A multi-parameter process optimization method and control system for cotton-like functional filament production lines

Abstract

This invention discloses a multi-parameter process optimization method and control system for a cotton-like functional filament production line. The method acquires raw material batch characteristics and online operating condition data, constructs a mechanistic model, introduces a learning model to compensate for prediction deviations, and forms a digital twin model. Based on historical stable batch errors combined with online error rolling updates, an upper bound for quality prediction uncertainty is obtained. Lexicographically robust solutions are performed under process constraints and interlocking constraints. Dual time-scale control is used to achieve parameter adjustment and rate-of-change limitation. When the risk reaches 90% of the upper limit or the overall quality falls below the lower limit three times consecutively, restricted disturbance identification is triggered, the feasible process range is updated, interlocks are tightened, the disabled parameter combination library is maintained, and the setpoints are written back. This scheme improves production stability and batch consistency, and reduces filament breakage and energy consumption.

Description

Technical Field

[0001] This invention belongs to the field of textile technology, specifically relating to a multi-parameter process optimization method and control system for cotton-like functional filament production lines. Background Technology

[0002] The production of synthetic fiber filaments is a continuous and strongly coupled process industry, typically including melt extrusion, spinning and side-blowing cooling, drafting and orientation, web processing, false-twist hot box setting, oiling, and winding. To achieve a cotton-like feel and moisture-wicking properties, the production line needs to achieve multi-parameter coordinated control within a narrow process window, and each stage exhibits significant thermal-tension coupling and time-varying hysteresis characteristics. Simultaneously, batch variations in raw materials, such as viscosity / melt index, moisture content, masterbatch ratio, and storage time, as well as environmental fluctuations, such as side-blowing air temperature and humidity, and workshop temperature and humidity, can lead to drift in quality indicators and filament breakage risk, resulting in increased energy consumption, frequent filament breakage, and decreased batch consistency.

[0003] In existing technologies, some literature has proposed using data-driven models for intelligent optimization design of spinning processes. Chinese patent CN101782771A discloses an intelligent optimization design method for spinning processes based on immune neural networks. This method uses neural network modeling and combines it with an expert system to uniformly configure production line parameters. Furthermore, it can synchronously correct the model based on real-time production data, thereby achieving optimized control of the production process. However, such solutions typically focus on a general framework of model fitting, parameter recommendation, and online correction, which still falls short in addressing key issues in continuous filament production: First, they lack a structured fusion mechanism that explicitly incorporates mechanistic constraints (heat transfer, tension coupling, cooling and solidification, drawing, orientation, and shaping, etc.) into optimization and control, making it prone to model extrapolation distortion; second, they lack an auditable upper bound characterization of the uncertainty of prediction errors, making it difficult to ensure that quality and safety boundaries are met even in the worst-case scenario when batch drift or sudden changes in operating conditions occur; third, multi-objective objective functions and constraint systems often do not prioritize "filament breakage risk" as a hard constraint, and they also lack a closed-loop execution mechanism that coordinates with interlocking protection, which may cause tension fluctuations, filament breakage, or frequent rollbacks in actual execution due to parameter optimization results.

[0004] On the other hand, digital twin technology has also been introduced into some stages of chemical fiber filament production to improve prediction and scheduling capabilities. For example, CN111924659B discloses a method for unwinding chemical fiber filaments based on a twin model. By constructing a digital twin model of the winding production line to predict parameters at future moments, and combining it with a path generation algorithm, an optimized unwinding path and unwinding scheme are obtained to control the unwinding operation of the production line. This type of technology helps to improve the efficiency of unwinding operations and the level of collaborative scheduling, but its focus is mainly on the scheduling and path planning of the winding end operation / logistics. It does not provide sufficient support for the full-process control of maintaining stable multi-parameter process windows, co-optimizing energy consumption and quality, and suppressing the risk of filament breakage required for the production of cotton-like functional filaments. Moreover, it usually does not provide robust optimization strategies for quality indicators, dual-time-scale rolling adjustment strategies, and trigger-based safety exploration and constraint self-updating mechanisms, making it difficult to adapt to the comprehensive requirements of cotton-like functional filaments for batch consistency, risk boundaries, and energy consumption constraints.

[0005] Therefore, there is an urgent need for a multi-parameter process optimization method and control system for cotton-like functional filament production lines. This system should be able to construct a suitable model based on raw material batches and online operating conditions, achieve multi-objective robust optimization, and take into account the overall quality of cotton-like products, energy consumption, and the risk of filament breakage. This would improve the stability of continuous production and batch consistency, and reduce filament breakage and energy consumption. Summary of the Invention

[0006] To address the aforementioned technical challenges, this paper presents an online adaptive multi-parameter process optimization and control method and system under conditions of raw material batch variations and environmental disturbances. This method aims to steadily improve the overall quality of cotton-like functional filaments and reduce unit energy consumption while meeting the risk boundary of filament breakage and the lower limit of quality.

[0007] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution.

[0008] This invention provides a multi-parameter process optimization method for a cotton-like functional filament production line, comprising the following steps:

[0009] S1 acquires raw material batch characteristics and online operating condition data;

[0010] S2 establishes a mechanistic model that includes the coupling relationship between heat transfer and tension, and establishes a learning model to compensate for the prediction bias of the mechanistic model. The two are then integrated to form a digital twin model. Based on the quantile statistics of historical stable batch prediction errors and combined with online error rolling updates, the upper bound of the uncertainty of quality prediction is obtained.

[0011] S3 performs lexicographically robust solution under process constraints and interlocking constraints: considering the worst case of the upper bound of the uncertainty of the quality prediction, first ensure that the risk of broken yarn does not exceed the risk upper limit, then ensure that the comprehensive quality evaluation index of imitation cotton is not lower than the quality lower limit, on this basis, prioritize minimizing the energy consumption per unit output, then improve the comprehensive quality evaluation index of imitation cotton, and finally minimize the setting change to obtain the optimal process setting value.

[0012] S4 employs dual time-scale control, with the slow cycle updating the optimal process setpoint and the fast cycle using rolling predictive control and applying a rate-of-change limit, so that key process parameters track the optimal process setpoint without triggering interlocking constraints.

[0013] When the risk of broken yarn is greater than or equal to 90% of the upper limit of risk or the comprehensive quality evaluation index of imitation cotton is lower than the lower limit of quality for three consecutive times, S5 will trigger restricted disturbance identification and narrow the feasible process range, tighten interlock constraints and maintain the library of prohibited parameter combinations, and write back the updated optimal process setting value.

[0014] Specifically, the raw material batch characteristics include at least two of the following: slice viscosity or melt index grade, slice moisture content, masterbatch or additive ratio, and batch storage time; the online operating condition data includes at least two of the following: temperature, humidity, speed, tension, and pressure.

[0015] Specifically, the process parameters corresponding to the process constraints include at least melt temperature, side-blowing temperature, side-blowing relative humidity, draw ratio, guide roller speed group, false twist hot box temperature, oiling rate, network pressure and winding tension; the guide roller speed group includes at least the first draw roller speed, the second draw roller speed, the false twist input roller speed, the false twist output roller speed and the winding speed.

[0016] Specifically, the quality indicators used to calculate the comprehensive quality evaluation index of the imitation cotton include at least linear density fluctuation index, breaking strength, breaking elongation, evenness index and / or hairiness index, as well as moisture absorption and wicking index; the comprehensive quality evaluation index of the imitation cotton is obtained by weighting the above quality indicators after they are converted to the same dimension, and the weights are determined by the product specification library or customer indicator library and are allowed to be updated adaptively based on historical stable batch statistics.

