier S4; A5: Low melting point modified polyester material, sourced from supplier S5; A6: White masterbatch; A7: Functional additive 1; A8: Functional additive 2.

[0174] 3. Test equipment

[0175] 10L laboratory high-speed mixer; 50kg / h pilot-scale melt extrusion spinning device; loss-in-weight feeder and small belt scale linkage unit; fiber stretching and setting device; universal material testing machine; colorimeter; ash / impurity analysis device; online temperature and humidity acquisition module; industrial control computer and the formulation optimization software system of this invention.

[0176] 4. Data Foundation

[0177] This embodiment uses historical data from a company over the past 12 months, compiling a total of 936 valid production batches, including: 936 sets of historical BOM formulas; 14,320 raw material procurement and inventory data entries; 936 sets of quality inspection records; 181,440 environmental monitoring records; 268,800 belt scale and material feeding records; and 2,880 raw material market price sequences. This data is used to construct a raw material quality instability exclusion matrix, train a deep learning prediction model, and perform online validation.

[0178] (II) Testing Items and Methods

[0179] 1. Fracture strength and elongation at break According to the company's internal tensile testing methods, at a temperature... relative humidity Tested after equilibration under the specified conditions for 24 hours. Twenty monofilaments were tested for each sample, and the average value was taken.

[0180] 2. Color difference value The color difference was measured using a colorimeter, with a standard sample as a reference, to obtain the comprehensive color difference value. .

[0181] 3. Impurity content Take a 100g sample, and calculate the percentage of impurities by separation, sieving, and weighing.

[0182] 4. Raw material cost per ton of product The raw material cost C per ton of product is calculated using the following formula: ; in: The mass of the i-th raw material being fed; Let T be the unit price of the i-th raw material; T be the transportation cost of this batch; U be the converted cost of public works for this batch; and Y be the output of finished products for this batch.

[0183] 5. Feeding deviation rate

[0184] Feeding deviation rate D is calculated using the following formula: ; in: This represents the actual amount of material fed into a given time window. This represents the theoretical feed rate for that time window.

[0185] 6. Quality volatility The coefficient of variation (CV) is calculated for key performance indicators of multiple consecutive batches under the same implementation scheme to characterize batch stability. The formula is as follows: ; in: The standard deviation is the sample standard deviation. This is the sample mean.

[0186] (III) Example 1: Intelligent proportioning and local re-optimization control using the complete process of the present invention

[0187] 1. Experimental Objective

[0188] This study verifies whether the present invention can achieve the following technical effects under actual production conditions through a raw material quality instability elimination matrix + candidate raw material pool + deep learning prediction + dual constraint optimization + local re-optimization of remaining batches: 1) reducing raw material costs; 2) improving product quality pass rate; 3) reducing inter-batch fluctuations; and 4) maintaining continuous production stability when feeding deviations occur.

[0189] 2. Construction of the raw material quality instability elimination matrix

[0190] The performance of batches A1 to A8 in the most recent 90 relevant batches was statistically analyzed, and the quality pass rate P, critical performance volatility W, and abnormal batch proportion E were calculated. The instability score Z was then calculated using the following formula: Where a=0.4, b=0.35, and c=0.25. The calculation results are shown in Table 1.

[0191] Table 1: Raw Material Quality Instability Scoring Results

[0192]

[0193] In this embodiment, an instability threshold is set. As shown in Table 1, A8's instability score exceeds the threshold, therefore it is excluded and will not be included in the candidate raw material pool for the current batch. Figure 3 and Figure 8 As shown, this invention does not directly solve for all 8 raw materials, but first selects 7 candidate raw materials to enter the subsequent modeling and optimization stage.

[0194] 3. Deep Learning Prediction Model and Parameter Settings

[0195] (1) Model structure

[0196] This embodiment uses an LSTM + static feature fully connected + multi-task output model: Time series inputs: raw material price series, environmental series, equipment status series; Static feature inputs: raw material candidate proportion, inventory, supplier category, product model; Two-layer LSTM: 64 and 32 hidden units respectively; Two-layer fused fully connected layer: 128 and 64 hidden units respectively; Four output heads: each outputs the predicted quality value. Cost forecast Efficiency prediction value And prediction confidence G.

