Process optimization method, system, device and medium based on multi-objective genetic algorithm

Abstract

The application discloses a process optimization method based on a multi-objective genetic algorithm, belonging to the technical field of quality control. The product specification and the characteristic parameter set of the rubber compound formula are obtained according to the work order instruction, the rheological properties of the rubber compound are sensed online, and the formula parameters are updated; the historical data of the process knowledge base are called to construct a sample set, a multi-objective fitness evaluation model is trained, and the quality, efficiency, energy efficiency and surface defect risk score prediction are realized. The decision variables such as the speed of the rubber mixing machine and the roller temperature are demarcated, and the parameter feasible region is constructed relying on multiple constraints; through the multi-objective genetic algorithm, the maximum of three evaluation scores and the condition that the surface defect risk is lower than the threshold value are taken as the conditions, the Pareto optimal solution set is solved, and the process formula is screened according to the production strategy and is issued for execution. In production, data are collected in real time, and a surface quality evaluation index is calculated; when the index exceeds the standard or the key parameters drift, local rapid optimization fine-tuning parameters are triggered, intelligent optimization of process parameters is realized, the product qualification rate is improved, and the production energy efficiency and the equipment operation stability are considered.

Description

Technical Field

[0001] This application relates to the technical field of quality control, and in particular to a process optimization method, system, equipment and medium based on a multi-objective genetic algorithm. Background Technology

[0002] Rubber products are widely used in new energy vehicles, deep-sea exploration, aerospace, and many other fields. Their performance and production efficiency directly depend on the precision of parameter control in core processes such as plasticizing, mixing, and extrusion. Process optimization is a key path to solving industry pain points and promoting high-quality industrial development. Currently, the rubber manufacturing industry is gradually transforming towards intelligent manufacturing, but many problems still need to be addressed in the production process: there is a strong coupling between the rubber mixing and extrusion processes; the rotation speed, roller temperature, and uniformity of rubber compound thickness in the rubber mixing stage affect extrusion stability; the multi-stage temperature zone settings, traction speed, and current load in the extrusion stage affect the final size and fluctuations of the semi-finished product. Furthermore, quality, efficiency, and energy consumption are mutually restrictive in production: increasing speed may increase output but cause dimensional deviations, increased mold trials, or increased downtime; it may also increase energy consumption. In addition, there are characteristic fluctuations between batches of rubber compound formulations, and rheological parameters are easily affected by environmental factors, leading to mismatches between process parameters and actual production conditions.

[0003] Currently, the commonly used conventional process optimization technology in the industry is a parameter control method based on orthogonal experiments and human experience. Its core process is as follows: First, based on product specifications and rubber compound formulations, and combined with operators' past production experience, the ranges of key process parameters such as the mixing mill speed, roller temperature, and extruder heating temperature are initially set. Then, through orthogonal experimental design, different parameter combinations are selected for small-batch trial production, collecting data on process parameters, product qualification rate, and energy consumption during the trial production process. Finally, through range analysis or single-factor variable analysis, the optimal process parameter combination for a single objective (such as the highest qualification rate or the lowest energy consumption) is selected and issued as the standard process to the production system for execution. This conventional technique still has significant limitations: First, it can only achieve optimization of a single objective, easily leading to a situation where one aspect is prioritized at the expense of others; second, it relies on human experience to set the initial parameter range, resulting in strong subjectivity. In summary, this can easily lead to batch product defects and is difficult to adapt to the demands of modern large-scale, high-precision production. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a process optimization method, system, equipment, and medium based on a multi-objective genetic algorithm.

[0005] Firstly, this application provides a process optimization method based on a multi-objective genetic algorithm, employing the following technical solution:

[0006] A process optimization method based on a multi-objective genetic algorithm includes:

[0007] Based on the work order instructions issued by the manufacturing execution system, obtain the product specification feature parameter set and the rubber compound formula feature parameter set; perform online rheological property sensing on the rubber corresponding to the rubber compound formula feature parameter set, obtain the actual rheological parameters, and update the rubber compound formula feature parameter set;

[0008] Based on the updated set of rubber compound formulation feature parameters and the set of product specification feature parameters, the parameter set of historical process parameters, historical work order qualification rate, historical work order equipment effective utilization rate, historical work order unit output energy consumption and historical surface quality rating are extracted from the process knowledge base to form a historical sample dataset.

[0009] Based on the historical sample dataset, a multi-objective fitness evaluation model is trained. The input of the multi-objective fitness evaluation model is the set of product specification feature parameters, the set of rubber compound formulation feature parameters, and the set of process parameters. The output is the predicted quality evaluation score, the predicted performance evaluation score, the predicted energy efficiency evaluation score, and the predicted surface defect risk score.

[0010] Define a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed, and construct the feasible domain of the set of decision variable parameters based on equipment physical constraints, process standard constraints, and experience linkage constraints.

[0011] The product specification feature parameter set of the current work order, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model are input into the multi-objective genetic algorithm. The multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than a set risk threshold. It performs global optimization within the feasible region and outputs a Pareto optimal solution set.

[0012] According to the preset production strategy, recommended process formulas are selected from the Pareto optimal solution set, and the recommended process formulas are sent to the manufacturing execution system and equipment control system for execution.

[0013] During production execution, real-time process data and product surface quality data are collected, and a surface quality assessment index is calculated based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or a key process parameter experiences uncontrollable drift, rapid local optimization is triggered, starting from the current execution parameters, to generate and issue process parameter fine-tuning increments.

[0014] By adopting the above technical solution, the rheological properties of the rubber compound are sensed online and the formulation characteristics are corrected in real time. This ensures that the generated process parameters closely match the actual state of the current production materials, solving the problem of production instability caused by batch fluctuations in the rubber compound. Simultaneously, a multi-objective fitness evaluation model trained on historical production data maps macroscopic statistical indicators to standardized evaluation scores, accurately predicting the performance of quality, efficiency, energy efficiency, and surface defect risks under different parameter combinations. Combined with a multi-objective genetic algorithm for global optimization within the feasible domain, this overcomes the shortcomings of single-objective optimization, achieving synergistic optimization and a triple balance of quality, efficiency, and energy consumption in the score dimension. Furthermore, the dynamic monitoring and rapid local optimization mechanism based on real-time surface quality data calculation of the evaluation index during production can respond promptly to production anomalies, curbing quality risks at their inception through parameter fine-tuning, further enhancing the stability and anti-interference capability of the production process.

[0015] Optionally, the calculation of the effective utilization rate of the historical work order equipment includes:

[0016] Based on the equipment status event log, the effective processing time of historical work orders is calculated, excluding startup and trial molding time.

[0017] Calculate the historical work order utilization rate, which is the percentage of the effective processing time of the historical work order to the planned production time.

[0018] Calculate the historical work order pass rate, which is the percentage of the number of qualified semi-finished products in the historical work orders to the total number of inspections;

[0019] Calculate the effective utilization rate of equipment for historical work orders, where the effective utilization rate of equipment for historical work orders = historical work order operation rate * historical work order qualification rate.

[0020] By adopting the above technical solutions, with effective processing time as the core statistical dimension, the system filters out valueless operating time such as equipment idling and waiting for materials, allowing the calculation of equipment utilization rate to truly reflect the actual proportion of time that the equipment creates value for production. At the same time, the qualified rate indicator is incorporated into the utilization rate calculation system, directly linking the equipment's production efficiency with output quality. This avoids the misleading effect of high utilization rate but low qualified rate, ensuring that the final effective utilization rate of the equipment can fully reflect the true performance of the equipment in terms of "effective output".

[0021] Optionally, the steps for training a multi-objective fitness evaluation model based on the historical sample dataset specifically include:

[0022] The parameter sets of product specification features, rubber compound formulation features, and historical process parameters in the historical sample dataset are used as training inputs.

[0023] The historical quality assessment score, historical performance assessment score, historical energy efficiency assessment score, and historical surface defect risk score in the historical sample dataset are used as training labels for the predicted quality assessment score, predicted performance assessment score, predicted energy efficiency assessment score, and predicted surface defect risk score, respectively.

[0024] The gradient boosting regression tree algorithm is used to train sub-models for predicting quality assessment scores, performance assessment scores, energy efficiency assessment scores, and surface defect risk scores, which together constitute a multi-objective fitness assessment model.

[0025] Optionally, performing global optimization within the feasible region includes:

[0026] An initial population is randomly generated within the feasible domain, and historically successful process parameters that are most similar to the current product specification feature parameter set and rubber compound formulation feature parameter set are retrieved from the process knowledge base and injected into the initial population as seed individuals.