[0017] Specifically, the mechanism model includes at least the effects of side-blowing temperature and humidity on the cooling and curing rate, the effects of draw ratio and guide roller speed group on orientation crystallization, and the effects of false twist hot box temperature on curling and shaping, and maps the effects to the prediction terms of linear density fluctuation, evenness or fuzz, and moisture absorption and conduction index; and / or, the learning model uses the difference between the quality index predicted by the mechanism model and the quality index measured online as the learning object, and adopts a recursive update or sliding window update method to ensure that the compensated error meets the error upper limit corresponding to each quality index within a preset time window.

[0018] Specifically, the upper bound of the quality prediction uncertainty is determined by a preset quantile of the historical stable batch prediction error distribution, and adaptively increases with the increase of online error, so as to ensure that the lexicographical robust solution still satisfies the lower quality limit and the upper risk limit in the worst case; and / or, the wire breakage risk is obtained by fusing the winding tension anomaly factor and the process deviation anomaly factor; wherein the winding tension anomaly factor is composed of at least the tension fluctuation amplitude, the tension peak value and the peak value duration, and the process deviation anomaly factor is composed of at least two of the following: melt temperature deviation, side blowing temperature and humidity deviation, network pressure deviation, oiling rate deviation and false twist hot box temperature deviation.

[0019] Specifically, the process constraints include the equipment's allowable range and stable operating range for each process parameter, wherein the stable operating range is obtained from historical stable batch statistics; the lexicographical robust solution prioritizes searching within the stable operating range, and expands the search within the equipment's allowable range only when the lower quality limit and upper risk limit are met; and / or, the interlocking constraints include at least one or a combination of the following rules: when the winding tension exceeds the upper tension limit or the tension fluctuation amplitude exceeds the upper fluctuation limit, increasing the winding speed is prohibited and the draw ratio is limited; when the risk of wire breakage exceeds the upper risk limit or reaches a preset proportion of the upper risk limit, the winding speed and / or draw ratio are forcibly reduced, and the network pressure is adjusted to the wire breakage prevention safety range; when the melt temperature deviates by more than a preset deviation, the rate of change of the guide roller speed group and the draw ratio is limited to meet the rate of change limit.

[0020] Specifically, in the dual time-scale control, the slow cycle is 5 to 30 minutes, used to update at least one of the optimal process setting, the lower quality limit, and the upper risk limit; the fast cycle is 0.5 to 5 seconds, used for the rolling predictive control output execution quantity; the execution quantity includes at least one of the following: heating power setting, side blowing air volume or temperature and humidity setting, guide roller speed setting, draw ratio setting, network pressure setting, and winding tension setting, and a change rate limit is applied to the execution quantity.

[0021] Specifically, when the risk of broken yarn reaches 90% or more of the risk limit or the comprehensive quality evaluation index of imitation cotton is lower than the quality limit for three consecutive times, the corresponding parameter group is written into the disabled parameter combination library, the feasible process range is narrowed, and the interlocking constraints are tightened; when the risk of broken yarn is lower than the risk limit for five consecutive slow cycles and the comprehensive quality evaluation index of imitation cotton is not lower than the quality limit, the feasible process range is expanded in a limited way and the updated optimal process setting value is written back.

[0022] Furthermore, the present invention provides a multi-parameter process optimization control system for a cotton-like functional filament production line, comprising the following modules:

[0023] The data acquisition module is used to collect raw material batch characteristics, online operating condition data, process parameters and quality indicators; the digital twin module is used to establish a mechanism model and compensate for prediction deviations through the learning model to form a digital twin model, and output the upper bound of the uncertainty of quality prediction.

[0024] The optimization decision module is used to perform lexicographically robust solving and output the optimal process setpoints under process constraints and interlocking constraints.

[0025] The dual-timescale control module is used to update the optimal process setpoint in a slow cycle and to perform rolling predictive control and apply a rate-of-change limit in a fast cycle.

[0026] The control execution module is used to drive the heating, side blowing, drafting, false twisting, networking and winding actuators; the safety exploration and constraint self-updating module is used to trigger restricted disturbance identification, update the feasible process range, tighten interlock constraints and maintain the prohibited parameter combination library when the risk of yarn breakage reaches 90% or more of the risk limit or the comprehensive quality evaluation index of imitation cotton is lower than the quality lower limit three times in a row, and write the updated optimal process setting value back to the dual time scale control module.

[0027] This invention addresses the highly coupled and multi-perturbation characteristics of continuous production of cotton-like functional filaments. Its core principles are data acquisition, digital twin, robust optimization, dual-timescale closed-loop, and self-updating of safety constraints. First, it collects raw material batch characteristics and online operating condition data to establish a mechanistic model incorporating the coupling relationship between heat transfer and tension. A learning model is then introduced to compensate for residual errors in mechanistic predictions, forming a digital twin model that can be corrected online. Further, based on historical stable batch error quantile statistics and combined with rolling updates of online errors, an upper bound on the uncertainty of quality prediction is constructed, ensuring that optimization remains verifiable even in the worst-case scenario. Subsequently, under process constraints and interlocking constraints… Lexicographical robust solution is adopted to prioritize ensuring that the risk of yarn breakage does not exceed the risk limit and that the comprehensive quality evaluation index of cotton imitation does not fall below the quality limit. On this basis, the energy consumption per unit output, quality, and set variation are optimized. Finally, the optimal process set value is updated in a slow cycle and the rolling predictive control is applied in a fast cycle with a limit on the rate of change to achieve stable tracking of key parameters. When the risk of yarn breakage reaches 90% or more of the risk limit or the quality falls below the lower limit three times in a row, the restricted disturbance identification, feasible process range shrinkage, interlock tightening, and maintenance of the disabled parameter combination library are triggered to achieve adaptive updating of the safety boundary, thereby forming a feasible closed-loop self-optimizing control system.

[0028] Compared to existing parameter recommendations based on pure data fitting or digital twin or centralized control schemes that only address local aspects, this invention uses a digital twin fusion model that integrates mechanistic constraints and residual learning. It introduces an upper bound on quality prediction uncertainty and robust worst-case verification, ensuring that optimization results consistently meet risk upper limits and quality lower limits even under batch drift and environmental fluctuations, thus reducing the probability of yarn breakage and instability from the source. Lexicographical robust solution prioritizes yarn breakage risk and quality lower limits as hard constraints, minimizing unit output energy consumption and suppressing large fluctuations in settings while ensuring safety and quality, achieving a balance between energy consumption reduction and stable operation. Dual-timescale rolling prediction closed-loop and rate-of-change limiting improve the tracking capability of the thermal-tension coupled hysteresis system, reducing tension shocks and frequent interlocking backoffs. Furthermore, through restricted disturbance identification, feasible region self-contraction, and a disabled parameter combination library mechanism, it can quickly isolate high-risk parameter combinations and adaptively update interlocking strength, significantly improving the continuous stability, batch consistency, and overall quality of finished products in cotton-like functional filament production, while reducing yarn breakage downtime, waste yarn rate, and parameter adjustment dependence. Attached Figure Description

[0029] Figure 1 System structure block diagram of the present invention;

[0030] Figure 2 The method flowchart of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0032] I. Definition of Terms

[0033] To enable those skilled in the art to implement this invention without creative effort, key terms, indicators, and controlled objects are first explained.