[0197] (2) Training parameters Training set / validation set / test set partitioning: 70% / 15% / 15%; Optimizer: Adam; Initial learning rate: 5×10 -5 ; Batch size: 64; Maximum number of training rounds: 180; Early stopping rounds: 15; Dropout: 0.3.

[0198] (3) Loss function

[0199] The total loss function is: ; The mean squared error is used for each sub-loss.

[0200] (4) Model effect

[0201] On the test set, the model achieved the following prediction accuracies: 1) Root mean square error of quality prediction: 0.083; 2) Mean absolute percentage error of cost prediction: 2.74%; 3) Mean absolute percentage error of efficiency prediction: 3.12%; 4) Correlation coefficient between confidence output and actual prediction error: 0.81. This indicates that the model can be used for subsequent allocation optimization decisions.

[0202] 4. Dual-constraint ratio optimization

[0203] The candidate raw material pool is A1 to A7. The optimization objectives are: minimum cost per ton; maximum comprehensive quality indicators of strength and overall color difference; and maximum production line stability and efficiency.

[0204] The constraints include: the sum of the proportions of each raw material is 1; each raw material meets the upper and lower limits of the process; inventory meets the current requirement of 10 tons per batch; and price fluctuation risk value. Prediction confidence level .

[0205] Price volatility risk is calculated using the following formula: ;in, As a percentage of raw materials, This represents the raw material price fluctuation coefficient. The final target formulation is shown in Table 2.

[0206] Table 2: Target Proportioning Scheme of Example 1

[0207]

[0208] Based on the planned total material input of 10.45 tons (considering normal process losses), the work order is broken down and divided into 120 material input time windows, each window being 1 minute long.

[0209] 5. Online material feeding and local re-optimization

[0210] When the feeding process reached the 51st time window, the actual feeding amount of A3 was too high, resulting in a window deviation rate of: ; That is, 6.7%, exceeding the set threshold of 5%. At the same time, online quality inspection found that the melt uniformity deviated from the model prediction range, thus triggering local re-optimization of the remaining batches.

[0211] In the partial re-optimization: the actual material input for the first 51 windows is frozen; only the theoretical material input for the subsequent 69 windows is recalculated; after re-optimization, the remaining stage proportions for A3 and A5 are slightly reduced, while the remaining stage proportions for A1 and A2 are slightly increased; the new control values ​​are synchronously written back to the remaining execution work orders and belt scale settings in the ERP system. The changes in material input deviation rate before and after re-optimization are shown below. Figure 9As shown, the maximum deviation rate before triggering was 6.7%, and after re-optimization, the subsequent window deviation rate stabilized between 1.2% and 2.4%.

[0212] 6. Test Results

[0213] This embodiment completed 6 parallel batches consecutively, and the results are shown in Table 3.

[0214] Table 3: Results of continuous batch testing in Example 1

[0215]

[0216] The statistical results are as follows: average cost per ton: 7311.7 yuan / t; quality pass rate: 100%; average breaking strength: 4.96 cN / dtex; coefficient of variation (CV) for breaking strength: 0.47%; average color difference: 0.89; coefficient of variation (CV) for color difference: 2.11%. This demonstrates that the present invention can achieve high-quality and stable production at a relatively low cost.

[0217] (iv) Example 2: Application example of the present invention without triggering local re-optimization under stable operating conditions

[0218] To verify the optimization capability of the present invention under normal operating conditions, the same model, candidate raw material pool and dual constraint conditions as in Example 1 were used. However, no feeding deviation exceeded the limit during this production process, so local re-optimization was not triggered.

[0219] The target ratio is as follows: A1: 0.22 A2: 0.18 A3: 0.17 A4: 0.13 A5: 0.12 A6: 0.10 A7: 0.08 Four batches were produced consecutively, and the results are shown in Table 4.

[0220] Table 4: Detection Results of Example 2

[0221]

[0222] Statistical results: Average cost per ton: 7349.8 yuan / t; Quality pass rate: 100%; Average fracture strength: 4.905 cN / dtex; Coefficient of variation (CV): 0.29%. This indicates that even under normal operating conditions without triggering re-optimization, this invention, relying on candidate pool screening and dual-constraint optimization, can achieve a good balance between cost and quality.