[0027] For each candidate decision variable parameter set in the initial population, the multi-objective fitness evaluation model is invoked to calculate the corresponding prediction quality evaluation score, prediction effectiveness evaluation score, prediction energy efficiency evaluation score, and prediction surface defect risk score.

[0028] Based on the predicted surface defect risk score, individuals in the population are pre-screened, and individuals whose predicted surface defect risk score exceeds the set risk threshold are removed.

[0029] Based on the predicted quality assessment score, predicted efficacy assessment score, and predicted energy efficiency assessment score of the individuals retained after pre-screening, non-dominated ranking and crowding degree calculation are used to evaluate and select individuals in the population.

[0030] Perform crossover and mutation operations on the selected individuals, and ensure that the newly generated individuals are located within the feasible region by the constraint repair operator;

[0031] The selection, crossover, and mutation processes are iteratively executed until convergence, and the Pareto optimal solution set is output.

[0032] By adopting the above technical solutions, the blindness of the completely random initialization of the population in traditional genetic algorithms is broken. By injecting historically successful process parameters that best match the current production needs as seed individuals, the initial population has the foundation to converge towards high-quality solutions from the beginning, shortening the optimization cycle of the algorithm and reducing the risk of the algorithm getting trapped in local optima. The evaluation and selection mechanism of non-dominated sorting and crowding calculation can screen out high-quality individuals that take into account quality, efficiency and energy efficiency evaluation scores in scenarios with multiple conflicting objectives. This ensures the directionality of the population's convergence towards the Pareto front and maintains the diversity of solutions, avoiding the loss of one aspect due to single-objective optimization. The constraint repair operator after crossover and mutation ensures that all candidate solutions meet the requirements of equipment physics and process standards, so that the algorithm optimization always revolves around feasible production solutions without having to deal with a large number of infeasible solutions, further improving the optimization efficiency.

[0033] Optionally, based on a preset production strategy, recommended process formulations can be selected from the Pareto optimal solution set, including:

[0034] Set quality assessment score thresholds;

[0035] If the production strategy prioritizes quality, then among all solutions where the predicted quality assessment score is greater than the quality assessment score threshold, a comprehensive evaluation index is calculated based on the predicted performance assessment score and the predicted energy efficiency assessment score, and the solution with the optimal comprehensive evaluation index is selected as the recommended process formula.

[0036] If the production strategy prioritizes efficiency, then among the solutions whose predicted quality assessment scores are greater than the preset lower limit threshold for quality scores, the solution with the highest predicted performance assessment score is selected as the recommended process formula.

[0037] If the production strategy prioritizes cost, then among all solutions with predicted quality assessment scores greater than the quality assessment score threshold, the solution with the highest predicted energy efficiency assessment score is selected as the recommended process formulation.

[0038] Optionally, the steps to trigger rapid local optimization and generate process parameter fine-tuning increments specifically include:

[0039] Use the current execution parameters as the initial point for local optimization;

[0040] Construct a local optimization objective, which is to minimize the predicted surface defect risk score so that the actual surface quality assessment index falls back, or to minimize the deviation of key process parameters from the recommended set value so as to stabilize the production state.

[0041] Construct local optimization constraints, wherein the changes in the predicted quality assessment score and the predicted energy efficiency assessment score relative to the current value are less than a set range;

[0042] A fast search is performed within a preset neighborhood of the initial point to obtain the process parameter fine-tuning increment that optimizes the local optimization objective.

[0043] By adopting the above technical solution, optimization starts from the current execution parameters, eliminating the need for global optimization from scratch. This shortens the response time for parameter adjustments, enabling immediate reactions to production anomalies and minimizing quality risks and production fluctuations. Local optimization focuses on reducing surface defect risks or stabilizing key parameters, avoiding additional fluctuations caused by irrelevant parameter adjustments. Simultaneously, by limiting the range of changes in quality assessment scores and energy efficiency assessment scores, it ensures that parameter fine-tuning does not negatively impact core production performance indicators, maintaining overall production stability and economy while addressing current issues. This small-scale, precise fine-tuning approach balances production flexibility and controllability while avoiding production interruptions and parameter oscillations that global optimization might cause. It ensures that the production process maintains high efficiency and stability even in the face of unexpected situations such as rubber material fluctuations and minor equipment malfunctions, ultimately achieving continuous product quality stability and maximizing production efficiency.

[0044] Optionally, the process optimization method further includes:

[0045] After the work order is completed, the actual work order qualification rate, actual work order equipment effective utilization rate, actual work order unit output energy consumption and actual surface quality rating are collected and mapped to generate actual quality assessment score, actual efficiency assessment score, actual energy efficiency assessment score and actual surface defect risk score, which constitute the actual performance record.

[0046] After the actual performance records are stored in the process knowledge base, the model incremental update process is triggered;

[0047] Outlier detection and data cleaning are performed on the newly added actual performance records to remove invalid or abnormal data records caused by sensor failure or extreme operating conditions.

[0048] The newly added actual performance records after cleaning are merged with the historical training data buffer of the current multi-objective fitness evaluation model to form an incremental learning sample set;

[0049] An online learning algorithm is used to iteratively update the multi-objective fitness evaluation model with the incremental learning sample set. During the update process, new samples are given higher weights, while the original key parameters of the model are subject to regularization constraints.

[0050] After the incremental model update is completed, the updated model is integrated and fused with the original model to generate a new multi-objective fitness evaluation model.

[0051] By adopting the above technical solutions, the real results of the production process are promptly fed back into the process knowledge base through the full-dimensional performance data collection after the work order is completed, providing the model with the most vivid data that is closest to the actual production scenario. The outlier detection and data cleaning process can effectively filter the interference caused by atypical data such as sensor failures and extreme working conditions, ensuring the authenticity and reliability of incremental learning samples. The online learning algorithm, combined with the update strategy of assigning high weights to new samples and regularizing the original parameters, allows the model to quickly absorb the latest production experience and adapt to changes in current production conditions, while avoiding the problem of the model overfitting new data and losing the original mature knowledge, thus finding a precise balance between adaptability and stability. The fusion mechanism of new and old model parameters further smooths the performance fluctuations during the model iteration process, ensuring that the updated model can seamlessly connect with the original production process and will not affect the continuity of subsequent process parameter optimization.

[0052] Secondly, this application provides a process optimization system based on a multi-objective genetic algorithm, employing the following technical solution:

[0053] A process optimization system based on a multi-objective genetic algorithm includes:

[0054] The process knowledge base module is used to store historical production process data, including historical product specification feature parameter sets, historical rubber compound formula feature parameter sets, historical process parameter sets, historical work order pass rates, historical work order equipment effective utilization rates, historical work order unit output energy consumption, and historical surface quality ratings.

[0055] The online sensing and data acquisition module communicates with the manufacturing execution system and field sensors to receive work order instructions, acquire product specification feature parameter sets and rubber compound formula feature parameter sets, and perform online sensing of the rheological properties of the rubber compound to obtain actual rheological parameters and update the rubber compound formula feature parameter set.

[0056] The model training and maintenance module is used to train a multi-objective fitness evaluation model based on the historical sample dataset. The input of the multi-objective fitness evaluation model is a set of product specification feature parameters, a set of rubber compound formulation feature parameters, and a set of process parameters. The output is a predicted quality evaluation score, a predicted performance evaluation score, a predicted energy efficiency evaluation score, and a predicted surface defect risk score.

[0057] The optimization solution module is used to: define a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed; and construct the feasible region of the decision variable parameter set based on equipment physical constraints, process standard constraints, and experience linkage constraints; receive the product specification feature parameter set of the current work order, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model, and input them into the multi-objective genetic algorithm; the multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than a set risk threshold, and performs global optimization within the feasible region to output a Pareto optimal solution set;

[0058] The strategy decision-making and distribution module is used to select recommended process formulas from the Pareto optimal solution set according to the preset production strategy, and distribute the recommended process formulas to the manufacturing execution system and equipment control system for execution;

[0059] The online monitoring and local optimization module communicates with the equipment control system and online detection equipment. It is used to collect real-time process data and product surface quality data during production execution, and calculate the surface quality assessment index based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or when key process parameters drift uncontrollably, it triggers rapid local optimization starting from the current execution parameters, generates process parameter fine-tuning increments, and issues them.

[0060] Thirdly, this application provides a computer device that adopts the following technical solution:

[0061] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the process optimization method based on a multi-objective genetic algorithm as described in the first aspect.

[0062] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0063] A computer-readable storage medium storing a computer program capable of being loaded by a processor and executed as described in the first aspect, which is based on a multi-objective genetic algorithm for process optimization.