[0034] 1. Cotton-like functional filament production line

[0035] The cotton-like functional filament production line refers to a production system that uses polyester or modified polyester as the main raw materials and produces continuous filaments with cotton-like feel (soft, fluffy, moisture-wicking, cotton-like touch) and functional properties (moisture-wicking, hydrophilic, antistatic, antibacterial, etc.) through continuous processes such as melt extrusion, spinning, cooling and curing, stretching and orientation, web treatment, false twisting hot box setting, oiling, and winding.

[0036] 2. Raw material batch characteristics

[0037] Raw material batch characteristics refer to a set of attributes that are directly related to batch differences and significantly affect spinning stability and filament quality. These may include, but are not limited to, chip viscosity grade or melt index grade, chip moisture content, masterbatch / additive ratio, batch storage time, chip crystallinity, and end carboxyl group content.

[0038] 3. Online operating condition data

[0039] Online operating condition data refers to signals reflecting the status of equipment and processes that are collected in real time by sensors and online monitoring devices, including at least two types of data such as temperature, humidity, speed, tension, pressure, current, and vibration. This invention uses at least a portion of the signals from temperature, speed, tension, and pressure for model calibration, risk calculation, and interlock determination.

[0040] 4. Process parameters

[0041] Process parameters are settable or adjustable process variables that have a significant impact on quality and stability. They include at least: melt temperature, side-blowing air temperature, side-blowing air relative humidity, draw ratio, guide roller speed group, false twisting hot box temperature, oiling rate, network pressure, and winding tension. Their values ​​are subject to the combined constraints of the equipment's allowable range, stable operating range, and interlocking rules.

[0042] 5. Quality Indicators and Comprehensive Quality Evaluation Indicators for Cotton Imitation

[0043] The quality indicators include at least: linear density fluctuation index, breaking strength, breaking elongation, yarn count and / or hairiness index, and moisture absorption and wicking index.

[0044] The comprehensive quality evaluation index for imitation cotton is a comprehensive score obtained by weighting the above quality indicators according to their dimensions, and is used to express the overall performance of cotton feel and function.

[0045] Dimensionality and weighting can be implemented in the following forms:

[0046] Let the first The measured values ​​of each quality indicator are The target value (or the center of the expected interval) is The allowed range is Then its score This can be defined as interval normalization or bias scoring. Scoring formula:

[0047] ;

[0048] in, This indicates that the value is limited to... .

[0049] Comprehensive quality evaluation indicators for imitation cotton It can be defined as:

[0050] ;

[0051] in, For the quantity of quality indicators; For the first Each quality indicator weight satisfies and .

[0052] 6. Digital twin models, mechanism models, and learning models

[0053] The digital twin model of this invention is a computable model that maps the input batches, operating conditions, and process parameters of a production line to the output quality indicators, risks, energy consumption, etc., including:

[0054] Mechanism model: Based on physical and chemical mechanisms, it includes at least the relationship between heat transfer and tension coupling, the effect of side blowing on cooling and solidification rate, the effect of stretching on orientation crystallization, and the effect of false twist hot box on curling and shaping.

[0055] Learning model: Used to compensate for systematic biases in mechanism models, it can be implemented using recursive least squares, incremental neural networks, gradient boosting trees, time series models, etc.; its training objective is the difference between mechanism predictions and actual measurements.

[0056] 7. Upper bound of uncertainty in quality forecasting

[0057] The upper bound of quality prediction uncertainty is a conservative upper bound on the prediction error of quality indicators by the digital twin model, used for worst-case constraint verification in robust solutions. It can be obtained from the quantile statistics of historical stable batch error distribution (e.g., the 95th quantile) and can be adaptively increased or decreased through online error rolling updates, typically increasing as the error increases to enhance safety.

[0058] 8. Risk of broken threads, risk ceiling, and 90% risk trigger.

[0059] The risk of wire breakage is an indicator that characterizes the probability or intensity of wire breakage within a short time window in the future. It is obtained by fusing at least the abnormal winding tension factor and the abnormal process deviation factor.

[0060] The risk ceiling is the upper limit of the permissible risk; exceeding or approaching this limit will trigger protective actions. When the risk of wire breakage reaches 90% or more of the risk ceiling (i.e., risk ≥ 90%), subsequent safety exploration and constraint updates will be triggered.

[0061] The fusion method adopted in this invention: Let the tension anomaly score be... The process deviation score is Then there is a risk of broken threads. It can be defined as:

[0062] ;

[0063] in, For weight fusion.

[0064] 9. Process constraints, interlock constraints, feasible process range, and prohibited parameter combination library

[0065] Process constraints: Upper and lower limits constrained by the allowable range and stable operating range of the equipment;

[0066] Interlocking constraints: Rule constraints that prioritize safety and stability, and when triggered, restrict speed, draw, etc., or force a rollback;

[0067] Feasible process range: Under the combined effect of process constraints and interlocking constraints, optimize and control the parameter space that allows for searching or operation;

[0068] Disabled parameter combination library: records parameter combinations that cause the risk to approach or exceed the risk limit or cause the comprehensive quality evaluation index of imitation cotton to fall below the quality lower limit three times in a row; disabled combinations cannot be selected by the optimizer again until the deactivation conditions are met.

[0069] 10. Dual-time-scale control and rolling predictive control

[0070] Dual timescale control includes:

[0071] Slow cycle: Used to update optimal process settings, quality lower limit, risk upper limit or weight, etc., such as once every 5-30 minutes;

[0072] Fast cycle: Used to execute rolling predictive control output and apply a rate-of-change limit, such as once every 0.5–5 seconds.

[0073] Rolling predictive control refers to using a digital twin model to predict the trends of quality, risk, and energy consumption over a future period in each fast cycle, solving for the control quantity that should be applied at the current time, and then rolling it forward and executing it cyclically.

[0074] II. System

[0075] like Figure 1 As shown, the system of the present invention includes at least the following modules, each of which can be deployed on the same industrial control computer or distributed across edge computing nodes, production line controllers and host computer servers.

[0076] 1. Data Acquisition Module

[0077] The data acquisition module is used to collect:

[0078] 1) Raw material batch characteristics: Read from the raw material management system, laboratory testing system, or batch barcode system;

[0079] 2) Online operating condition data: collected from temperature sensors, humidity sensors, speed encoders, tension sensors, pressure sensors, etc.

[0080] 3) Process parameters and setpoints: Read the current settings and actual feedback from the control execution terminal;

[0081] 4) Quality indicators: obtained from online testing devices, such as online yarn / hair analyzers, online linear density estimators, online moisture conduction or moisture content characterization devices; and offline laboratory sampling, such as strength, elongation, moisture conduction, etc.

[0082] Quality indicators can be obtained by combining online and offline methods: online for fast-cycle control and offline for slow-cycle correction and weight update; moisture absorption and conduction indicators that are difficult to measure directly online can be obtained by establishing a mapping model using measurable surrogate quantities, such as moisture content change, surface resistance, and water absorption rate related signals.

[0083] 2. Digital Twin Module

[0084] The digital twin module comprises: a mechanistic model submodule, a learning model submodule, an online calibration submodule, and an uncertainty upper bound estimation submodule. Its outputs include: predicted values ​​of quality indicators; upper bounds on quality prediction uncertainty; predictions or estimates of risk and energy consumption-related quantities; and model credibility and error monitoring results.