[0223] (v) Comparative Example 1: Using empirical fixed proportions, without using exclusion matrices, predictive models, or re-optimization.

[0224] 1. Proportioning method

[0225] Comparative Example 1 uses the company's existing experience in formulating the product, with the formula fixed as follows: A1: 0.18 A2: 0.16 A3: 0.18 A4: 0.16 A5: 0.14 A6: 0.10 A7: 0.04 A8: 0.04 The proposed plan does not exclude A8, nor does it consider price volatility risk or forecast confidence level.

[0226] 2. Production Results

[0227] Six batches were produced consecutively, and the results are shown in Table 5.

[0228] Table 5: Detection Results of Comparative Example 1

[0229]

[0230] Statistical results: Average cost per ton: 7526.2 yuan / t; Quality pass rate: 50%; Average breaking strength: 4.802 cN / dtex; Coefficient of variation (CV): 0.66%; Average color difference: 1.132. Compared with Example 1: the cost per ton is 214.5 yuan / t higher; the quality pass rate is 50 percentage points lower; the color difference is greater; and the impurity content fluctuates more. This indicates that relying solely on experience to fix the proportions is insufficient to balance cost, quality, and stability.

[0231] (vi) Comparative Example 2: Using deep learning for prediction and optimization, but without building an exclusion matrix and without performing local re-optimization.

[0232] 1. Purpose of comparison

[0233] Verify the contribution of the two key technical features in this invention: raw material quality instability elimination matrix and local re-optimization of remaining batches.

[0234] 2. Implementation Method

[0235] In Comparative Example 2: the same historical data and deep learning model as in Example 1 were used; all A1 to A8 were allowed to enter the optimization; cost, quality and efficiency were jointly optimized; but no exclusion matrix was constructed and no candidate raw material pool was set up; if a deviation occurred during the feeding process, only a small correction was made to the current flow rate, and the remaining batches of the executed part were not frozen for local re-optimization.

[0236] 3. Results

[0237] Six batches were produced consecutively, and the results are shown in Table 6.

[0238] Table 6: Detection Results of Comparative Example 2

[0239]

[0240] Statistical results: Average cost per ton: 7410.8 yuan / t; Quality pass rate: 66.7%; Average maximum feed deviation rate: 6.35%. Compared with Example 1: The cost per ton is 99.1 yuan / t higher; the pass rate is 33.3 percentage points lower; and the maximum feed deviation rate is significantly higher. This indicates that relying solely on prediction and optimization without pre-screening of the exclusion matrix and local re-optimization results in a significantly inferior actual effect compared to the complete solution of this invention.

[0241] (vii) Comparative Example 3: Using an exclusion matrix and optimization, but without setting an upper limit for price risk and a lower limit for confidence.

[0242] To verify the effectiveness of the dual constraints, a comparative example 3 was set up. This comparative example uses an exclusion matrix to filter the candidate pool and also uses deep learning prediction and multi-objective optimization, but does not include a price volatility risk cap. and lower bound of prediction confidence Four batches were produced consecutively, and the results are shown in Table 7.

[0243] Table 7: Detection Results of Comparative Example 3

[0244]

[0245] While this comparative example appears to have a lower cost per ton, it reveals several low-cost but unreliable solutions, leading to a significant increase in the risk of quality mismatch. This invention is designed to address this issue. and The dual constraints are not redundant, but rather key conditions for ensuring industrial feasibility and quality stability.

[0246] (viii) Comprehensive Results Analysis

[0247] 1. Cost vs. Quality Comparison

[0248] The core results of Examples 1, 2 and each comparative example are summarized in Table 8.

[0249] Table 8: Comprehensive Comparison of Various Implementation Plans

[0250]

[0251] 2. Technical effectiveness proof

[0252] Based on the above experimental data, it can be proven that the present invention has at least the following technical effects:

[0253] (1) Reduce the cost per ton of product Example 1 showed an average reduction of 214.5 yuan / t compared to Comparative Example 1, and an average reduction of 99.1 yuan / t compared to Comparative Example 2.