[0064] In summary, this application includes at least one of the following beneficial technical effects:

[0065] 1. This method uses online sensing of rubber rheological properties to correct formulation characteristics in real time, matching the actual batch status of materials, effectively solving the production instability problem caused by batch fluctuations of rubber. At the same time, it maps historical macroscopic statistical indicators into standardized quality, efficiency, energy efficiency, and risk assessment scores, and combines gradient boosting regression tree model and multi-objective genetic algorithm for global optimization. This makes the prediction of the multi-objective fitness assessment model more consistent with the intrinsic relationship between process parameters and output performance, breaking through the limitations of traditional single-objective and direct prediction of macroscopic statistics, and achieving a precise balance and synergistic optimization of product quality, equipment utilization efficiency, and production energy consumption in terms of score dimensions.

[0066] 2. By collecting surface quality data in real time during the production process and calculating the evaluation index, the system can accurately capture quality fluctuations and parameter drifts, triggering rapid local parameter fine-tuning to immediately curb production anomalies and quality risks. During the fine-tuning process, constraints are imposed based on predicted quality and energy efficiency evaluation scores to ensure that local parameter adjustments do not impair overall core efficiency. At the same time, the system achieves closed-loop iterative updates of the model through work order production data retrieval, score mapping, cleaning, and online incremental learning, continuously adapting to changes in production conditions and improving the anti-interference capability of the production process and the long-term accuracy and continuity of process optimization. Attached Figure Description

[0067] Figure 1 This is a first flowchart of an embodiment of the method of this application;

[0068] Figure 2 This is a second flowchart of an embodiment of the method of this application;

[0069] Figure 3 This is a third flowchart of an embodiment of the method of this application;

[0070] Figure 4 This is the fourth flowchart of an embodiment of the method of this application;

[0071] Figure 5 This is the fifth flowchart of an embodiment of the method of this application. Detailed Implementation

[0072] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-5 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0073] The first embodiment of this application discloses a process optimization method based on a multi-objective genetic algorithm. (Refer to...) Figure 1 The process optimization method includes S110-S170:

[0074] S110: Based on the work order instructions issued by the manufacturing execution system, obtain the product specification feature parameter set and the rubber compound formula feature parameter set; perform online sensing of the rheological properties of the rubber corresponding to the rubber compound formula feature parameter set, obtain the actual rheological parameters, and update the rubber compound formula feature parameter set.

[0075] S120, based on the updated set of rubber compound formulation feature parameters and product specification feature parameters, extracts the parameter set of historical process parameters, historical work order qualification rate, historical work order equipment effective utilization rate, historical work order unit output energy consumption and historical surface quality rating from the process knowledge base to form a historical sample dataset.

[0076] S130, based on historical sample dataset, trains a multi-objective fitness evaluation model. The input of the multi-objective fitness evaluation model is the set of product specification feature parameters, the set of rubber compound formulation feature parameters, and the set of process parameters. The output is the predicted quality evaluation score, the predicted performance evaluation score, the predicted energy efficiency evaluation score, and the predicted surface defect risk score.

[0077] S140 defines a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed, and constructs the feasible domain of the decision variable parameter set based on equipment physical constraints, process standard constraints, and experience linkage constraints.

[0078] S150: Input the current work order's product specification feature parameter set, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model into the multi-objective genetic algorithm; The multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than the set risk threshold. It performs global optimization within the feasible region and outputs the Pareto optimal solution set;

[0079] S160, based on the preset production strategy, selects recommended process formulas from the Pareto optimal solution set and sends the recommended process formulas to the manufacturing execution system and equipment control system for execution;

[0080] S170 collects real-time process data and product surface quality data during production execution, and calculates the surface quality assessment index based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or when key process parameters drift uncontrollably, it triggers rapid local optimization starting from the current execution parameters, generates process parameter fine-tuning increments, and issues them.

[0081] Specifically, for step S110, a communication connection is established with the Manufacturing Execution System (MES) via Industrial Ethernet. The OPC UA protocol is used to read work order instructions issued by the MES in real time. These instructions contain basic information such as work order number, product model, and production batch. Based on a pre-set product specification feature library, the corresponding specification feature parameter set is extracted. This parameter set includes key parameters such as the target size of the semi-finished product, hardness, and tensile strength. Each parameter corresponds to a standardized value, forming the product specification feature parameter set. The rubber compound formulation feature parameter set is obtained from the MES's formulation management module and includes parameters such as the mass ratio, initial viscosity, and Mooney viscosity of each component, including natural rubber, carbon black, and vulcanizing agent, constituting the rubber compound formulation feature parameter set.

[0082] An online rheometer is installed at the discharge port of the rubber mixing mill to sense the rheological properties of the rubber compound in real time. The sampling frequency is set to 1 time / minute to obtain actual rheological parameters, including Mooney viscosity and vulcanization curve. If the actual Mooney viscosity is detected to be outside the range of the initial formulation feature parameter set, the Mooney viscosity parameter in the rubber compound formulation feature parameter set is updated to the actual value to ensure that the rubber compound formulation feature parameter set is consistent with the actual state of the rubber compound.

[0083] Reference Figure 2 In S120, the calculation of the effective utilization rate of equipment in historical work orders includes S210-S240:

[0084] S210, based on the equipment status event log, calculates the effective processing time of historical work orders, excluding start-up and trial molding time;

[0085] S220, calculate the historical work order utilization rate, which is the percentage of the effective processing time of the historical work order to the planned production time;

[0086] S230, calculate the historical work order pass rate, which is the percentage of the number of qualified semi-finished products in the historical work orders / the total number of inspections;

[0087] S240, calculate the effective utilization rate of equipment in historical work orders. Effective utilization rate of equipment in historical work orders = operating rate of historical work orders * qualification rate of historical work orders.

[0088] Specifically, for step S120, based on the updated set of rubber compound formulation feature parameters and the set of product specification feature parameters, the process knowledge base is retrieved using structured query language (SQL). This process knowledge base is stored in a relational database (such as MySQL), and is classified and indexed by product specifications and rubber compound formulation. Historical work order data with a similarity of ≥ set similarity threshold (such as 85%) with the current parameter set are extracted to construct a historical sample dataset.

[0089] The historical process parameter set includes actual operating parameters such as the mixing mill speed, mixing mill roller temperature, extruder heating temperatures at each stage (e.g., stage 1 120-140℃, stage 2 140-160℃, stage 3 160-180℃), and traction speed. The average value is taken as the dimension value of the parameter set based on the work order cycle. Historical work order pass rate, historical work order equipment utilization rate, historical work order unit output energy consumption, and historical surface quality rating are all statistically analyzed by work order and used as the original performance indicators for subsequent model training. These are used to map and generate corresponding historical quality assessment scores, historical efficiency assessment scores, historical energy efficiency assessment scores, and historical surface defect risk scores in S130.

[0090] The calculation of the effective utilization rate of historical work orders follows the work order-level statistical standard: First, the time window (work order start time to end time) corresponding to the current historical work order is filtered from the equipment status event log of the MES. Based on the three status fields of Run, Mode, and Trial in the log, the effective processing time is calculated. Only when Run=1, Mode='Production', and Trial=0, the corresponding time slice is included in the effective processing time, excluding the time corresponding to startup (Mode='Startup'), trial molding (Trial=1), and downtime. For example, if a historical work order has a planned production time of 8 hours, and the effective processing time after filtering is 7.2 hours, then the historical work order utilization rate = 7.2 / 8 × 100% = 90%. The historical work order pass rate is obtained through the quality inspection module of the MES. The total number of semi-finished products inspected and the number of qualified products are counted. For example, if the total number of inspections is 200 and the number of qualified products is 180, then the pass rate = 180 / 200 × 100% = 90%. Finally, the effective utilization rate of historical work order equipment = 90% × 90% = 81%.

[0091] The energy consumption per unit output of historical work orders is calculated by dividing the total energy consumption of the rubber mixing section and the extrusion section of the work order (obtained by integrating the equipment power meter, in kWh) by the actual output of the work order (in kg). For example, if the total energy consumption is 640 kWh and the output is 8000 kg, then the energy consumption per unit output = 640 / 8000 = 0.08 kWh / kg.

[0092] Historical surface quality ratings are output by a machine vision inspection system, categorized into three levels: A (no defects), B (minor defects), and C (serious defects). Defect area percentage data is recorded simultaneously to support continuous quantification during subsequent model training. Each historical sample includes an input set of "product specification feature parameters + rubber compound formulation feature parameters + historical process parameters," and a label set of "historical quality assessment score + historical performance assessment score + historical energy efficiency assessment score + historical surface defect risk score." This results in a historical sample dataset containing over 1000 valid samples.