[0085] 3. Optimize the decision-making module

[0086] The optimization decision module runs in a slow cycle, performs lexicographically robust solving, outputs the optimal process settings, and provides suggestions on feasible process ranges, interlock tightening / relaxation suggestions, and updated candidates for the disabled parameter combination library.

[0087] 4. Dual time scale control module

[0088] The dual-timescale control module includes a slow-cycle manager and a fast-cycle controller:

[0089] Slow Cycle Manager: Receives output from the optimization decision module and confirms and smoothly updates it by combining capacity, order specifications, and equipment status;

[0090] Fast cycle controller: performs rolling predictive control, outputs the executed quantity, and ensures stability through rate of change limiting, anti-saturation, and anti-disturbance strategies.

[0091] 5. Control Execution Module

[0092] The control and execution module communicates with the field actuators and includes at least: heating control (melt temperature, hot box temperature), side blowing control (temperature, humidity, air volume), drafting and guide roller speed control, network pressure control, oiling rate control, winding speed and winding tension control, etc. The execution method can be conventional industrial components such as frequency converters, servo drives, proportional valves, and temperature controllers.

[0093] 6. Safety Exploration and Constraint Self-Update Module

[0094] The security exploration and constraint self-updating module is used for:

[0095] Triggering conditions: The risk of broken fibers reaches 90% or more of the risk limit, or the comprehensive quality evaluation index of imitation cotton is lower than the quality lower limit for three consecutive times;

[0096] Constrained perturbation identification: Apply small perturbations to selected parameters within a safety margin and estimate sensitivity and uncertainty;

[0097] Update the feasible process range: shrinkage or limited expansion;

[0098] Tighten interlocking constraints: Adjust interlocking thresholds, rate of change limits, etc.;

[0099] Maintain a library of disabled parameter combinations: write, remove, version management, and traceability.

[0100] III. Specific Technical Route of the Method of the Invention

[0101] like Figure 2 As shown, the overall approach of this invention is a data-twin-robust optimization-dual-timescale closed-loop-secure exploration self-updating system. Its core contributions mainly focus on:

[0102] 1) The digital twin model adopts a structured fusion of mechanistic model and learning model compensation, and introduces error quantile statistics and online updates to form an upper bound on the uncertainty of quality prediction;

[0103] 2) Optimize by using lexicographical robust solution. In the worst case, first satisfy the hard constraints of "risk upper limit / quality lower limit", and then optimize energy consumption, quality and setting changes.

[0104] 3) The control adopts a dual time scale, with slow-cycle update settings and fast-cycle rolling predictive control with amplitude limiting;

[0105] 4) Introduce a safety mechanism of "90% trigger and three consecutive triggers" to perform restricted disturbance identification, feasible domain shrinkage / limited expansion, interlock tightening and disabled parameter combination library maintenance, so as to achieve adaptive steady-state operation against batch drift, environmental disturbance and aging error.

[0106] 1. S1: Obtain raw material batch characteristics and online operating condition data

[0107] The goal of Step One is to provide the input data foundation for digital twins, risk assessment, and optimized control.

[0108] 1.1 Acquisition and Preprocessing of Raw Material Batch Characteristics

[0109] Raw material batch characteristics must include at least one of the following two categories: chip viscosity grade or melt index grade; chip moisture content; masterbatch or additive ratio; batch storage time.

[0110] Implementation method:

[0111] 1) When a batch is received into the warehouse, the moisture content, viscosity, or melt index of the slices are measured in the laboratory; the proportion of masterbatch is recorded by the batching system.

[0112] 2) Before production begins, the production line scans the batch barcode, and the data acquisition module reads the corresponding batch characteristics from the database;

[0113] 3) Standardize and cache batch features for subsequent model input.

[0114] 1.2 Online operating condition data acquisition and synchronization

[0115] Online operating condition data should include at least two of the following: temperature, humidity, velocity, tension, and pressure. To improve control and model calibration effectiveness, it is preferable to collect at least [number missing] data.

[0116] Temperature-related: melt temperature, hot box temperature;

[0117] Temperature and humidity categories: side-blowing air temperature, side-blowing air relative humidity;

[0118] Tension type: winding tension;

[0119] Speed-related: Critical roller speed;

[0120] Stress-related: Network stress.

[0121] Data synchronization method:

[0122] All sampled signals are timestamped and uniformly enter the data bus;

[0123] Set the fast cycle to control the sampling period, such as 1 second. For faster signals, you can first sample at high frequency, then filter at low frequency, and then downsample.

[0124] Missing or outlier points can be handled using conventional methods such as forward filling and moving median filtering to avoid introducing abrupt changes.

[0125] 2. S2: Digital twin model construction and online calibration, generation of upper bounds on uncertainty

[0126] Step two forms the basis for robust optimization and rolling predictive control. The mechanistic model of this invention must include the coupling relationship between heat transfer and tension, and the learned model is used to compensate for deviations; and based on error quantile statistics and online rolling updates, an upper bound on the uncertainty of quality prediction is obtained, making the worst-case scenario calculable.

[0127] 2.1 Construction of the Mechanism Model

[0128] The cotton-like functional filament production line can be divided into several sections according to the process: melting and extrusion section; spinning and side-blowing cooling section; drawing and orientation section; web processing section; false twisting hot box setting section; oiling and winding section.

[0129] The mechanistic model does not require high-fidelity CFD or finite element analysis for every segment, but it must at least express the direction of thermal-tension coupling and the influence of key parameters in an engineering-usable form. A combination of state-space modeling and empirical correlation can be used.

[0130] The temperature and humidity of the side-blown airflow affect the fiber cooling rate and the location of the curing points, thus influencing forming tension, orientation, and linear density fluctuations. Fiber temperature states can be defined. With cooling intensity coefficient Side-blowing air temperature and humidity:

[0131] ;

[0132] in, For fiber temperature, This refers to the side-blowing air temperature. For time.

[0133] Draw ratio, roll speed, and winding tension directly affect the risk of yarn breakage and yarn evenness / hair. The tension coupling sub-model adopts a simplified tension transmission model:

[0134] Define the tension state of each segment It is determined by the velocity difference and the viscoelasticity of the material; the tension fluctuation is linked to the rate of velocity change, and a rate of change limit is introduced.

[0135] ;

[0136] in, For the first Duan Zhangli; For the first Segment speed; This is an empirical coefficient; This is a reference temperature.

[0137] Regarding orientation crystallization and false twist setting, the draw ratio and oven temperature affect the degree of orientation crystallization and crimp setting, thus influencing breaking strength, elongation, and cotton feel. These can be introduced as influencing factors for predicting quality indicators.

[0138] ;

[0139] in, For orientation / shape-fixed integrated state factors; It is a monotonic function or a piecewise linear function; and As an intermediate variable input for predicting quality indicators.

[0140] The mechanistic model output includes at least predicted values ​​for quality indicators, such as linear density fluctuation, yarn / hair trend, strength / elongation trend, and moisture conductivity trend. It can also simultaneously output energy consumption estimates (heating and motor) and risk-related predictions (tension trend).

[0141] 2.2 Learning Model

[0142] The learning model, also known as the residual compensation model, aims to compensate for systematic biases in the mechanistic model, making the output of the digital twin model more closely resemble the real production line while maintaining interpretability and controllability. Multiple implementation options are available for the learning model; a simplified approach (A) can be used for small-scale production, while approach B is recommended for more complex production lines.

[0143] Option A: Recursive Least Squares Residual Model

[0144] Inputs: Batch characteristics, online operating data, process parameters, and mechanistic model prediction output;

[0145] Output: Predicted residuals of quality indicators;

[0146] Form: Linear or piecewise linear model;

[0147] Update: Recursive least squares with a forgetting factor.

