Re-drying section process parameter optimization control method based on machine learning and genetic algorithm

Abstract

The invention discloses a re-drying section process parameter optimization control method based on machine learning and a genetic algorithm, and the method comprises the following working steps: 1, constructing a tobacco re-drying process parameter database, and standardizing the parameters; 2, real-time technological parameter values in redrying batch production are collected, and the parameters are digitized; and step 3, finding out a function relationship between the quality value and the process parameter value by using a machine learning method, and recommending an optimal process parameter combination under the current objective condition in real time by using a genetic algorithm to achieve parameter intelligence. Processing parameters are in a process controlled state, the parameters are systematically managed, and necessary related work such as parameter collection and analysis is carried out; a parameter setting and adjusting basis is provided; and parameter management has continuity, parameters are optimized and improved sustainably, and the technological level is improved.

Description

Technical Field

[0001] The present invention relates to an optimization control method for process parameters in the tobacco leaf re-drying process section, specifically an optimization control method for re-drying process parameters based on machine learning and genetic algorithms. Background Technique

[0002] Threshing and re-drying is a necessary processing process for tobacco leaves to transform from agricultural products into raw materials for the cigarette industry. It can generally be divided into several process sections such as vacuum conditioning, primary leaf conditioning, secondary leaf conditioning, foreign matter removal, threshing and pneumatic conveying, leaf re-drying, stem re-drying, and finished product packaging. Among them, the parameter management of the leaf re-drying process section belongs to one of the core process contents of threshing and re-drying, involving a series of specific tasks such as parameter setting, adjustment, and optimization. The level of parameter management directly determines whether the process performance of the equipment can be fully exerted and whether the product quality is stable and controllable. With the continuous improvement of the quality requirements of re-dried products by cigarette industrial enterprises, the importance of parameter management in the re-drying link has become increasingly prominent. Work such as the construction of regional processing centers for key brand raw materials in the industry and major special projects for threshing and re-drying technology upgrades has put forward higher and clearer requirements for parameter management.

[0003] At present, parameter management work in domestic re-drying enterprises is generally lacking, and the processing parameters are basically not in a process-controlled state. Usually, during processing, equipment operators set and adjust parameters according to their personal subjective experience. The problems brought about are: first, the systematicness of parameter management is insufficient. Parameter setting and adjustment mainly aim to meet basic production operations, and necessary related work such as parameter collection and analysis is not carried out; second, the basis for parameter setting and adjustment is not rigorous, and the implementation of process discipline is not serious. The setting and adjustment of parameters during the production process are relatively arbitrary; third, the continuity of parameter management is insufficient, process experience cannot be effectively accumulated, and parameters cannot be continuously optimized and improved. In the short term, these problems have caused quality fluctuations and instability among teams. In the long run, such a parameter management method will lead to difficulties in carrying out process analysis and parameter optimization, effective process experience cannot be summarized and refined, and the process technology will fall into low-level repetition. Summary of the Invention

[0004] The purpose of the present invention is to provide an optimization control method for process parameters in the tobacco leaf re-drying process section based on machine learning and genetic algorithms, which can automatically set various relevant process parameters to achieve or be closest to the given quality indicators in production.

[0005] To achieve the above purpose, the present invention provides the following technical solutions, including the following steps:

[0006] Step1. Construct a database of tobacco leaf re-drying process parameters and standardize the parameters;

[0007] Step2. Collect the real-time process parameter values during re-drying batches of production and digitize the parameters;

[0008] Step 3. Use machine learning methods to find the functional relationship between quality values and process parameter values, and use genetic algorithms to recommend the optimal process parameter combination under current objective conditions in real time to achieve parameter intelligence.

[0009] The parameters in Step 1 are further divided into process parameter values and objective parameters. The process of parameter standardization is to set the importance of each parameter as: critical, important, general; and set its upper limit, lower limit, and recommended value.

[0010] Step 2 includes 2 steps:

[0011] Step 2-1. Parameter maintenance, including: maintenance entry, and issuing, uploading, archiving, and exporting data after the production process, as well as data visualization.