[0093] If there is insufficient historical data in the process knowledge base (e.g., the number of valid samples is less than 100), it is determined to be a cold start state. The training and global optimization of the multi-objective fitness evaluation model are suspended, and the default process parameters in the process knowledge base are issued for execution. Production data is continuously accumulated until the sample quantity requirement is met.

[0094] S130, the specific steps for training a multi-objective fitness evaluation model based on historical sample datasets include:

[0095] The training input consists of the set of product specification feature parameters, rubber compound feature parameters, and historical process parameters from the historical sample dataset. The training labels are the historical quality assessment scores, historical performance assessment scores, historical energy efficiency assessment scores, and historical surface defect risk scores from the historical sample dataset, respectively. The gradient boosting regression tree algorithm is used to train sub-models for predicting the quality assessment scores, performance assessment scores, energy efficiency assessment scores, and surface defect risk scores, which together constitute the multi-objective fitness assessment model.

[0096] Specifically, for step S130, the set of product specification feature parameters, rubber compound formulation feature parameters, and historical process parameters from the historical sample dataset are used as joint training inputs. First, the input data is standardized, and the Min-Max normalization method is used to map all parameters to the [0,1] interval to eliminate the influence of dimensions. Four key score indicators from the historical sample dataset are used as training labels for the corresponding prediction targets. The historical quality assessment score is generated by linearly mapping the historical work order pass rate (e.g., a pass rate of 90%-100% is mapped to a score of 0.9-1.0), the historical efficiency assessment score is generated by linearly mapping the historical work order equipment utilization rate, and the historical energy efficiency assessment score is generated by inversely mapping the historical work order unit output energy consumption (the lower the energy consumption, the higher the score, e.g., energy consumption of 0.08-0.06 kWh / kg is mapped to a score of 0.8-1.0). The historical surface defect risk score is generated by continuously processing the rating based on the defect area ratio: rating A (no defects) corresponds to a risk score of 0.0, rating B (minor defects) corresponds to a risk score range of [0.1, 0.4] (linearly mapped according to the defect area ratio), and rating C (serious defects) corresponds to a risk score range of [0.5, 1.0], thus providing continuously distributed training labels.

[0097] The multi-objective fitness evaluation model is constructed using the Gradient Boosting Regression Tree (GBRT) algorithm, implemented using the XGBoost framework. Four sub-models are trained, each corresponding to a prediction objective. During training, uniform hyperparameters are set: learning rate 0.1, decision tree depth 5, minimum number of samples per leaf node 10, and number of iterations 100. Five-fold cross-validation is used to avoid overfitting. For example, the sub-model predicting quality assessment scores uses the input parameter set as the independent variable and historical quality assessment scores as the dependent variable. After training, it is validated using a validation set to ensure that the model's coefficient of determination R² ≥ 0.85 and mean squared error (MSE) ≤ 0.005. The sub-model predicting surface defect risk scores focuses on optimizing the sample imbalance problem, assigning a 1.5x weight to C-level (severe defect) samples to ensure the model's prediction accuracy for high-risk samples. After the four sub-models are trained, they are integrated through a model fusion module to form the multi-objective fitness evaluation model. Inputting any set of candidate process parameters will simultaneously output four predicted scores.

[0098] For step S140, the decision variable parameter set explicitly includes the core parameters of rubber mill speed (x1), rubber mill roller temperature (x2), extruder section heating temperature (xn), and traction speed (x(n+1)), forming the decision variable parameter set [x1,x2,x3,...,x(n+1)]. All parameters are encoded with real numbers, which meets the solution requirements of the genetic algorithm.

[0099] The feasible domain is constructed based on three types of constraints to ensure the rationality and feasibility of decision variables: Physical constraints are determined by the technical parameters provided by the equipment manufacturer, such as the mixing mill speed range of 50-150 r / min, roller temperature range of 80-120℃, the upper limit of heating temperature for each section of the extruder not exceeding 180℃, and the traction speed range of 3-8 m / min. Parameters exceeding these ranges will lead to equipment failure or damage. Process standard constraints refer to the mixing-extrusion process standard table established by the company, such as the vulcanization temperature of the rubber compound needing to be controlled at 100-110℃ (corresponding to roller temperature x2), and the upper limit of heating temperature for each section of the extruder... The temperature difference should not exceed 20℃ (e.g., |x3-x4|≤20, |x4-x5|≤20); the empirical linkage constraint is based on the summary of on-site production experience. For example, the traction speed x6 is positively correlated with the temperature x5 of the three sections of the extruder. When x5=160℃, the reasonable range of x6 is 3.4-4.5m / min. If x5 increases to 170℃, the reasonable range of x6 can be adjusted to 3.8-4.8m / min, and vice versa. The linkage relationship is determined by the linear fitting formula to determine the benchmark value, i.e., x6base=0.03*x5-0.8, and the traction speed is allowed to fluctuate around the benchmark value.

[0100] The three types of constraints are structured and represented by a system of inequalities, ultimately constructing a set of decision variable parameters. This ensures that all decision variables within the feasible region meet actual production requirements. For example, x1∈[50,150], x2∈[80,120], x3∈[120,140], x4∈[140,160], x5∈[160,180], x6∈[3,8], and satisfies |x3-x4|≤20, |x4-x5|≤20, x6∈[0.03*x5-1.3, 0.03*x5-0.3].

[0101] Reference Figure 3 In S150, the global optimization solution within the feasible region includes S310-S360:

[0102] S310, randomly generate an initial population within the feasible domain, and retrieve historical successful process parameters that are most similar to the current product specification feature parameter set and rubber compound formulation feature parameter set from the process knowledge base, and inject them as seed individuals into the initial population;

[0103] S320: For each candidate decision variable parameter set in the initial population, call the multi-objective fitness evaluation model to calculate the corresponding prediction quality evaluation score, prediction effectiveness evaluation score, prediction energy efficiency evaluation score, and prediction surface defect risk score.

[0104] S330, based on the predicted surface defect risk score, pre-screens individuals in the population and removes individuals whose predicted surface defect risk score exceeds the set risk threshold;

[0105] S340, based on the predicted quality assessment score, predicted efficacy assessment score, and predicted energy efficiency assessment score of the individuals retained after pre-screening, non-dominated ranking and crowding calculation are used to evaluate and select individuals in the population.

[0106] S350 performs crossover and mutation operations on the selected individuals and ensures that the newly generated individuals are within the feasible region through the constraint repair operator;

[0107] S360 iteratively executes the selection, crossover, and mutation process until convergence, outputting the Pareto optimal solution set.

[0108] Specifically, for step S150, the product specification feature parameter set of the current work order, the updated rubber compound formula feature parameter set, and the feasible region constructed in S140 are obtained. These three types of data, along with the multi-objective fitness evaluation model trained in S130, are input into the multi-objective genetic algorithm. The multi-objective genetic algorithm selected is NSGA-II (Non-dominated sorting genetic algorithm II), which has the advantages of fast convergence speed and uniform solution set distribution, meeting the needs of multi-objective optimization. Its optimization objectives are clearly set to three: maximizing the prediction quality evaluation score (max QScore), maximizing the prediction efficiency evaluation score (max EScore), and maximizing the prediction energy efficiency evaluation score (max EnScore). The three objectives have equal weights and no additional weight coefficients need to be set.

[0109] The initial population is generated using a "random generation + seed injection" approach. First, 100 candidate individuals are randomly generated within the feasible region Ω, each corresponding to a set of decision variable parameters. Random generation ensures a uniform distribution of parameters, avoiding concentration in any particular interval. Then, historical successful process parameters (defined as those with a historical work order pass rate ≥90%, equipment utilization rate ≥80%, and unit output energy consumption ≤0.08 kWh / kg) with ≥90% similarity to the current product specification feature parameter set and rubber compound formulation feature parameter set are retrieved from the process knowledge base. The five parameters with the highest similarity are selected as seed individuals and injected into the initial population, forming an initial population of 105 individuals together with the randomly generated individuals. The addition of seed individuals improves the quality of the initial population and shortens the algorithm's convergence time.