[0148] ;

[0149] in, For residuals; The input feature vector; For parameter vectors.

[0150] Recurrence formula:

[0151] ;

[0152] Parameter definition: For the first Time parameters; For gain; For actual measurement; For prediction.

[0153] Option B: Lightweight Temporal Neural Network Residual Model

[0154] Input feature set:

[0155] Batch features, i.e., static features;

[0156] Online operating conditions, such as temperature, humidity, tension, pressure, and speed;

[0157] Process setpoints and actual feedback;

[0158] Quality indicators of mechanistic model predictions.

[0159] Network structure:

[0160] First layer: One-dimensional convolutional or gated recurrent units, used to extract temporal features;

[0161] The second layer: a fully connected layer, mapped to the residual output;

[0162] Output layer: Outputs a residual value for each quality indicator.

[0163] Training objective: To minimize the mean squared error between the predicted and actual residuals; and to weight different quality metrics to match production priorities. Loss function:

[0164] ;

[0165] in, For loss; For the number of indicators; For the first True residual of the indicator; To predict residuals; For weights.

[0166] Training dataset organization: The time series data of each batch is sliced ​​according to a fixed time window; the labels are obtained from the measured quality index—mechanism prediction quality index.

[0167] 2.3 Fusion Output of Digital Twin Models

[0168] Digital twin model output quality index prediction It is obtained by superimposing mechanism prediction and residual compensation:

[0169] ;

[0170] in, For digital twin prediction; For mechanism prediction; Output residuals for the learned model.

[0171] 2.4. Online calibration strategy

[0172] The core of online correction lies in continuously monitoring the prediction error and the post-compensation error, and adjusting the learning model update rate, forgetting factor, or re-initializing the model parameters as necessary. The following strategy combination is employed:

[0173] Sliding window correction: with the most recent Minute data is updated in a window;

[0174] Error threshold: When the error after compensation exceeds the preset threshold and continues to occur, increase the update frequency of the learning model or trigger retraining;

[0175] Batch migration correction: When batch characteristics change significantly, increase the weight of batch characteristics or enable batch-specific parameters.

[0176] 2.5. Generation of the Upper Bound of Quality Prediction Uncertainty

[0177] To ensure robust solutions hold true in the worst-case scenario, this invention introduces an upper bound on the uncertainty of quality prediction. The specific implementation is as follows:

[0178] 1) For historically stable batches, statistically analyze the prediction error sequence for each quality indicator. ;

[0179] 2) Calculate the absolute value of the error The 95th percentile value is used as the initial upper bound. ;

[0180] 3) During online operation, the error quantile value within the nearest window is calculated in a rolling manner. ;

[0181] 4) Obtain the current upper bound using a smooth update method. :

[0182] ;

[0183] in, For the first Upper bound of indicator uncertainty; For smoothing coefficients; This is for window quantile estimation.

[0184] When the online error increases Increase, thus Adaptive enhancement improves robustness.

[0185] 3. S3: Lexicographically robust solution

[0186] Step 3 requires considering the worst-case scenario of the upper bound of the uncertainty in quality prediction under process constraints and interlocking constraints, and outputting the optimal process setpoint by satisfying the hard constraints and the target sorting in lexicographical order.

[0187] 3.1. Construction of the constraint system

[0188] 3.1.1 Process Constraints

[0189] Process constraints are a combination of the equipment's permissible range and its stable operating range.

[0190] For each process parameter ,definition:

[0191] Equipment permissible range: ;

[0192] Stable operating range: (Based on statistics from historically stable batches).

[0193] Methods for obtaining the stable operating range:

[0194] Select a set of batches with low fiber breakage rate and qualified quality over a period of time;

[0195] For each parameter, take a quantile interval, such as 5%–95%, as the stable range;

[0196] For new batches, fine-tuning can be made based on batch characteristics, such as offsetting by viscosity grade.

[0197] 3.1.2 Interlocking Constraints

[0198] Interlocking constraints are a set of rules, typically including:

[0199] When the winding tension exceeds the upper limit or the amplitude of the tension fluctuation exceeds the upper limit, it is prohibited to increase the winding speed and the draw ratio is limited to be adjusted upward.

[0200] When the risk of wire breakage exceeds the risk limit or reaches the preset ratio of the risk limit, the winding speed and / or draw ratio are forcibly reduced, and the network pressure is adjusted to the safe range for preventing wire breakage.

[0201] When the melt temperature deviates from the preset deviation, the guide roller speed group and the rate of change of the draw ratio are limited, while providing a basis for subsequent rate of change limitation.

[0202] Interlocking rules can be transformed into feasible domain constraints and allowed action set constraints in optimization. For example, certain parameters may be prohibited from being increased under certain states, or a limit may be set on the rate of change of a given parameter.

[0203] 3.2 Worst-case quality and risk verification

[0204] The prediction quality index of digital twin models is The upper bound of uncertainty is Then the worst-case scenario can be verified using a conservative approach:

[0205] For indicators where a higher is always better, such as certain scores for strength and moisture wicking capacity, take... ;

[0206] For indicators where smaller is better, such as linear density fluctuation and certain quantities of yarn / hair, take... ;

[0207] For the comprehensive score, after standardizing it to the same dimension, add uncertainty to the most unfavorable direction.

[0208] To facilitate implementation, the comprehensive quality evaluation indicators for imitation cotton can be directly applied. Set an upper bound for uncertainty. (by each) (synthetic or empirical estimation), and in the worst case, the following requirements are made:

[0209] Values ​​of the comprehensive quality evaluation index for cotton-like materials under the worst-case scenario. ;

[0210] The value of the wire breakage risk under the worst-case scenario. .

[0211] 3.3 Lexicographically Robust Solution Algorithm

[0212] Lexicographically robust solution is equivalent to hierarchical constraints plus hierarchical objectives. It is implemented as follows:

[0213] Construct a comprehensive evaluation function, but still maintain the lexicographical order hard constraint as the priority:

[0214] Candidates that violate the risk ceiling or quality floor will be directly eliminated;

[0215] For those that meet the requirements, calculate the comprehensive target value, prioritizing energy consumption, followed by quality, and then minimal change.

[0216] Let the energy consumption per unit output be The cotton-like rating is Set the variable as Then it can be minimized under the hard constraints:

[0217] ;

[0218] in, For the overall objective; For quality weights; For variable weights; Energy consumption per unit of output; As a comprehensive quality evaluation index for imitation cotton; To set the variable amount.

[0219] in It can be defined as a weighted distance relative to the currently set value:

[0220] ;

[0221] in, The number of parameters; For the first Parameter weights; Set for candidates; This is the current setting.

[0222] 3.4 Calculation of energy consumption per unit output

[0223] Energy consumption per unit output can be calculated as the cumulative energy consumption / effective output of the statistical window. It can be obtained from the power estimation of electricity meters and frequency converters, and the heating power measurement. Ineffective output is eliminated based on the shutdown due to wire breakage.

[0224] 4. S4: Dual-time-scale control and rolling predictive control

[0225] The goal of step four is to stably track the optimal process setpoint under disturbances and batch errors without triggering interlocking constraints; the key points are: slow cycle update strategy, fast cycle rolling predictive controller, rate of change limit, and interlocking coordination.