[0012] Step 2-2. Real-time parameter collection: Real-time parameter collection comes from associated systems and the production line real-time control system, including the production management system MIS and the quality system QA; use the Java microservice framework Java Spring Cloud + Spring Boot to complete data interaction and docking with them; the production line real-time control system directly communicates with Siemens PLC (programmable logic controller) using C# to obtain production equipment parameters, process parameters, quality parameters, and environmental data in real time. After data collection, data cleaning, time alignment, etc. are also completed according to predefined rules. During cleaning, it mainly depends on the upper and lower limits of the parameters, and part of the time alignment work depends on experiments. For example, after multiple tests, it is known that the average time for tobacco leaves to pass from the entrance to the exit of the re-drying machine is 7 minutes and 13 seconds, so the moisture content value at the entrance and the exit of the infrared moisture meter is separated by 7 minutes and 13 seconds.

[0013] Step 3 includes 2 steps:

[0014] Step 3-1. Machine learning, forward prediction: Find the functional correspondence between the quality index values, process parameter values, and objective parameters in the re-drying process section; among them, there are 46 quality indexes with direct connections and 10 objective parameters.

[0015] We need to find two function models based on the collected data:

[0016] y1 (moisture content at the exit of tobacco leaves) = f(x1, x2, x3.....x56);

[0017] y2 (temperature at the exit of tobacco leaves) = f(x1, x2, x3.....x56).

[0018] Among them, x1 - x56 refer to 46 process parameter values + 10 objective parameter values, a total of 56 characteristic values.

[0019] That is, given x, find y. For example, assume x1 (customer) = China Tobacco Jiangsu Industrial Co., Ltd., x2 (temperature) = 28, x3 (humidity) = 35, x4 (temperature in drying zone 1) = 100...., find y1 (moisture content at the outlet) =?, y2 (temperature at the outlet) =?.

[0020] Step3 - 2. Optimize the algorithm and perform reverse optimization: Use the genetic algorithm to recommend the optimal process parameter combination under the current objective conditions in real time;

[0021] In the above - mentioned Step3 - 1, there are two quality index values, namely the moisture content of tobacco leaves at the outlet and the temperature of tobacco leaves at the outlet, which are represented by y1 and y2 respectively; Use the XGboost algorithm to build a model and obtain:

[0022] y1 = ∑Wi * Xi + b1;

[0023] y2 = ∑Qi * Xi + b2;

[0024] Where i represents the i - th parameter, Wi and Qi refer to the weights of the i - th parameter characteristic value, b1 and b2 refer to the bias, the weights and the bias are obtained through model training, and can be understood as constants in the function, and Xi is the i - th parameter value;

[0025] In the above - mentioned Step3 - 2, optimize the algorithm and perform reverse optimization; It is to find the combination of process parameters given the quality index, and includes the following 6 steps:

[0026] Step3 - 2 - 1. Determine that the goal is multi - objective optimization, and it is necessary to optimize the moisture content of tobacco leaves at the outlet and the temperature at the outlet simultaneously. Its initial optimization function is: |Ymoisture set - Ymoisture real|+|Ytemperature set - Ytemperature real|<Threshold;

[0027] Among them, Ymoisture set is: the set value of the moisture content quality of tobacco leaves;

[0028] Ytemperature set is: the set value of the temperature quality of tobacco leaves;

[0029] Ymoisture real is: the actual value of the moisture content of tobacco leaves;

[0030] Ytemperature real is: the actual value of the temperature of tobacco leaves;

[0031] Ytemperature real is: the actual value of the temperature of tobacco leaves;

[0032] Threshold refers to the tolerance threshold; when the absolute value of the difference between the actual values of moisture content and temperature and their quality set values is within the threshold range, it is the goal of optimization control;

[0033] Step3-2-2, determine the process parameters that need to be optimized and controlled this time:

[0034] Divide the variables into two parts, those that need to be optimized and controlled and those that do not need to be optimized and controlled, and represent them with Xoptimized and Xconstant respectively. Then y = f(Xoptimized, Xconstant);

[0035] Step3-2-3, randomly generate the first-generation population; the value of Xconstant is fixed; Xoptimized is assigned values according to the following formula: Xi = Lower + Adjust Step * Random Range;

[0036] Where Lower refers to the lower limit of the parameter value, Adjust Step refers to the recommended value, and Random Range refers to the random adjustment range of the recommended value, and its value = a random integer from 0 to (upper limit - lower limit) / Adjust Step;

[0037] Step3-2-4, adopt the roulette wheel algorithm, substitute the gene variable values (Xoptimized and Xconstant) of each individual into the prediction function y = f(Xoptimized, Xconstant), and use the difference between the y value and the expected quality target value as the judgment condition. Prioritize selecting those combinations with the smallest difference as the genetic candidate combinations for the next generation;

[0038] Step3-2-5, adopt gene exchange in the genetic candidate combinations, and add gene mutation with a certain probability to generate a new population;

[0039] Step3-2-6, go back to Step3-2-4 and Step3-2-5 to continue the loop. When a process parameter combination smaller than the threshold is detected, terminate the loop to obtain the recommended value of optimization control.