[0110] For each candidate decision variable parameter set in the initial population, the multi-objective fitness evaluation model is sequentially invoked to calculate fitness. Before calculation, the candidate decision variable parameter sets need to be standardized to maintain consistency with the input format during model training, avoiding prediction errors caused by dimensional issues. The calculation process adopts a batch processing method, inputting 10 candidate individuals at a time. The model simultaneously outputs four prediction indicators for each individual: predicted quality assessment score (QScore), predicted efficiency assessment score (EScore), predicted energy efficiency assessment score (EnScore), and predicted surface defect risk score (RScore). For example, if the decision variable parameter set of a candidate individual is [105,108,132,148,168,5.4], after inputting it into the model, the output prediction results are QScore=0.93, EScore=0.85, EnScore=0.88, and RScore=0.12. To ensure computational robustness, each sub-model of the multi-objective fitness evaluation model uses 5-fold cross-validation during the training phase to generate 5 base learners. During prediction, each candidate individual is input into one of five base learners, and the average of the five outputs is taken as the final prediction result to eliminate the prediction variance of a single model. At the same time, the prediction time for each individual is recorded to ensure that the computation time for a single individual is ≤0.5s, which meets the algorithm iteration efficiency requirements.

[0111] Based on the individuals retained after S330 pre-screening and their three core optimization objectives (QScore, EScore, EnScore), the core logic of the NSGA-II algorithm is used to perform non-dominated ranking and crowding calculation on the individuals in the population, thus completing the evaluation and selection of individuals. High-R-value individuals that are removed do not participate in non-dominated ranking, ensuring that the evolutionary direction does not evolve towards high-risk areas of surface defects. The core of non-dominated ranking is to divide the individuals in the population into different non-dominated layers according to their dominance relationships. If individual A's three optimization objectives are all no worse than individual B's, and at least one objective is better than individual B's, then individual A dominates individual B, and the dominated individual moves to the next layer. The first layer is the optimal non-dominated layer, containing all individuals not dominated by other individuals; the second layer contains individuals dominated by individuals in the first layer but not dominated by other individuals, and so on. For example, in a population, individual A (QScore=0.93, EScore=0.85, EnScore=0.88) and individual B (QScore=0.91, EScore=0.83, EnScore=0.86) dominate individual B. Individual A enters the first layer, and individual B enters the second layer.

[0112] Crowding degree calculation is used to measure the diversity of individuals within the same non-dominated layer. It calculates the Euclidean distance between each individual and its neighbors in the three target spaces. The larger the distance, the greater the crowding degree, which represents better individual diversity and can prevent the solution set from converging to a local optimum. The selection process follows the principle of "non-dominated layer priority + crowding degree priority". It starts by selecting individuals from the first layer. If the number of individuals in the first layer is less than the preset population size (100), it continues to select from the second layer until the population size is reached. During selection, individuals with high crowding degree are given priority to ensure the diversity and quality of the population.

[0113] For the selected individuals, crossover and mutation operations are performed sequentially to generate new offspring individuals. Simultaneously, a constraint repair operator ensures that the newly generated individuals always remain within the feasible region Ω, avoiding the generation of invalid individuals. The crossover operation uses the SBX (Simulated Binary Crossover) algorithm, suitable for individuals with real-number encoding. The crossover probability is set to 0.8, and the crossover factor is set to 1.2. During crossover, two parent individuals are randomly selected, and their decision variable parameter sets are crossovered to generate two offspring individuals.

[0114] For example, parent individuals A [105,108,132,148,168,5.4] and parent individuals B [110,110,135,150,170,5.6] cross over to produce offspring individuals C [107,109,133,149,169,5.5] and offspring individuals D [108,109,134,149,169,5.5].

[0115] The mutation operation uses a multinomial mutation algorithm with a mutation probability of 0.1 and a mutation factor of 0.5. A decision variable of an individual is randomly selected and a small perturbation is performed within its feasible region. For example, if the x1 (speed of the rubber mixing mill) of the offspring individual C is 107 r / min, it becomes 106 r / min after mutation, ensuring that the mutation amplitude is within a reasonable range.

[0116] The constraint repair operator repairs invalid individuals (individuals outside the feasible region) that may appear after crossover and mutation. The repair principle is "nearest repair + linkage repair": For individuals with a single parameter exceeding the limit, they are repaired to the boundary value of the feasible region. For example, after the mutation of x5 (extruder three-stage temperature) of offspring individual D, it is 185℃, which exceeds the upper limit of 180℃, so it is repaired to 180℃. For individuals that violate linkage constraints, the traction speed x(n+1) is repaired first. Based on the actual value of xn, the baseline value is calculated through the linkage formula x(n+1)=0.03*xn-0.8, and the repair value is taken within the range of 0.5 above and below the baseline value. x(n+1) is adjusted to a reasonable range. For example, when x5=180℃, x6 is repaired to 4.6-5.6m / min to ensure that the repaired individuals fully meet the three types of constraints.

[0117] The algorithm iteratively executes the selection of S340 and the crossover and mutation processes of S350 until the convergence condition is met, outputting the Pareto optimal solution set. Two convergence conditions are set: first, the number of iterations reaches a preset maximum value (100 generations); second, the change in the Pareto optimal solution set for 10 consecutive generations is less than 0.01 (the change is determined by calculating the average Euclidean distance between the solutions of consecutive generations). Iteration stops when either condition is met. During iteration, the Pareto optimal solution set is recorded every 10 generations to monitor the convergence trend. If excessive fluctuations occur, the crossover and mutation probabilities are adjusted appropriately (e.g., increasing the mutation probability to 0.12) to ensure stable convergence. After iteration stops, the Pareto optimal solution set is output, containing 20-30 optimal individuals. Each individual corresponds to a set of process parameters, and there is no dominance relationship between any two individuals; that is, it is impossible to find an individual that is superior to another individual in all three optimization objectives. For example, in the Pareto optimal solution set, there are two individuals, individual 1 (QScore=0.95, EScore=0.82, EnScore=0.92) and individual 2 (QScore=0.92, EScore=0.88, EnScore=0.89). Individual 1 is better in terms of quality and energy efficiency, while individual 2 is better in terms of efficiency. The two constitute a trade-off relationship.

[0118] If the Pareto optimal solution set output by NSGA-II is empty, the constraints are gradually relaxed (e.g., the similarity threshold is lowered by 5% or the feasible region boundary is expanded by 5%), and the global optimization solution is re-executed; if it is still empty, the individual with the best prediction comprehensive evaluation score (EScore+EnScore) within the feasible region is selected as the compromise solution and distributed.

[0119] S160, based on the preset production strategy, selects recommended process formulations from the Pareto optimal solution set, including:

[0120] Set a quality assessment score threshold; if the production strategy is quality-first, then among all solutions with predicted quality assessment scores greater than the quality assessment score threshold, calculate a comprehensive evaluation index based on the predicted performance assessment score and the predicted energy efficiency assessment score, and select the solution with the optimal comprehensive evaluation index as the recommended process formula; if the production strategy is efficiency-first, then among solutions with predicted quality assessment scores greater than the preset lower limit of the quality score threshold, select the solution with the highest predicted performance assessment score as the recommended process formula; if the production strategy is cost-first, then among all solutions with predicted quality assessment scores greater than the quality assessment score threshold, select the solution with the highest predicted energy efficiency assessment score as the recommended process formula.

[0121] Specifically, for step S160, a quality assessment score threshold is set. This threshold is determined based on the company's production standards and product applications, and is generally set to 0.9. The threshold can be manually adjusted through the MES system to adapt to different production needs.

[0122] Based on the company's current production strategy, recommended process formulations are selected from the Pareto optimal solution set. The specific implementations of the three production strategies are as follows: **Quality-First Strategy:** First, all solutions with a predicted quality assessment score QScore > threshold (0.9) are selected. Among the selected solutions, a comprehensive evaluation index (EScore + EnScore; the larger the sum, the better the overall efficiency and energy consumption performance) is calculated for each solution. The solution with the highest comprehensive evaluation index is selected as the recommended process formulation. **Efficiency-First Strategy:** Among solutions that meet the preset lower limit of the quality score (e.g., QScore ≥ 0.85 to avoid excessively low quality), the solution with the highest predicted efficiency assessment score EScore is selected from the Pareto optimal solution set as the recommended process formulation. **Cost-First Strategy:** Similarly, solutions with a QScore > 0.9 are first selected. Among the selected solutions, the solution with the highest predicted energy efficiency assessment score EnScore is selected as the recommended process formulation. After selection, detailed parameters of the recommended process formulation are output, along with the corresponding three predicted scores, for process engineers' reference.

[0123] The selected recommended process formula is mapped according to the field requirements of the MES system and the equipment control system to build a data package. The data package includes the set values ​​of decision variables, upper and lower limits (set at ±5% of the feasible region for early warning during equipment operation), and corresponding prediction score indicators (QScore, EScore, EnScore, RScore).