[0226] 4.1 Slow-cycle update strategy

[0227] Slow cycle (5-30 minutes) execution content:

[0228] 1) Receive the optimal process setting value output from step three;

[0229] 2) Smooth the setpoint to avoid tension fluctuations caused by abrupt changes;

[0230] 3) Conduct feasibility verification based on equipment status and interlocking constraints;

[0231] 4) Issue new target settings to the fast cycle controller.

[0232] Apply first-order filtering or segmented ramping to the setpoint:

[0233] If a parameter changes beyond the allowed step size, it will be updated incrementally in multiple steps.

[0234] After each step forward, observe whether the tension and risk increase; if they do, pause the forward movement.

[0235] 4.2 Fast Cycle Rolling Predictive Controller

[0236] The fast-cycle (0.5–5 seconds) controller executes at each sampling time:

[0237] 1) Read the current online operating status data and current settings;

[0238] 2) Use digital twin models to predict the trends in quality, risk, and energy consumption in short-term future windows (e.g., the next 20–60 seconds);

[0239] 3) Under the constraints of "rate of change limit + interlocking allowed action", solve for the execution quantity that should be applied at the current time;

[0240] 4) Output the execution quantity to the control execution module;

[0241] 5) Scroll forward and execute repeatedly.

[0242] 4.2.1 Selection of Control Variable and Controlled Variable

[0243] The controllable parameters can be: heating power setting, side blowing air volume or temperature and humidity setting, guide roller speed setting, draft ratio setting, network pressure setting, winding tension setting, etc. The controlled parameters can be: winding tension stability, risk indicators, comprehensive quality evaluation indicators for imitation cotton, energy consumption, etc.

[0244] 4.2.2 Rate of Change Limit

[0245] To reduce the risk of yarn breakage, an upper limit must be set for the rate of change of key settings:

[0246] The draw ratio changes by no more than a certain threshold per second;

[0247] The winding speed changes by no more than a certain threshold per second;

[0248] The temperature setting of the hot box should not change beyond a certain threshold.

[0249] The specific threshold can be obtained from the equipment manual and historical stable batch statistics.

[0250] 4.2.3 Coordination with Interlocking Constraints

[0251] Interlocking constraints have higher priority than controllers. Implementation method:

[0252] Before the controller outputs, the interlocking verification module checks whether the interlocking rules are violated.

[0253] If interlocking may be triggered, the control quantity should be pruned, such as prohibiting speed increase;

[0254] If the risk has reached the upper limit, then a conservative retreat action (slowing down / reducing stretching / adjusting network pressure) should be implemented directly.

[0255] 5. S5: Restricted disturbance identification under 90% trigger rate and three consecutive triggers, feasible region shrinkage, interlock tightening and disabled library maintenance.

[0256] 5.1 Trigger Decision

[0257] For the risk of broken wires: calculate the risk value in real time and compare it with the risk limit. If the risk reaches 90% or more, it will be triggered.

[0258] For the comprehensive quality evaluation index of cotton-like products: during slow cycles or quality inspection updates, the number of consecutive occurrences is counted, and a trigger is established when three consecutive occurrences are reached. The number of consecutive occurrences can be defined by a continuous window or continuous sample points within a batch. In production, it is recommended to use the slow cycle update point as the statistical granularity to avoid noise-induced false triggering.

[0259] 5.2 Restricted Disturbance Identification

[0260] 5.2.1 Selection of Disturbance Target

[0261] Preferably select parameters that are sensitive to risk and quality and can recover quickly as perturbation targets, such as network pressure, draw ratio fine-tuning, winding tension setting fine-tuning, and side blower temperature fine-tuning. Avoid making large perturbations to parameters with high thermal inertia and slow recovery, such as melt temperature.

[0262] 5.2.2 Disturbance Amplitude and Window

[0263] The disturbance amplitude shall not exceed a preset proportion of the allowable adjustment step size, and the disturbance duration shall not exceed a preset identification window; termination shall be initiated immediately if the safety margin is insufficient. Specifically, the following settings are available:

[0264] Disturbance amplitude: 10%–30% of the allowable step size;

[0265] Recognition window: 20–120 seconds;

[0266] Safety margins: Both the tension safety margin and the temperature safety margin must be higher than the lower limit.

[0267] 5.2.3 Identification Output

[0268] Identification output:

[0269] The sensitivity of key process parameters to risk, i.e. whether increasing the parameter significantly increases the risk;

[0270] The sensitivity of key process parameters to cotton-like indicators;

[0271] Does the uncertainty of current digital twin models increase?

[0272] Based on this, update the upper bound of quality prediction uncertainty, the range of feasible processes, and the strength of interlocking constraints.

[0273] 5.3 Feasible process range for shrinkage

[0274] To mitigate risk, a conservative strategy should be prioritized when the trigger occurs:

[0275] Tighten the range of parameters that are positively correlated with risk, such as lowering the upper limit of the maximum winding speed and the upper limit of the maximum draw ratio;

[0276] Tighten tension-related parameters, such as improving tension control accuracy and limiting the rate of change;

[0277] The parameter group at the trigger time and its neighborhood are marked as high-risk areas to avoid re-exploration.

[0278] 5.4 Tighten interlocking constraints

[0279] To strengthen the rules, stricter interlocking constraints may include:

[0280] Lower the tension limit / fluctuation limit;

[0281] Adjust the risk trigger ratio from a higher value to a more conservative value, such as temporarily setting it to an 80% warning level after a 90% trigger rate.

[0282] Tighten the rate of change limit threshold to reduce transient shocks;

[0283] Forced rollback strategies are more aggressive, such as simultaneously reducing speed and stretch.

[0284] 5.5 Maintain the library of disabled parameter combinations

[0285] 5.5.1 Writing Rules

[0286] When the trigger condition is met (90% trigger rate) or the quality falls below the lower limit three times consecutively, the key parameter group at that time is written into the disabled library. The following parameters can be recorded:

[0287] Parameters such as melt temperature, side blowing temperature and humidity, draw ratio, guide roller speed group, hot box temperature, oiling rate, network pressure, and winding tension;

[0288] Batch characteristics and environmental conditions;

[0289] Trigger type (risk trigger / quality trigger);

[0290] Evidence includes tension fluctuations, energy consumption, and quality scores at the time of triggering.

[0291] 5.5.2 Conditions for cancellation

[0292] The release condition can be set as follows: the confidence level reaches the threshold in subsequent restricted disturbance identification and the interlocking feasibility is verified before release; the optimizer must not output disabled combinations before release.

[0293] Version 5.5.3 Management and Traceability

[0294] The disabled library should include a version number to facilitate tracing the reason for a policy change; in production, a subset of the disabled library can be maintained separately according to product specifications / batch types.

[0295] 5.6 Write back the updated optimal process settings

[0296] After completing the identification of constrained disturbances and the updating of constraints, under the updated feasible process range and interlocking constraints, the lexicographical robust solution of S3 is re-executed to output the updated optimal process settings.

[0297] 6. Typical models

[0298] 6.1 Dataset Definition and Collection Organization

[0299] 6.1.1 Data Table Structure

[0300] Each batch generates a "batch data package", which contains at least:

[0301] Batch static characteristics table: viscosity / melt index grade, moisture content, masterbatch ratio, storage time, etc.;

[0302] Process timing table (by timestamp): temperature, humidity, speed, tension, pressure, setpoint, execution quantity, motor current, etc.;

[0303] Quality label table: online quality indicators (strip / hair, linear density fluctuation, etc.) and offline sampling inspection results (strength, elongation, moisture wicking, etc.);

[0304] Event table: wire breakage event time, downtime, alarm, etc.