[0040] Preferably, the optimization function in Step3-2-1 is:

[0041] (Wmoisture * |Ymoisture set - Ymoisture predict| + Wtemperature * |Ytemperature set - Ytemperature predict|) / (Wmoisture + Wtemperature) < 0.01;

[0042] Where Wmoisture is: the weight of the moisture content of tobacco leaves;

[0043] Ymoisture set is: the set value of the quality of the moisture content of tobacco leaves;

[0044] Ymoisture predict is: the moisture content of tobacco leaves;

[0045] Wtemperature is: the weight of the temperature of tobacco leaves;

[0046] Ytemperature set is: the set value of the quality of the temperature of tobacco leaves;

[0047] Ytemperature predict is: the predicted value of the temperature of tobacco leaves;

[0048] Preferably, there are 46 process parameters involved in the re-drying process section, and 10 objective parameters.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] The processing parameters are in a process-controlled state, the parameters are systematically managed, and necessary related work such as parameter collection and analysis is carried out; there is a basis for parameter setting and adjustment; parameter management has continuity, the parameters can be continuously optimized and improved, and the process technology level is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] It includes the following steps:

[0054] Step1. Construct a database of re-drying process parameters of tobacco leaves and standardize the parameters;

[0055] Step2. Collect the real-time process parameter values in the production of re-drying batches and digitize the parameters;

[0056] Step3. Use machine learning methods to find the functional relationship between the quality value and the process parameter value, and use genetic algorithms to recommend the optimal process parameter combination under the current objective conditions in real time to achieve parameter intelligence;

[0057] The parameters in Step1 are further divided into process parameter values and objective parameters. The process of parameter standardization is to set the importance of each parameter as: critical, important, general; and set its upper limit, lower limit, and recommended value.

[0058] Step2 includes 2 steps:

[0059] Step2-1, Parameter maintenance, including: maintenance input, and production process parameter distribution, parameter upload, data archiving and data export after batch end, and data visualization presentation;

[0060] Step2-2, Real-time parameter collection: Real-time parameter collection comes from associated systems and the production line real-time control system, including the production management system MIS, quality system QA; Use the Java microservices framework Java Spring Cloud + Spring Boot to complete data interaction and docking with them; The production line real-time control system directly communicates with Siemens PLC (programmable logic controller) using C# and then obtains in real-time, collects equipment parameters, process parameters, quality parameters, and environmental data. After data collection, according to predefined rules, the data is also cleaned, time-aligned, etc. When cleaning, it mainly depends on the upper and lower limits of the parameters, and part of the work of time alignment depends on experiments. For example, through multiple tests, it is known that the average passing time of tobacco leaves from the entrance to the exit of the re-drying machine is 7 minutes and 13 seconds, then the moisture content value at the entrance and the value at the exit of the infrared moisture meter are separated by 7 minutes and 13 seconds.

[0061] Step3 includes 2 steps:

[0062] Step3-1, Machine learning, forward prediction: Find the functional correspondence between the quality index values in the re-drying process section and process parameter values, objective parameters; Among them, there are 46 quality indicators with direct connections and 10 objective parameters.

[0063] Table 1 Process parameters of the tobacco leaf re-drying process section

[0064]

[0065]

[0066] The objective parameters include: workshop temperature, humidity, commissioning customer, original tobacco grade, selected grade, tobacco leaf origin, tobacco leaf variety, product name, target value of export moisture content, target value of export temperature, 10 parameters.

[0067] We need to find two function models based on the collected data:

[0068] y1 (Moisture content of tobacco leaves for export) = f(x1, x2, x3.....x56);

[0069] y2 (Temperature of tobacco leaves for export) = f(x1, x2, x3.....x56).

[0070] Among them, x1 - x56 refer to 46 process parameter values + 10 objective parameter values, a total of 56 characteristic values.

[0071] That is, given x, find y. For example, assume x1 (customer) = China Tobacco Jiangsu Industrial Co., Ltd., x2 (temperature) = 28, x3 (humidity) = 35, x4 (temperature in drying zone 1) = 100...., find y1 (moisture content for export) =?, y2 (temperature for export) =?.