[0124] The data packet distribution process adopts a three-level workflow: "MES review → Equipment control system reception → Parameter confirmation". First, the data packet is sent to the MES system, where process engineers review and confirm the rationality of the parameters. After the review is passed, the MES system forwards the data packet to the equipment control system (PLC) via industrial Ethernet. After receiving the data packet, the equipment control system automatically verifies whether the parameters are within the feasible range. If the verification is successful, the parameters are written into the corresponding equipment control module, and the operating parameters of the rubber mixing mill and extruder are adjusted, waiting for the production instruction to start execution. If the verification fails, the equipment control system sends out an error message, and the process engineer readjusts the parameters and distributes the data packet again.

[0125] Reference Figure 4 S170, the steps that trigger rapid local optimization and generate process parameter fine-tuning increments specifically include S410-S440:

[0126] S410, using the current execution parameters as the initial point for local optimization;

[0127] S420, construct local optimization objectives. The local optimization objectives are to minimize the predicted surface defect risk score so that the actual surface quality assessment index falls back, or to minimize the deviation of key process parameters from the recommended set values ​​so as to stabilize the production status.

[0128] S430, Construct local optimization constraints, where the change in the predicted quality assessment score and the predicted energy efficiency assessment score relative to the current value is less than a set range;

[0129] S440 performs a fast search within a preset neighborhood of the initial point to obtain the process parameter fine-tuning increment that optimizes the local optimization objective.

[0130] Specifically, for step S170, a real-time data acquisition system is constructed during production execution. Sensors (speed sensor, temperature sensor, power sensor) collect real-time operating parameters of the mixing mill and extruder at a sampling frequency of once per minute. The collected data includes real-time process data such as mixing mill speed, roller temperature, extruder section temperatures, and traction speed. Simultaneously, a machine vision inspection device (selected model: Keyence IV2 series) is installed at the extruder outlet to detect the surface quality of the semi-finished product in real time. An image recognition algorithm calculates the surface quality index, which ranges from 0 to 1, where 0 represents no defects and 1 represents severe defects. A warning threshold of 0.3 is set (adjustable according to product requirements). The criterion for determining uncontrollable drift of key process parameters is: a parameter deviating from the recommended set value by more than ±5% for a duration of ≥5 minutes. For example, if the recommended mixing mill speed is 105 r / min, and the speed remains below 99.75 r / min or above 110.25 r / min for 5 consecutive minutes, it is considered uncontrollable drift. When the surface quality index is detected to exceed 0.3, or when a critical process parameter drifts uncontrollably, a local optimization process is immediately triggered. The process parameters currently being executed by the equipment are used as the starting point for local optimization, ensuring that local optimization is based on the current production status and avoiding production interruptions caused by parameter mutations.

[0131] Based on the detected anomaly type, a corresponding local optimization objective is constructed to ensure that the objective is clear and quantifiable. If the anomaly type is that the actual surface quality assessment index exceeds the warning threshold, the local optimization objective is set to minimize the predicted surface defect risk score RScore. This is achieved by guiding the process parameters to be fine-tuned towards the low-risk range, causing the actual surface quality assessment index to fall back below the warning threshold. If the anomaly type is that the critical process parameter drifts uncontrollably (such as the rubber mixing mill speed drift), the local optimization objective is set to minimize the sum of squared normalized offsets of the critical process parameters relative to the recommended settings, i.e., the objective function is... , where xi is the current parameter and xiref is the recommended setting value. The production status is stabilized by suppressing parameter offset, while taking into account the decline in the predicted quality assessment score and the predicted performance assessment score.

[0132] To avoid a decline in overall production performance due to localized optimization, localized optimization constraints are established. These constraints require that the changes in the predicted quality assessment score (QScore) and the predicted energy efficiency assessment score (EnScore) relative to their current values ​​be within a set range of ±3%. Specifically, if the current predicted QScore is 0.93 and EnScore is 0.88, the localized optimization constraints are 0.9021 ≤ QScore ≤ 0.9579 and 0.8536 ≤ EnScore ≤ 0.9064, ensuring that quality and energy consumption do not fluctuate significantly during localized optimization. These constraints are based on the company's production baseline. If the predicted QScore or EnScore exceeds the constraint range during optimization, the optimization direction is immediately terminated, and the optimization strategy is adjusted to prioritize the stability of overall production performance. Furthermore, the constraints can be manually adjusted via the MES system according to actual production conditions to adapt to the production needs of different products.

[0133] A preset neighborhood is constructed centered on a given initial point. The neighborhood range is determined based on the sensitivity of the parameters. For sensitive parameters (such as roller temperature and traction speed), the neighborhood range is set to ±3%, while for insensitive parameters (such as rubber mixing mill speed), the neighborhood range is set to ±5%. For example, for an initial point x1 = 104 r / min, the neighborhood range is 98.8-109.2 r / min; for x2 = 105℃, the neighborhood range is 101.85-108.15℃. A gradient-based heuristic discrete search algorithm is used for fast search. The learning rate is set to 0.05, and the number of iterations is set to 20. Each iteration searches within the neighborhood for the parameter increment that optimizes the local target. The step size of the increment is determined based on the parameter type: 1 r / min for speed, 1℃ for temperature, and 0.1 m / min for traction speed. For example, in the case of a surface quality assessment index exceeding the warning level, the initial predicted RScore is 0.35. Through iterative search using the gradient descent algorithm, the fine-tuning increments are obtained as x2 = +1℃ and x6 = +0.2m / min. The adjusted parameters are [x1 = 104r / min, x2 = 106℃, x3 = 129℃, x4 = 144℃, x5 = 164℃, x6 = 5.3m / min]. At this point, the predicted RScore is 0.19, which satisfies the local optimization objective. Meanwhile, QScore = 0.925 and EnScore = 0.875, both within the constraints. This fine-tuning increment is the final solution result, which is sent to the equipment control system for execution, achieving rapid local optimization.

[0134] If the local optimization fails to find a solution that satisfies the local optimization constraints (variation range less than ±3%) after iterative search within the preset neighborhood, the local optimization is deemed to have failed, automatic fine-tuning is stopped and a system alarm is triggered, and manual intervention is required to adjust the process parameters.

[0135] Reference Figure 5 Furthermore, the process optimization method also includes S510-S560:

[0136] S510 After the work order is completed, the actual work order qualification rate, actual work order equipment effective utilization rate, actual work order unit output energy consumption and actual surface quality rating are collected and mapped to generate actual quality assessment score, actual efficiency assessment score, actual energy efficiency assessment score and actual surface defect risk score, which constitute the actual performance record.

[0137] S520 triggers the incremental update process of the model after the actual performance record is stored in the process knowledge base;

[0138] S530 performs outlier detection and data cleaning on newly added actual performance records, removing invalid or abnormal data records caused by sensor failure or extreme operating conditions.

[0139] S540 merges the cleaned new actual performance records with the historical training data buffer of the current multi-objective fitness evaluation model to form an incremental learning sample set;

[0140] S550 employs an online learning algorithm to iteratively update the multi-objective fitness evaluation model using an incremental learning sample set. During the update process, new samples are given higher weights, while the original key parameters of the model are subject to regularization constraints.

[0141] S560 integrates and merges the updated model with the original model after the incremental model update is completed, generating a new multi-objective fitness evaluation model.

[0142] Specifically, for step S510, after the work order production is completed, the actual performance record collection process is initiated. The collected data comes from the MES system, equipment control system, and quality inspection system, ensuring the authenticity and accuracy of the data. The actual work order pass rate is statistically analyzed through the quality inspection system, using a combination of manual and machine vision inspection. The total number of inspections and the number of qualified semi-finished products for the work order are counted. For example, if the total number of inspections is 1000 and 940 are qualified, then the actual work order pass rate = 940 / 1000 × 100% = 94%. The calculation of the actual work order equipment utilization rate is consistent with S120. The effective processing time, operating rate, and pass rate of the work order are recalculated. For example, if the effective processing time is 7.5 hours, the planned production time is 8 hours, the operating rate is 93.75%, and the actual pass rate is 94%, then the actual equipment utilization rate = 93.75% × 94% = 88.125%. The actual unit output energy consumption of a work order is calculated by integrating historical data from the equipment power meter. This involves calculating the total energy consumption of the rubber mixing and extrusion sections and dividing it by the actual output of the work order. For example, if the total energy consumption is 780 kWh and the output is 10,000 kg, then the actual unit output energy consumption = 780 / 10,000 = 0.078 kWh / kg. The actual surface quality rating is determined by the quality inspection department based on the surface defects of the semi-finished products, and is divided into three levels: A, B, and C, which are simultaneously quantified into corresponding scores. These indicators are compiled into an actual performance record, including the work order number, production time, actual quality assessment score, actual efficiency assessment score, actual energy efficiency assessment score, and actual surface defect risk score, forming a standardized record format.