[0305] 6.1.2 Time Alignment and Window Slicing

[0306] For the process timing table, align with a fixed sampling period (1 second); divide the continuous sequence according to length. Sliding window slices ( (seconds), step size is ( (seconds). Each window sample contains:

[0307] Input: Operating conditions and parameter sequences within the window, and batch static characteristics;

[0308] Tags: Measured values ​​of quality indicators at the end of the window or the average of the window.

[0309] 1.3 Training / Validation / Testing Partition

[0310] Divide the data into batches rather than randomly at specific times to avoid information leakage. Set partitioning:

[0311] Training set: 70% of historical batches;

[0312] Validation set: 15%;

[0313] Test set: 15%.

[0314] If there are differences in product specifications, stratified sampling can be performed according to specifications.

[0315] 6.2 Mechanism Model Parameter Calibration

[0316] Empirical coefficients in the mechanistic model, such as the cooling intensity coefficient and the tension coupling coefficient, can be calibrated in the following ways:

[0317] 1) Select several stable batches;

[0318] 2) The objective is to minimize the error between the predicted and measured quality indicators.

[0319] 3) Use least squares or grid search to determine the coefficients;

[0320] 4) Group the coefficients according to batch characteristics, such as calibrating a set of coefficients according to viscosity grade to achieve batch migration correction.

[0321] 6.3 Learning Model Training

[0322] 6.3.1 Training of the recursive least squares residual model

[0323] Initial stage: Fit initial parameters offline using the training set;

[0324] Online phase: After obtaining new samples in each time window, the parameters are updated according to the recursive formula;

[0325] Forgetting factor selection: 0.95–0.995 to make the model more sensitive to recent data.

[0326] 6.3.2 Training of Lightweight Temporal Neural Network Residual Model

[0327] Specific network configuration:

[0328] Input dimensions:

[0329] Batch static feature dimensions: 4–10 dimensions;

[0330] Process feature dimensions per second: 20–60 dimensions;

[0331] Window duration: 60 seconds.

[0332] Network: 1) Temporal coding layer: 1 gated recurrent unit layer, 32–128 hidden units;

[0333] 2) Fully connected layers: 2 layers, widths 64 and 32;

[0334] 3) Output layer: Outputs the residuals of each quality index.

[0335] Activation function: conventional nonlinear, such as rectification;

[0336] Loss function: Weighted mean squared error;

[0337] Optimizer: Conventional adaptive gradient;

[0338] Training rounds: stop early based on the validation set;

[0339] Regularization: Weight decay and discarding can be used to avoid overfitting.

[0340] Online incremental update method:

[0341] After each slow cycle (5–30 minutes) is completed, the data from that cycle is appended as a new training sample.

[0342] Perform incremental training in several steps with a small learning rate;

[0343] If the error continues to rise, a re-initialization or rollback to the previous stable model version will be triggered.

[0344] 6.3.3 Parameter Selection for the Upper Bound of Uncertainty

[0345] Initial quantile: 95th percentile;

[0346] Online window length: 30–120 minutes;

[0347] Smoothing factor: 0.8–0.95 is recommended.

[0348] 6.3.4. Risk ceiling, quality floor and weight selection

[0349] Risk ceiling: can be determined by working backward from historical breakage rate targets, combined with tolerance for tension fluctuations;

[0350] Quality lower limit: set by customer specifications or internal control standards, such as the overall score not being lower than a certain threshold;

[0351] Weights: Initial weights are provided by the product specification library / customer metric library, and fine-tuned by combining the historical best batches;

[0352] Nine out of ten triggers: fixed at 90% of the risk limit; three consecutive triggers: statistics are based on slow cycle update points.

[0353] 6.3.5 Forecast Length and Weights in Rolling Forecast Control

[0354] Prediction length: 20–120 seconds;

[0355] Control weights: Risk and tension stability take precedence over energy consumption, and energy consumption optimization is mainly completed by slow cycles;

[0356] Change rate limit: determined by the equipment and historical stable range, and can be temporarily tightened after being triggered.

[0357] Example 1: Energy consumption optimization and quality maintenance under stable batch conditions

[0358] IV. Examples

[0359] Example 1: Energy consumption optimization and quality maintenance under stable batch conditions

[0360] 1) Step 1: Read the batch viscosity grade and moisture content; collect data on side blowing temperature and humidity, winding tension, guide roller speed, and network pressure;

[0361] 2) Step Two: Use the mechanistic model to predict the fluctuation trend of bark / hair and linear density, and learn the model to compensate for the bias; calculate the upper bound of the 95th percentile of error and update it on a rolling basis;

[0362] 3) Step 3: Search for candidate settings within the stable operating range, and eliminate those that exceed the upper limit of risk or fall below the lower limit in the worst case scenario; among the remaining candidates, prioritize the one with the lowest energy consumption;

[0363] 4) Step 4: The slow cycle updates the settings every 10 minutes, and the fast cycle rolls predictive control every 1 second to limit the rate of change of draw ratio and winding speed;

[0364] 5) Step five: Stable operation, no triggering; the system continuously updates the upper bound of uncertainty and the parameters of the learning model.

[0365] Results: Under the premise that the quality meets the lower limit and the risk meets the upper limit, the energy consumption per unit output decreases, the setting changes are small, and the operation is stable.

[0366] Example 2: Triggering three consecutive triggers for quality degradation due to batch instability

[0367] When the comprehensive quality evaluation index of imitation cotton is below the lower limit of quality for three consecutive slow cycles:

[0368] The system triggers step five, selecting network pressure and stretch ratio for constrained disturbance identification;

[0369] It was found that the draw ratio sensitivity increased and the model error increased, so the upper bound of the uncertainty was raised;

[0370] Narrow the feasible process range (reduce the upper limit of draw ratio and the upper limit of winding speed), and tighten interlocking constraints;

[0371] Write the trigger-time parameter group into the disabled parameter combination library;

[0372] Repeat step three to output the new conservative optimal setting and write it back.

[0373] Results achieved: The risk of yarn breakage decreased, and the quality score rebounded to above the lower limit of quality.

[0374] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, which are used to illustrate the technical solutions and implementation process of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the essence and spirit of the technical solution of this invention, those skilled in the art can make various equivalent substitutions or modifications to the system composition, module division, parameter value range, model form, training and update strategy, control cycle and triggering conditions in the above embodiments. For example, the learning model can be replaced by a gradient boosting tree, support vector regression or temporal neural network instead of a residual model; the implementation of lexicographical robust solution can be replaced by Bayesian optimization or heuristic search instead of grid / sampling screening; and rolling predictive control can be replaced by gain scheduling control or piecewise linear predictive control. As long as the technical feature combination of "digital twin based on the fusion of mechanism model and learning model and online correction, construction of upper bound of quality prediction uncertainty, lexicographical robust solution under process constraints and interlocking constraints, adoption of dual time scale closed-loop rolling adjustment and application of rate of change limit, identification of restricted disturbance and updating of feasible process range when risk or quality is triggered, interlocking constraints and disabled parameter combination library and write back set value" is still satisfied, it should all fall within the protection scope of this invention.