[0072] In the process of parameter collection in Step 2, we have collected a large amount of data required for machine learning modeling (that is, the matrix data corresponding to x1....x56 characteristic values, y1 moisture content for export, and y2 temperature for export). 50 million records were collected during the first - phase training of the model, and it was extended to 120 million records in the second phase. The training process is solved through C# + Microsoft Ml.Net machine learning framework. The main reason for choosing ML.Net is the convenience of mutual calling with the C# code for data collection. The algorithm model selected is XGboost. The prediction model trained has good effects in production practice, and the prediction values and actual values have a high degree of coincidence.

[0073] XGBoost is one of the most excellent algorithms in recent machine learning competitions. Based on GBDT, some innovative improvements have been made at the algorithm level and system design level. XGBoost can be regarded as a better and faster implementation of GBDT. It requires a long space to introduce the detailed algorithm implementation details of XGBoost, so it will not be elaborated in this article. In practice, we have also fully compared various algorithms, including Random Forest, Fast Forest Regression, etc., and finally selected the XGBoost algorithm with the best effect.

[0074] After training through the XGBoost algorithm, the above two function model expressions are obtained as follows:

[0075] y1 (Moisture content of tobacco leaves for export) = f(x1, x2, x3.....x56) = w1 * x1 + w2 * x2 +... + w56 * x56 + b1, abbreviated as y1 = ∑Wi * Xi + b1.

[0076] y2 (Temperature of tobacco leaves for export) = f(x1, x2, x3.....x56) = q1 * x1 + q2 * x2 +... + q56 * x56 + b2,

[0077] Abbreviated as y2 = ∑Qi*Xi + b2.

[0078] In the said Step3-1, there are two quality index values, namely the moisture content of the tobacco leaves at the outlet and the temperature of the tobacco leaves at the outlet, which are represented by y1 and y2 respectively; using the XGboost algorithm for modeling, we get:

[0079] y1 = ∑Wi*Xi + b1;

[0080] y2 = ∑Qi*Xi + b2;

[0081] Where i represents the i-th parameter, Wi and Qi refer to the weights of the i-th parameter eigenvalue, b1 and b2 refer to the bias, the weights and biases are obtained through model training, and can be understood as constants in the function, and Xi is the i-th parameter value.

[0082] In the forward prediction modeling of Step3-1, based on the collected historical data, the function mapping relationship between the independent variable x and the dependent variable y in the tobacco leaf redrying process section is found, where y refers to the quality index concerned in the tobacco leaf redrying process section.

[0083] The forward prediction model of the function mapping relationship between the variable x and the dependent variable y can give real-time reference to the operators during the production process, but this is not enough. We still need to go further. On this basis, feedback optimization is carried out to find the optimal combination of process parameters, that is, an optimization algorithm needs to be given. The factory specifies one (or more) quality indexes that it wants to achieve. Under the limited objective conditions, how should the relevant process parameters be set to achieve (or be closest to) this quality index value.

[0084] The forward prediction model in Step 3-1 is f(x) → y, that is, given the x process parameters, the y quality index is obtained. In Step3-2, on the contrary, given the y quality index, the combination of x process parameters is obtained. For example, assuming that the manufacturer hopes that the moisture content of the tobacco leaves after the tobacco leaf redrying process section is 11% and the temperature is 40°C, under the known but unchangeable objective conditions (temperature and humidity) and the natural properties of the tobacco leaves (grade, variety, origin, etc.), how should the multiple controllable process parameters involved in the tobacco leaf redrying be set to achieve (be closest to) our expected value.

[0085] This problem can be abstractly understood as an NP problem in mathematics. The NP problem refers to the Nondeterministic Polynomial class of problems. The characteristic of this type of problem is that it is not easy to directly find the best answer, but given a set of numerical values, it can be easily verified whether it is the correct answer.

[0086] The solutions of the NP problem can be divided into two categories:

[0087] Category 1: If all answers can be verified within the allowed time, it is called NPC (Nondeterministic Polynomial complete problem), and it can be completed by the exhaustive method. For example, assume that there are 6 relevant process parameters found in the cut tobacco leaf redrying stage, and each parameter has 10 possible adjustment values. Then the possible combinations of the 6 parameters are 10 * 10 * 10 * 10 * 10 * 10 = 1,000,000, one million. One million combinations are not a big problem for today's computing power. We can completely substitute these one million combinations into the forward prediction model in turn to determine which combination (or some combinations) of results is closest to our expectations, that is, the optimal parameter combination we are looking for.