[0254] (2) Improve product qualification rate Example 1 had all six batches pass the test, while Comparative Example 1 had only 50% and Comparative Example 2 had 66.7%.

[0255] (3) Reduce batch fluctuations Example 1 shows that the coefficient of variation (CV) of fracture strength is 0.47%, which is significantly lower than that of empirically formulated methods; the color difference value is also smaller, and the impurity content is more stable.

[0256] (4) Improve continuous production capacity under abnormal operating conditions In Example 1, when a 6.7% feeding deviation occurred, the deviation rate in the subsequent window was reduced to 1.2%–2.4% by freezing the already executed portion and locally re-optimizing the remaining batches. Although Comparative Example 2 underwent ordinary correction, the maximum deviation rate remained above 5.7%, indicating that the local re-optimization mechanism of this invention has significant advantages.

[0257] (5) Improve industrial feasibility Comparative Example 3 illustrates that simply pursuing low cost can easily lead to the selection of low-reliability solutions, resulting in quality mismatch. This invention, by introducing a dual constraint of a price risk ceiling and a confidence level floor, can improve execution robustness while maintaining cost competitiveness.

[0258] (ix) Explanation in conjunction with the accompanying drawings

[0259] like Figure 3 and Figure 8 As can be seen, this invention first screens out high-risk raw materials using a raw material quality instability exclusion matrix, and then constructs a candidate raw material pool, thereby narrowing the optimization space and reducing the probability of inferior raw materials entering the scheme. For example... Figure 4 It is evident that the deep learning model employed in this invention is not a simple cost prediction model, but a multi-task model that simultaneously processes price series, historical formulas, environmental and equipment data. It can simultaneously output quality, cost, efficiency, and confidence levels, providing input for subsequent dual-constraint optimization. For example... Figure 5 As can be seen, the target proportioning scheme is further broken down into the total batch feed amount and the theoretical feed amount for each time window, and synchronously linked with ERP work orders and belt scale control values ​​to ensure the scheme's feasibility. For example... Figure 6 and Figure 9 As can be seen, when the material feeding deviation exceeds the limit, this invention freezes the already executed portion and only performs local re-optimization on the remaining batches. Therefore, it can correct subsequent control quantities without interrupting the executed data. Figure 7As can be seen, in the embodiment, the actual material input data, online quality data, and re-optimization results generated in each batch are fed back into the model update module, forming a self-learning closed loop for continuous optimization. For example... Figure 10 It can be seen that the present invention is superior to the main comparative example in terms of cost, pass rate and stability, which proves the effectiveness of the overall technical solution of the present invention.

[0260] The above application examples and experimental data demonstrate that this invention does not merely propose an intelligent proportioning approach at the theoretical level, but rather achieves quantifiable and verifiable technical effects under laboratory / pilot-scale conditions through a comprehensive set of technical measures, including a raw material quality instability exclusion matrix, a candidate raw material pool, deep learning joint prediction, dual-constraint optimization, work order decomposition, online deviation sensing, local re-optimization of remaining batches, and variable-weight update closed loop. Specifically, this manifests in: 1) significantly reducing the cost per ton of product; 2) significantly improving the quality pass rate; 3) significantly reducing batch fluctuations; 4) significantly enhancing online control capabilities under deviation conditions; and 5) forming a sustainable iterative self-learning optimization mechanism.

[0261] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0262] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0263] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0264] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0265] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0266] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0267] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0268] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for intelligent proportioning of recycled fiber raw materials based on deep learning technology, characterized in that, Includes the following steps: S1. Collects multi-source data from ERP, MES, field acquisition devices, and external data interfaces; S2. The multi-source data is cleaned, missing data is filled in, time is aligned and features are standardized, and a raw material quality instability exclusion matrix is ​​constructed based on historical quality detection results. According to the raw material quality instability exclusion matrix, raw materials whose quality instability scores exceed the instability threshold within a preset statistical period are excluded to form a candidate raw material pool for the current production batch. S3. Input the cost characteristics, price fluctuation characteristics, historical BOM formula characteristics, historical quality characteristics, environmental characteristics, and equipment feeding characteristics corresponding to the candidate raw material pool into the deep learning prediction model, and output the quality prediction value, cost prediction value per ton of product, production efficiency prediction value, and prediction confidence level corresponding to each candidate ratio scheme. S4. Based on the output results, construct a ratio optimization model with dual constraints. Take the ratio of each candidate raw material as the decision variable, and take the minimum cost per ton of product, the maximum quality prediction value, and the maximum production efficiency prediction value as the joint optimization objectives. At the same time, take the upper limit of price fluctuation risk and the lower limit of prediction confidence as the feasible region constraints, and solve to obtain the target ratio scheme for the current production batch. S5. Decompose the target proportioning scheme into the theoretical total amount of each raw material to be fed in the current production batch, and further decompose it into the theoretical amount of each feeding time window. Write the decomposition results back to the current batch work order and feeding task in the ERP system, and send them to the belt scale control system at the same time.