Claims

1. A multi-parameter process optimization method for a cotton-like functional filament production line, characterized by, Includes the following steps: S1 acquires raw material batch characteristics and online operating condition data; S2 establishes a mechanistic model that includes the coupling relationship between heat transfer and tension, and establishes a learning model to compensate for the prediction bias of the mechanistic model. The two are integrated to form a digital twin model. Based on the quantile statistics of historical stable batch prediction errors and combined with online error rolling updates, the upper bound of the uncertainty of quality prediction is obtained. S3 performs lexicographically robust solution under process constraints and interlocking constraints: considering the worst case of the upper bound of the uncertainty of the quality prediction, first ensure that the risk of broken yarn does not exceed the risk upper limit, then ensure that the comprehensive quality evaluation index of imitation cotton is not lower than the quality lower limit, on this basis, prioritize minimizing the energy consumption per unit output, then improve the comprehensive quality evaluation index of imitation cotton, and finally minimize the setting change to obtain the optimal process setting value. S4 employs dual time-scale control, with the slow cycle updating the optimal process setpoint and the fast cycle using rolling predictive control and applying a rate-of-change limit, so that key process parameters track the optimal process setpoint without triggering interlocking constraints. When the risk of broken yarn is greater than or equal to 90% of the risk limit or the comprehensive quality evaluation index of imitation cotton is lower than the quality limit for three consecutive times, the S5 triggers the restricted disturbance identification and shrinks the feasible process range, tightens the interlock constraints and maintains the prohibited parameter combination library, and writes back the updated optimal process setting value. The upper bound of the quality prediction uncertainty is obtained through the following steps: 1) For historically stable batches, statistically analyze the prediction error sequence for each quality indicator. ; 2) Calculate the absolute value of the error The 95th percentile value is used as the initial upper bound. ; 3) During online operation, the error quantile value within the nearest window is calculated in a rolling manner. ; 4) Obtain the current upper bound using a smooth update method. : ； in, For the first Upper bound of indicator uncertainty; For smoothing coefficients; For window quantile estimation; When the online error increases Increase, thus Adaptive enhancement improves robustness.

2. The method according to claim 1, characterized in that: The raw material batch characteristics include at least two of the following: slice viscosity or melt index grade, slice moisture content, masterbatch or additive ratio, and batch storage time; the online operating condition data include at least two of the following: temperature, humidity, speed, tension, and pressure.

3. The method according to claim 1, characterized in that: The process parameters corresponding to the process constraints include at least melt temperature, side-blowing temperature, side-blowing relative humidity, draw ratio, guide roller speed group, false twist hot box temperature, oiling rate, network pressure and winding tension; the guide roller speed group includes at least the first draw roller speed, the second draw roller speed, the false twist input roller speed, the false twist output roller speed and the winding speed.

4. The method according to claim 1, characterized in that: The quality indicators used to calculate the comprehensive quality evaluation index of the imitation cotton include at least linear density fluctuation index, breaking strength, breaking elongation, evenness index and / or hairiness index, as well as moisture absorption and wicking index; the comprehensive quality evaluation index of the imitation cotton is obtained by weighting the above quality indicators after they are converted to the same dimension, and the weights are determined by the product specification library or customer indicator library and are allowed to be updated adaptively based on historical stable batch statistics.

5. The method according to claim 1, characterized in that: The mechanistic model includes at least the effects of side-blowing temperature and humidity on the cooling and curing rate, the effects of draw ratio and guide roller speed group on orientation crystallization, and the effects of false twist hot box temperature on curling and shaping. The effects are mapped to the prediction terms of linear density fluctuation, evenness or fuzz, and moisture absorption and conduction index. And / or, the learning model uses the difference between the quality index predicted by the mechanistic model and the quality index measured online as the learning object, and adopts a recursive update or sliding window update method to ensure that the compensated error meets the upper limit of the error corresponding to each quality index within a preset time window.

6. The method according to claim 1, characterized in that: The upper bound of the uncertainty of the quality prediction is determined by the preset quantile value of the historical stable batch prediction error distribution, and is adaptively increased as the online error increases, so as to ensure that the lexicographically robust solution still satisfies the lower bound of quality and the upper bound of risk in the worst case. And / or, the risk of wire breakage is obtained by fusing the winding tension anomaly factor and the process deviation anomaly factor; The abnormal winding tension factor consists of at least the tension fluctuation amplitude, the tension peak value, and the peak value duration. The abnormal process deviation factor consists of at least two of the following: melt temperature deviation, side blowing temperature and humidity deviation, network pressure deviation, oiling rate deviation, and false twisting hot box temperature deviation.

7. The method according to claim 1, characterized in that: The process constraints include the equipment allowable range and stable operating range of each process parameter, wherein the stable operating range is obtained from the statistics of historical stable batches; the lexicographical robust solution prioritizes the search within the stable operating range, and when the lower quality limit and upper risk limit are satisfied, the search is expanded within the equipment allowable range; and / or, the interlocking constraints include at least one or a combination of the following rules: when the winding tension exceeds the tension upper limit or the tension fluctuation amplitude exceeds the fluctuation upper limit, the winding speed is prohibited from being increased and the draw ratio is limited to be adjusted upward; When the risk of wire breakage exceeds the risk limit or reaches the preset ratio of the risk limit, the winding speed and / or draw ratio are forcibly reduced, and the network pressure is adjusted to the safe range for preventing wire breakage. When the melt temperature deviates beyond the preset deviation, the rate of change of the guide roller speed group and the draw ratio is limited to meet the rate of change limit.

8. The method according to claim 1, characterized in that: In the dual timescale control, the slow cycle is 5 to 30 minutes, used to update at least one of the optimal process setting, the lower quality limit, and the upper risk limit; the fast cycle is 0.5 to 5 seconds, used for the rolling predictive control output execution quantity; the execution quantity includes at least one of the following: heating power setting, side blowing air volume or temperature and humidity setting, guide roller speed setting, draw ratio setting, network pressure setting, and winding tension setting, and a change rate limit is applied to the execution quantity.

9. The method according to claim 1, characterized in that: When the risk of broken yarn reaches 90% or more of the risk limit or the comprehensive quality evaluation index of imitation cotton is lower than the quality lower limit three times in a row, the corresponding parameter group will be written into the prohibited parameter combination library, the feasible process range will be narrowed, and the interlocking constraints will be tightened. When five consecutive slow cycles simultaneously meet the preset ratio where the risk of broken fibers is lower than the upper risk limit and the comprehensive quality evaluation index of imitation cotton is not lower than the lower quality limit, the feasible process range is expanded to a limited extent and the updated optimal process setting value is written back.

10. A multi-parameter process optimization control system for a cotton-like functional filament production line, characterized in that, To implement the method according to any one of claims 1 to 9, comprising: The data acquisition module is used to collect raw material batch characteristics, online operating condition data, process parameters and quality indicators; the digital twin module is used to establish a mechanism model and compensate for prediction deviations through the learning model to form a digital twin model, and output the upper bound of the uncertainty of quality prediction. The optimization decision module is used to perform lexicographically robust solving and output the optimal process setpoints under process constraints and interlocking constraints. The dual-timescale control module is used to update the optimal process setpoint in a slow cycle and to perform rolling predictive control and apply a rate-of-change limit in a fast cycle. The control execution module is used to drive the heating, side blowing, drafting, false twisting, networking and winding actuators; the safety exploration and constraint self-updating module is used to trigger restricted disturbance identification, update the feasible process range, tighten interlock constraints and maintain the prohibited parameter combination library when the risk of yarn breakage reaches 90% or more of the risk limit or the comprehensive quality evaluation index of imitation cotton is lower than the quality lower limit three times in a row, and write the updated optimal process setting value back to the dual time scale control module.

Images