[0088] Category 2: If the combined computational amount is too large to be verified within the allowed time, the exhaustive method won't work. For example, in the actual cut tobacco leaf redrying stage, there are more than 56 relevant parameters. Even if each parameter has only 10 possible values, the final combination of 10 to the 57th power is an astronomical figure and cannot be calculated by exhaustion. In this case, it can be solved by the genetic algorithm.

[0089] The genetic algorithm is a search algorithm used in computational mathematics to solve optimization problems and is a type of evolutionary algorithm. Evolutionary algorithms were initially developed by drawing on some phenomena in evolutionary biology, including inheritance, mutation, natural selection, and hybridization. The genetic algorithm is usually implemented as a computer simulation. For an optimization problem, a population of abstract representations (called genes or chromosomes) of a certain number of candidate solutions (called individuals) evolves towards better solutions. Traditionally, solutions are represented in binary (i.e., strings of 0s and 1s), but other representation methods can also be used. The evolution starts from a population of completely random individuals and occurs generation by generation. In each generation, the fitness of the entire population is evaluated, and multiple individuals are randomly selected from the current population (based on their fitness), and a new population of life is generated through natural selection and mutation. This population becomes the current population in the next iteration of the algorithm.

[0090] In the optimization control of process parameters in the cut tobacco leaf redrying process section, the program logic of the genetic algorithm runs according to the following steps:

[0091] Step 3 - 2: Optimization algorithm, reverse optimization; Given the quality index, find the combination of process parameters, including the following 6 steps:

[0092] Step 3 - 2 - 1: Determine that the goal is multi - objective optimization, and it is necessary to optimize the moisture content and outlet temperature of the tobacco leaves at the same time. Its initial

[0093] The initial optimization function is: |Ymoisture set - Ymoisture real| + |Ytemperature set - Ytemperature real| < Threshold;

[0094] Where Ymoisture set is: the set value of the tobacco leaf moisture content quality;

[0095] Ytemperature set is: the set value of the tobacco leaf temperature quality;

[0096] Ymoisture real is: the actual value of the tobacco leaf moisture content;

[0097] Ytemperature real is: the actual value of the tobacco leaf temperature;

[0098] Threshold refers to the tolerance threshold; the absolute value of the difference between the actual values of the moisture content and temperature and their set values of quality is within the threshold range. For example, taking a value of 0.01 is the goal of the optimization control;

[0099] Ymoisture real and Ytemperature real cannot be obtained in the optimization control (when obtained, the result has already been generated and the recommended parameters are useless), but in Step3 - 1 we have already found their prediction function models. Assuming our prediction model is accurate and the predicted value is equal to the actual value, that is, Ypredict = Yreal. After substituting into the equation, we get |Ymoistureset - Ymoisture predict| + |Ytemperature set - Ytemperature predict| < Threshold0.01. And Ypredict = f(x1, x2, x3.....x56) = ∑Wi*Xi + b, so the optimization control is associated with the previous prediction model.

[0100] Since the importance of the tobacco leaf moisture content and the tobacco leaf temperature is different, the weight of the two needs to be added to the objective function. That is, the final optimization objective function is:

[0101] (Wmoisture * |Ymoisture set - Ymoisture predict| + Wtemperature * |Ytemperature set - Ytemperature predict|) / (Wmoisture + Wtemperature) < 0.01. Here, W refers to the weight. For example, Wmoisture = 8 and Wtemperature = 2, indicating that in the optimization control, the priority (weight) of the tobacco leaf moisture content is 4 times that of the tobacco leaf outlet temperature.