2. The method according to claim 1, characterized in that, In step S1, the multi-source data includes at least: raw material purchase unit price, supplier information, inventory and batch work order information obtained from the ERP system; historical BOM formula, historical output and historical quality test results obtained from the MES system; temperature, humidity, dust concentration, real-time feed amount of belt scale and equipment operating status data obtained from the field acquisition device; and real-time unit price and price fluctuation data of raw materials in the market obtained from the external data interface. And / or, the method for constructing the raw material quality instability exclusion matrix in step S2 is as follows: For each raw material, the quality pass rate, key performance indicator volatility, and abnormal batch proportion of the corresponding batch during production within a preset statistical period are statistically analyzed, and a raw material quality instability score Z is formed; when Z is greater than the preset instability threshold Z0, the raw material is marked as an excluded raw material and is not included in the candidate raw material pool construction for the current production batch; wherein: ； Where P is the quality pass rate, W is the volatility of key performance indicators, E is the proportion of abnormal batches, and a, b, and c are weighting coefficients, satisfying: a+b+c=1.

3. The method according to claim 1, characterized in that, In step S3, the deep learning prediction model includes a time series encoding layer, a feature fusion layer, and a multi-task output layer. The time series encoding layer is used to extract the time-dependent features of the purchase unit price, the real-time market unit price, and price fluctuation data. The feature fusion layer is used to fuse historical BOM formula features, historical quality features, environmental features, and equipment feeding features. The multi-task output layer simultaneously outputs the quality prediction value, the cost per ton of product prediction value, the production efficiency prediction value, and the prediction confidence level.

4. The method according to claim 1, characterized in that, In step S3, the predicted cost per ton of product C is calculated according to the following formula: ;in, Let be the mass ratio of the i-th raw material. Let T be the unit price of the i-th raw material, T be the transportation cost of the current batch, U be the converted cost of utilities for the current batch, and Y be the output of the finished product for the current batch. And / or, in step S4, the price volatility risk value R is calculated according to the following formula: ； in, Let represent the proportion of the i-th raw material. Let be the price fluctuation coefficient of the i-th raw material within a preset time window; And / or, in step S4, the proportioning optimization model with dual constraints further includes the following constraints: The sum of the proportions of each candidate raw material is 1; The proportion of each candidate raw material shall not be lower than the lower limit of the corresponding process proportion and not higher than the upper limit of the corresponding process proportion; The planned total amount of each candidate raw material shall not exceed the available inventory of the corresponding raw material; Price volatility risk value R shall not exceed the preset risk limit. ; The prediction confidence level G is not lower than the preset confidence lower limit. .

5. The method according to claim 1, characterized in that, The method also includes: S6. During production execution, the actual feeding amount, cumulative feeding deviation, and online quality inspection results within each feeding time window are obtained in real time. When the deviation between the actual feeding amount and the theoretical feeding amount exceeds the feeding deviation threshold, or the deviation between the online quality inspection result and the quality prediction value exceeds the quality deviation threshold, the actual feeding amount of the executed part in the current production batch is frozen, and a local re-optimization model for the remaining batch is constructed only for the unfinished part. The proportion of the remaining raw materials and the theoretical feeding amount for each subsequent feeding time window are recalculated. Feeding deviation rate D is calculated according to the following formula: ;in, This represents the actual amount of material fed into the current feeding window. This represents the theoretical feed rate for the current feeding time window; S7. Synchronously write back the remaining batch local re-optimization results obtained in step S6 to the remaining batch work orders, feeding tasks and belt scale control settings in the ERP system, and add the actual feeding amount, online quality inspection results and local re-optimization results of the current production batch as incremental samples to the training dataset, and perform variable weight updates on the raw material quality instability exclusion matrix and deep learning prediction model to form a self-learning closed loop for subsequent batches.