[0143] For step S520, the actual performance records of S510 are reviewed and confirmed by the process engineer (after confirming that the data is normal and complete) and then stored in the historical sample database of the process knowledge base. They are stored together with the original historical samples and classified and indexed by work order number, product specifications, and rubber compound formula for easy retrieval and use later.

[0144] After actual performance records are stored in the process knowledge base, the system automatically triggers the incremental model update process. The triggering method is "event-triggered," meaning that when the number of newly added actual performance records in the knowledge base reaches a certain number (e.g., 5 or more), the update process is automatically initiated. If fewer than 5 new records are added, they are temporarily stored in a buffer until a total of 5 records are accumulated before triggering the update, thus avoiding frequent updates that could impact system performance. Simultaneously, the system sends model update notifications to process engineers, informing them of the update start time, the number of new samples, and other information, facilitating engineer monitoring of the update process.

[0145] For step S530, outlier detection and data cleaning are performed on the newly added actual performance records to ensure the quality of the incremental learning sample set and avoid outlier data affecting the model update effect. Outlier detection adopts the 3σ criterion, first calculating the mean of each indicator (actual quality assessment score, actual performance assessment score, actual energy efficiency assessment score, and actual surface defect risk score). with standard deviation When the indicator value of a certain record exceeds When the data falls within the specified range, it is considered an outlier. Invalid data includes: data missing due to sensor malfunction (e.g., energy consumption data is 0), data generated under extreme conditions (e.g., a sudden drop in pass rate due to a power outage), and data entry errors causing anomalies (e.g., a pass rate exceeding 100%). Data cleaning methods include: removing outliers and invalid data; supplementing missing data using linear interpolation (e.g., if a record lacks a surface quality rating, linear interpolation is performed using the ratings of two adjacent work orders); and correcting erroneous data entry to ensure that each cleaned record is complete, accurate, and valid.

[0146] For step S540, a historical training data buffer is constructed to store the training data of the current multi-objective fitness evaluation model. The buffer capacity is set to 1000 records, and a "first-in, first-out" (FIFO) principle is adopted. When the buffer reaches its capacity limit, the oldest 100 records are deleted to ensure the timeliness of the buffer data. The newly added actual performance records cleaned in S530 are merged with the data in the historical training data buffer. Deduplication is performed during the merging process. If a new record has a duplicate work order number with a record in the buffer (e.g., a duplicate retrieval record), the latest record is retained, and duplicate records are removed. After merging, an incremental learning sample set is formed. The format of this sample set is consistent with the sample format used during the original model training, including input features (product specification feature parameter set, rubber compound feature parameter set, and actual process parameter set) and labels (actual performance indicators) to ensure compatibility for subsequent model updates.

[0147] For step S550, an online learning algorithm is used to iteratively update the multi-objective fitness evaluation model. The online learning algorithm selected is the Online Gradient Boosting (OLGB) algorithm, which is suitable for incremental learning scenarios and can quickly update the model using new samples without retraining the entire model, thus improving update efficiency. During the update process, to highlight the timeliness of new samples (new samples better reflect changes in current production conditions), new samples are assigned higher weights, such as setting the weight of new samples to 1.5 and the weight of historical buffer samples to 1.0, ensuring that new samples have a greater impact on model parameters. At the same time, to prevent model overfitting, regularization constraints are applied to the original key parameters of the model (such as decision tree depth and learning rate) using L2 regularization with a regularization parameter λ=0.1. By penalizing the absolute value of model parameters, excessively large parameters are limited, thereby improving the model's generalization ability. The iterative update process is as follows: each iteration uses 100 samples from the incremental learning sample set, and the number of iterations is set to 50. After each iteration, the prediction accuracy (R² and MSE) of the model is calculated. If the accuracy reaches the preset standard (R²≥0.85), the iteration stops; otherwise, the iteration continues until the accuracy reaches the standard.

[0148] For step S560, after the incremental model update is completed, the updated model and the original model are combined into an integrated model. During prediction, the outputs of the new and old models are weighted and averaged. The formula is: final predicted value = w × new model predicted value + (1-w) × old model predicted value, where w is the dynamic fusion weight, which is determined based on the proportion of the coefficient of determination R² of the new and old models on the validation set. The calculation formula is w = Rnew² / (Rnew² + Rold²), ensuring that the model with better prediction performance occupies a higher weight in the fusion.

[0149] After fusion, a new multi-objective fitness evaluation model is generated. The performance of the new model is validated using a validation set (20% of the samples drawn from the incremental learning set) to test its prediction accuracy, ensuring that the new model's accuracy is no lower than the original model and that its prediction stability is better. If validation passes, the integrated new model is saved to the model library, replacing the original model, for use in subsequent work order process optimization. If validation fails, the process returns to step S550, adjusting the regularization parameters or the number of iterations, and updating the model again until validation passes.

[0150] Based on the above method embodiments, the second embodiment of this application discloses a process optimization system based on a multi-objective genetic algorithm. The process optimization system based on a multi-objective genetic algorithm of this application embodiment can implement any of the above-described process optimization methods based on multi-objective genetic algorithms, and the specific working process of each module in the process optimization system based on multi-objective genetic algorithms can be referred to the corresponding process in the above method embodiments.

[0151] For ease of understanding, an example is given below: A process optimization system based on a multi-objective genetic algorithm includes:

[0152] The process knowledge base module is used to store historical production process data, including historical product specification feature parameter sets, historical rubber compound formula feature parameter sets, historical process parameter sets, historical work order pass rates, historical work order equipment effective utilization rates, historical work order unit output energy consumption, and historical surface quality ratings.

[0153] The online sensing and data acquisition module communicates with the manufacturing execution system and field sensors to receive work order instructions, acquire product specification feature parameter sets and rubber compound formula feature parameter sets, and perform online sensing of the rheological properties of the rubber compound to obtain actual rheological parameters and update the rubber compound formula feature parameter set.

[0154] The model training and maintenance module trains a multi-objective fitness evaluation model based on historical sample datasets. The input of the multi-objective fitness evaluation model is the set of product specification feature parameters, the set of rubber compound feature parameters, and the set of process parameters. The output is the predicted quality evaluation score, the predicted performance evaluation score, the predicted energy efficiency evaluation score, and the predicted surface defect risk score.

[0155] The optimization solution module is used to: define a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed; and construct the feasible region of the decision variable parameter set based on equipment physical constraints, process standard constraints, and experience linkage constraints; receive the product specification feature parameter set of the current work order, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model, and input them into the multi-objective genetic algorithm; the multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than a set risk threshold, and performs global optimization within the feasible region to output the Pareto optimal solution set;

[0156] The strategy decision-making and distribution module is used to select recommended process formulas from the Pareto optimal solution set according to the preset production strategy, and distribute the recommended process formulas to the manufacturing execution system and equipment control system for execution.

[0157] The online monitoring and local optimization module communicates with the equipment control system and online detection equipment. It is used to collect real-time process data and product surface quality data during production execution, and calculate the surface quality assessment index based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or when key process parameters drift uncontrollably, it triggers rapid local optimization starting from the current execution parameters, generates process parameter fine-tuning increments, and issues them.

[0158] The third embodiment of this application provides a computer device, which may include a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement a process optimization method based on a multi-objective genetic algorithm.

[0159] The memory can communicate with the processor via a communication bus, which can be an address bus, a data bus, a control bus, etc.

[0160] Additionally, the memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.

[0161] Furthermore, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0162] The fourth embodiment of this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a process optimization method based on a multi-objective genetic algorithm.

[0163] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0164] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A process optimization method based on a multi-objective genetic algorithm, characterized in that, include: Based on the work order instructions issued by the manufacturing execution system, obtain the product specification feature parameter set and the rubber compound formulation feature parameter set; The rheological properties of the rubber compound corresponding to the set of rubber compound formulation feature parameters are sensed online to obtain the actual rheological parameters and update the set of rubber compound formulation feature parameters. Based on the updated set of rubber compound formulation feature parameters and the set of product specification feature parameters, the parameter set of historical process parameters, historical work order qualification rate, historical work order equipment effective utilization rate, historical work order unit output energy consumption and historical surface quality rating are extracted from the process knowledge base to form a historical sample dataset. Based on the historical sample dataset, a multi-objective fitness evaluation model is trained. The input of the multi-objective fitness evaluation model is the set of product specification feature parameters, the set of rubber compound formulation feature parameters, and the set of process parameters. The output is the predicted quality evaluation score, the predicted performance evaluation score, the predicted energy efficiency evaluation score, and the predicted surface defect risk score. Define a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed, and construct the feasible domain of the set of decision variable parameters based on equipment physical constraints, process standard constraints, and experience linkage constraints. The product specification feature parameter set of the current work order, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model are input into the multi-objective genetic algorithm. The multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than a set risk threshold. It performs global optimization within the feasible region and outputs a Pareto optimal solution set. According to the preset production strategy, recommended process formulas are selected from the Pareto optimal solution set, and the recommended process formulas are sent to the manufacturing execution system and equipment control system for execution. During production execution, real-time process data and product surface quality data are collected, and a surface quality assessment index is calculated based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or a key process parameter experiences uncontrollable drift, rapid local optimization is triggered, starting from the current execution parameters, to generate and issue process parameter fine-tuning increments.