[0102] Step 3-2-2, determine the process parameters that need to be optimized and controlled this time:

[0103] In the prediction function of Step 3-1, y=f(x1....x56). Theoretically, except for the 10 objective parameters x47-x56 (workshop temperature, humidity, entrusted customers, raw smoke grade, etc.), the other x1-x46 process parameters can be optimized and controlled. However, this is not done in actual work. Usually, only 10 parameters are selected for optimization and control according to needs, and the others use the current values without change. The result of this step is to divide the x1........x56 variables in y=f(x1....x56) into two parts, those that need to be optimized and those that do not need to be optimized, represented by Xoptimized and Xconstant respectively, then y=f(Xoptimized,Xconstant);

[0104] Step3-2-3, randomly generate the first generation population, for example, generate 1000 individuals. The genes of each individual are composed of 56 variables, Xoptimized and Xconstant, in Step3-2-2. The values of Xconstant (variables that do not need or cannot be optimized) are the same, such as the temperature and humidity of the current workshop, the grade and origin of the processed tobacco leaves, etc. The Xoptimized variable is assigned according to the following formula:

[0105] Xi=Lower+Adjust Step*Random Range, where Lower refers to the lower limit of the parameter value, AdjustStep refers to the recommended value, and Random Range refers to the random adjustment range, whose value = a random integer from 0 to (upper limit - lower limit) / AdjustStep.

[0106] For example, the lower limit of the temperature of drying zone 1 is 50, the upper limit is 120, and the recommended value is 0.5, then its control range is 0 to (120-50) / 0.5=140, and the random value of 80 in the range of 0-140, then the temperature variable value of drying zone 1 of a certain individual (assuming the number is 1) is x=50+80*0.5=90. And the random value of another individual (assuming the number is 2) is 100, and its temperature variable value of zone 1 is x=50+100*0.5=100.

[0107] Step3-2-4. Use the roulette wheel algorithm. Substitute the gene variable values (Xoptimized and Xconstant) of each individual into the prediction function y = f(Xoptimized, Xconstant). Take the difference between the y value and the expected quality target value as the judgment condition, and preferentially select the combinations with the smallest differences as the genetic candidate combinations for the next generation.

[0108] For example, assume that after substituting the parameter values of No. 1 into the prediction function, the difference from the quality target value is 0.2, and the difference after substituting the parameter values of No. 2 is 0.8. Then, compared with No. 2, the gene of No. 1 has a higher probability of being selected by the program as the genetic candidate combination for the next generation.

[0109] Step3-2-5. In the genetic candidate combinations, use gene exchange and, with a certain probability, gene mutation to generate a new population.

[0110] The gene exchange process can be understood as follows. Assume that the genes of No. 1 and No. 2 are both selected as the genetic candidate combinations for the next generation. The drying zone 1 temperature value of No. 1 is 80, and the drying zone 2 temperature value is 100, while the zone 1 temperature and zone 2 temperature of No. 2 are 110 and 120 respectively. For the next generation of No. 1 and No. 2 (assume the number is 012), the drying zone 1 temperature can be 80 (from No. 1), and the drying zone 2 temperature can be 120 (from No. 2). This simulates the gene inheritance mechanism in the biological world.

[0111] The gene mutation process can be exemplified as follows. The system sets a mutation probability. Assume that No. 012 just meets this probability. The drying zone 2 temperature can neither come from the parental No. 1 nor from the parental No. 2, but instead, a new value is regenerated according to the formula Xi = Lower + Adjust Step * Random Range during initialization. This simulates the gene mutation mechanism in the biological world.

[0112] Step3-2-6. Go back to Step3-2-4 and Step3-2-5 to continue the loop. When a process parameter combination smaller than the threshold is detected, terminate the loop to obtain the optimized control recommended value. For example, the set value of the tobacco leaf outlet moisture content (i.e., the quality specified value) is 11, and the difference threshold is 0.1. Then, if the function value is 10.99 after substituting the process parameter values of a certain offspring individual in the current loop into the prediction function f(x), which is less than (11 - 0.1), stop the loop and record the process parameter values of this individual, which are the optimized control recommended values. If the threshold is never reached, terminate the loop after looping for the specified n generations (such as 300 generations), and find the individual closest to the quality set value from the candidate genetic combinations in the previous generations of loops. Its process parameter values are the optimized recommended values.