6. The method according to claim 5, characterized in that, The method for establishing the local re-optimization model for the remaining batch in step S6 is as follows: the cumulative actual amount of each raw material in the executed part of the current production batch is kept unchanged, the remaining raw material demand corresponding to the remaining planned output of the current production batch is taken as the optimization object, and only the theoretical amount of material input for the subsequent unexecuted material input time window is recalculated to avoid the overall cancellation and rearrangement of executed work orders. And / or, the weighted update in step S7 includes: assigning higher sample weights to the current batch of incremental samples than to historical samples, in order to improve the responsiveness of the raw material quality instability exclusion matrix and the deep learning prediction model to the latest changes in raw material quality, the latest changes in market price, and the latest changes in equipment status; its training loss function Represented as: ； in, Let j be the sample weight of the j-th sample. Let be the single-sample loss value corresponding to the j-th sample, and m be the total number of samples participating in this iteration of training.

7. A smart proportioning system for recycled fiber raw materials for implementing the method according to any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to collect multi-source data from ERP systems, MES systems, field acquisition devices, and external data interfaces; The data preprocessing module is used to clean, fill in missing data, align time and standardize features of the multi-source data, and to construct a raw material quality instability exclusion matrix and a candidate raw material pool. The deep learning prediction module is used to output the quality prediction value, cost per ton of product prediction value, production efficiency prediction value and prediction confidence value for each candidate ratio scheme based on the candidate raw material pool and the processed multi-source feature data. The proportioning optimization module is used to solve the target proportioning scheme based on the predicted quality value, the predicted cost per ton of product, the predicted production efficiency value, the prediction confidence level, and the constraints. The work order integration module is used to decompose the target proportioning scheme and synchronously write it back to the current batch of work orders and material feeding tasks in the ERP system; The feeding control module is used to control the belt scale to feed materials according to the theoretical feeding amount and to collect the actual feeding amount; The local re-optimization module is used to freeze the actual feed amount of the executed part when the deviation exceeds the threshold, and only recalculates the remaining raw material ratio and subsequent theoretical feed amount for the part that has not yet been completed in the current production batch. The model update module is used to update the raw material quality instability exclusion matrix and the deep learning prediction model by taking the actual feed amount, online quality detection results and local re-optimization results as incremental samples.

8. The system according to claim 7, characterized in that, The data acquisition module includes an ERP interface unit, a MES interface unit, an environmental acquisition unit, a belt scale acquisition unit, and a market data interface unit. The ERP interface unit is used to acquire purchase price, supplier information, inventory levels, and batch work order information; the MES interface unit is used to acquire historical BOM formulas, historical output, and historical quality test results; the environmental acquisition unit is used to acquire temperature, humidity, and dust concentration; the belt scale acquisition unit is used to acquire real-time feed rates; and the market data interface unit is used to acquire real-time raw material market unit prices and price fluctuation data. And / or, a consistent control relationship is established between the work order integration module and the material feeding control module for the remaining execution segments of the same batch. The local re-optimization results of the remaining batch output by the local re-optimization module include: the target total material feeding amount in the remaining batch work order, the theoretical material feeding amount in the subsequent material feeding time window, and the belt scale control setting value, and the three are updated synchronously using the same batch identifier or the same version identifier. And / or, the model update module is connected to both the data preprocessing module and the deep learning prediction module, and is used to: update the raw material quality instability exclusion matrix based on the online quality inspection results of the current batch; update the deep learning prediction model parameters based on the actual feed amount of the current batch, the online quality inspection results, and the local re-optimization results; and use the updated raw material quality instability exclusion matrix and deep learning prediction model parameters for the next batch of candidate raw material pool screening and target ratio scheme solution.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6.

Images