2. The process optimization method based on multi-objective genetic algorithm according to claim 1, characterized in that, The calculation of the effective utilization rate of equipment in the historical work orders includes: Based on the equipment status event log, the effective processing time of historical work orders is calculated, excluding startup and trial molding time. Calculate the historical work order utilization rate, which is the percentage of the effective processing time of the historical work order to the planned production time. Calculate the historical work order pass rate, which is the percentage of the number of qualified semi-finished products in the historical work orders to the total number of inspections; Calculate the effective utilization rate of equipment for historical work orders, where the effective utilization rate of equipment for historical work orders = historical work order operation rate * historical work order qualification rate.

3. The process optimization method based on multi-objective genetic algorithm according to claim 1, characterized in that, The specific steps for training a multi-objective fitness evaluation model based on the aforementioned historical sample dataset include: The parameter sets of product specification features, rubber compound formulation features, and historical process parameters in the historical sample dataset are used as training inputs. The historical quality assessment score, historical performance assessment score, historical energy efficiency assessment score, and historical surface defect risk score in the historical sample dataset are used as training labels for the predicted quality assessment score, predicted performance assessment score, predicted energy efficiency assessment score, and predicted surface defect risk score, respectively. The gradient boosting regression tree algorithm is used to train sub-models for predicting quality assessment scores, performance assessment scores, energy efficiency assessment scores, and surface defect risk scores, which together constitute a multi-objective fitness assessment model.

4. The process optimization method based on multi-objective genetic algorithm according to claim 1, characterized in that, The global optimization solution within the feasible region includes: An initial population is randomly generated within the feasible domain, and historically successful process parameters that are most similar to the current product specification feature parameter set and rubber compound formulation feature parameter set are retrieved from the process knowledge base and injected into the initial population as seed individuals. For each candidate decision variable parameter set in the initial population, the multi-objective fitness evaluation model is invoked to calculate the corresponding prediction quality evaluation score, prediction effectiveness evaluation score, prediction energy efficiency evaluation score, and prediction surface defect risk score. Based on the predicted surface defect risk score, individuals in the population are pre-screened, and individuals whose predicted surface defect risk score exceeds the set risk threshold are removed. Based on the predicted quality assessment score, predicted efficacy assessment score, and predicted energy efficiency assessment score of the individuals retained after pre-screening, non-dominated ranking and crowding degree calculation are used to evaluate and select individuals in the population. Perform crossover and mutation operations on the selected individuals, and ensure that the newly generated individuals are located within the feasible region by the constraint repair operator; The selection, crossover, and mutation processes are iteratively executed until convergence, and the Pareto optimal solution set is output.

5. The process optimization method based on multi-objective genetic algorithm according to claim 4, characterized in that, Recommended process formulations are selected from the Pareto optimal solution set based on a pre-defined production strategy, including: Set quality assessment score thresholds; If the production strategy prioritizes quality, then among all solutions where the predicted quality assessment score is greater than the quality assessment score threshold, a comprehensive evaluation index is calculated based on the predicted performance assessment score and the predicted energy efficiency assessment score, and the solution with the optimal comprehensive evaluation index is selected as the recommended process formula. If the production strategy prioritizes efficiency, then among the solutions whose predicted quality assessment scores are greater than the preset lower limit threshold for quality scores, the solution with the highest predicted performance assessment score is selected as the recommended process formula. If the production strategy prioritizes cost, then among all solutions with predicted quality assessment scores greater than the quality assessment score threshold, the solution with the highest predicted energy efficiency assessment score is selected as the recommended process formulation.

6. The process optimization method based on multi-objective genetic algorithm according to claim 1, characterized in that, The specific steps for triggering rapid local optimization and generating fine-tuning increments for process parameters include: Use the current execution parameters as the initial point for local optimization; Construct a local optimization objective, which is to minimize the predicted surface defect risk score so that the actual surface quality assessment index falls back, or to minimize the deviation of key process parameters from the recommended set value so as to stabilize the production state. Construct local optimization constraints, wherein the changes in the predicted quality assessment score and the predicted energy efficiency assessment score relative to the current value are less than a set range; A fast search is performed within a preset neighborhood of the initial point to obtain the process parameter fine-tuning increment that optimizes the local optimization objective.

7. The process optimization method based on multi-objective genetic algorithm according to claim 1, characterized in that, The process optimization method further includes: After the work order is completed, the actual work order qualification rate, actual work order equipment effective utilization rate, actual work order unit output energy consumption and actual surface quality rating are collected and mapped to generate actual quality assessment score, actual efficiency assessment score, actual energy efficiency assessment score and actual surface defect risk score, which constitute the actual performance record. After the actual performance records are stored in the process knowledge base, the model incremental update process is triggered; Outlier detection and data cleaning are performed on the newly added actual performance records to remove invalid or abnormal data records caused by sensor failure or extreme operating conditions. The newly added actual performance records after cleaning are merged with the historical training data buffer of the current multi-objective fitness evaluation model to form an incremental learning sample set; An online learning algorithm is used to iteratively update the multi-objective fitness evaluation model with the incremental learning sample set. During the update process, new samples are given higher weights, while the original key parameters of the model are subject to regularization constraints. After the incremental model update is completed, the updated model is integrated and fused with the original model to generate a new multi-objective fitness evaluation model.

8. A process optimization system based on multi-objective genetic algorithm, characterized by, Performing the process optimization method based on a multi-objective genetic algorithm as described in any one of claims 1 to 7, comprising: The process knowledge base module is used to store historical production process data, including historical product specification feature parameter sets, historical rubber compound formula feature parameter sets, historical process parameter sets, historical work order pass rates, historical work order equipment effective utilization rates, historical work order unit output energy consumption, and historical surface quality ratings. The online sensing and data acquisition module communicates with the manufacturing execution system and field sensors to receive work order instructions, acquire product specification feature parameter sets and rubber compound formula feature parameter sets, and perform online sensing of the rheological properties of the rubber compound to obtain actual rheological parameters and update the rubber compound formula feature parameter set. The model training and maintenance module trains a multi-objective fitness evaluation model based on the historical sample dataset. The input of the multi-objective fitness evaluation model is a set of product specification feature parameters, a set of rubber compound formulation feature parameters, and a set of process parameters. The output is a predicted quality evaluation score, a predicted performance evaluation score, a predicted energy efficiency evaluation score, and a predicted surface defect risk score. The optimization solution module is used to: define a set of decision variable parameters including the speed of the rubber mixing mill, the temperature of the rubber mixing mill rollers, the heating temperature of each section of the extruder, and the traction speed; and construct the feasible region of the decision variable parameter set based on equipment physical constraints, process standard constraints, and experience linkage constraints; receive the product specification feature parameter set of the current work order, the updated rubber compound formulation feature parameter set, the feasible region, and the multi-objective fitness evaluation model, and input them into the multi-objective genetic algorithm; the multi-objective genetic algorithm aims to maximize the predicted quality evaluation score, maximize the predicted efficiency evaluation score, and maximize the predicted energy efficiency evaluation score, with the constraint that the predicted surface defect risk score is lower than a set risk threshold, and performs global optimization within the feasible region to output a Pareto optimal solution set; The strategy decision-making and distribution module is used to select recommended process formulas from the Pareto optimal solution set according to the preset production strategy, and distribute the recommended process formulas to the manufacturing execution system and equipment control system for execution; The online monitoring and local optimization module communicates with the equipment control system and online detection equipment. It is used to collect real-time process data and product surface quality data during production execution, and calculate the surface quality assessment index based on the collected quality information. When the surface quality assessment index is detected to exceed the warning threshold or when key process parameters drift uncontrollably, it triggers rapid local optimization starting from the current execution parameters, generates process parameter fine-tuning increments, and issues them.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the process optimization method based on a multi-objective genetic algorithm as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7, based on a multi-objective genetic algorithm, to optimize the process.

Images