Claims

1. A method for optimizing and controlling the technological parameters of the re-drying section based on machine learning and genetic algorithms, characterized in that: It includes the following working steps: Step1. Construct a database of tobacco leaf redrying process parameters and standardize the parameters; Step2. Collect the real-time process parameter values during the production of redrying batches and digitize the parameters; Step3. Use machine learning methods to find the functional relationship between the quality values and process parameter values, and use genetic algorithms to recommend the optimal process parameter combination under the current objective conditions in real time to achieve parameter intelligence; The parameters in Step1 are further divided into process parameter values and objective parameters. The process of standardizing the parameters is to set the importance of each parameter as: critical, important, general; And set its upper limit, lower limit, and recommended value; Step2 includes 2 steps: Step2-1. Parameter maintenance, including: maintenance input, as well as the issuance of process parameters during production, parameter upload, data archiving and data export after the batch ends, and data visualization presentation; Step2-2. Collection of real-time parameters: The collection of real-time parameters comes from associated systems and the production line real-time control system; and according to predefined rules, the data is cleaned and time-aligned; Step3 includes 2 steps: Step3-1. Machine learning, forward prediction: Find the functional correspondence between the quality index values, process parameter values, and objective parameters in the redrying process section; Step3-2. Optimization algorithm, reverse optimization: Use genetic algorithms to recommend the optimal process parameter combination under the current objective conditions in real time; In Step3-1, there are two quality index values, the moisture content of tobacco leaves at the outlet and the temperature of tobacco leaves at the outlet, which are represented by y1 and y2 respectively; Use the XGboost algorithm to build a model and get: y1 = ∑Wi*Xi + b1; y2 = ∑Qi*Xi + b2; Where i represents the i-th parameter, Wi and Qi refer to the weights of the i-th parameter eigenvalue, b1 and b2 refer to the offset bias, and the weights and offsets are obtained through model training and can be understood as constants in the function, and Xi is the i-th parameter value; Step3-2. Optimization algorithm, reverse optimization; is to find the combination of process parameters given the quality index, and includes the following 6 steps: Step3-2-1. Determine that the goal is multi-objective optimization, and it is necessary to optimize the moisture content of tobacco leaves at the outlet and the temperature at the outlet simultaneously. Its initial optimization function is: |Ymoistureset - Ymoisturereal| + |Ytemperatureset - Ytemperaturereal| < Threshold; Where Ymoistureset is: the quality set value of the moisture content of tobacco leaves; Ytemperatureset is: the quality set value of the temperature of tobacco leaves; Ymoisturereal is: the actual value of the moisture content of tobacco leaves; Ytemperaturereal is: the actual value of the temperature of tobacco leaves; Threshold refers to the tolerance threshold; the absolute value of the difference between the actual values of the moisture content and temperature and their quality set values within the threshold range is the goal of optimization control; Step3-2-2. Determine the process parameters that need to be optimized and controlled this time: Divide the variables into two parts, those that need to be optimized and controlled and those that do not. Represent them with Xoptimized and Xconstant respectively, then y = f(Xoptimized, Xconstant); Step3-2-3. Randomly generate the first-generation population; the value of Xconstant is fixed; Xoptimized is assigned values according to the following formula: Xi = Lower + AdjustStep * RandomRange; Where Lower refers to the lower limit of the parameter value, AdjustStep refers to the recommended value, and RandomRange refers to the random adjustment range of the recommended value, and its value = a random integer from 0 to (upper limit - lower limit) / AdjustStep; Step3-2-4. Use the roulette wheel algorithm to substitute the gene variable values (Xoptimized and Xconstant) of each individual into the prediction function y = f(Xoptimized, Xconstant), and use the difference between the y value and the expected quality target value as the judgment condition. Prioritize selecting those combinations with the smallest differences as the genetic candidate combinations for the next generation; Step3-2-5. Use gene exchange in the genetic candidate combinations, and add gene mutation with a certain probability to generate a new population; Step3-2-6. Go back to Steps 3-2-4 and 3-2-5 to continue the loop. When a process parameter combination smaller than the threshold is detected, terminate the loop to obtain the optimized control recommended value.

2. The optimized control method for the process parameters of the re-drying section based on machine learning and genetic algorithm according to claim 1, characterized in that: The optimization function in Step 3-2-1 is: (Wmoisture * |Ymoistureset - Ymoisturepredict| + Wtemperature * |Ytemperatureset - Ytemperaturepredict|) / (Wmoisture + Wtemperature) < 0.01; Where Wmoisture is: the weight of the tobacco leaf moisture content; Ymoistureset is: the quality set value of the tobacco leaf moisture content; Ymoisturepredict is: the tobacco leaf moisture content; Wtemperature is: the weight of the tobacco leaf temperature; Ytemperatureset is: the quality set value of the tobacco leaf temperature; Ytemperaturepredict is: the predicted value of the tobacco leaf temperature.

3. The optimized control method for the process parameters of the re-drying section based on machine learning and genetic algorithms according to claim 1, characterized in that: There are a total of 46 process parameters involved in the re-drying process section, and 10 objective parameters.

Images