Cut tobacco dryer predictive control method based on multi-model fusion

Abstract

The invention relates to a cut tobacco dryer predictive control method based on multi-model fusion. The method comprises the following steps: acquiring preprocessing data; combining correlation analysis and random forest to screen features; a prediction model is constructed by using three algorithms of random forest, XGBoost and AdaBoost respectively; generating a training set and a verification set by using the historical production data, training the three prediction models by using the training set, and verifying the three prediction models by using the verification set; determining the weight according to the mean square error of the three prediction models, and carrying out weighted average on the prediction result of the fusion model to obtain a cut tobacco dryer outlet moisture prediction model based on multi-model fusion; and according to the predicted outlet moisture change trend, optimizing through a genetic algorithm, and adjusting the cut tobacco dryer control parameters. The advantages of the three models are fully utilized for fusion prediction, dynamic parameter optimization is carried out according to the predicted trend in combination with the genetic algorithm, and accurate prediction and self-adaptive control of outlet moisture are achieved.

Description

Technical Field

[0001] This invention relates to the field of tobacco processing technology, and specifically to a predictive control method for a tobacco drying machine based on multi-model fusion. Background Technology

[0002] In tobacco processing, the moisture content of tobacco shreds is a key process indicator, and its stability directly affects the quality of the finished product. However, the moisture content at the outlet of the tobacco drying machine is easily fluctuated due to multiple factors, including the production environment (such as temperature and humidity, and airflow disturbance) and equipment parameters (such as temperature, wind speed, and material conveying speed). Existing control methods mainly rely on human experience, which has drawbacks such as strong subjectivity, lack of quantitative standards, and large predictive uncertainties, making it difficult to adapt to complex production scenarios.

[0003] With the development of information technology, intelligent model analysis technology has been introduced to solve the problem of moisture control at the outlet. For example, the study "Research on Moisture Prediction at the Outlet of Thin Plate Drying Machine Based on Model Fusion" uses a weighted average fusion of three models: random forest, support vector machine, and backpropagation neural network, with the mean square error of the validation set (MSE) as the basis for weight calculation. However, this method relies on manual experience for feature selection (directly using process flow rate, steam parameters, etc.), without analyzing the importance of features through models, which may miss key influencing factors; data alignment uses a fixed time difference without dynamically optimizing the time lag, which may lead to misalignment of time-series data; it does not consider external variables such as environmental temperature and humidity, and only completes the prediction study without linking it to control decisions, resulting in poor control performance. For example, the study "Prediction of Moisture Content at the Outlet of Drying Machine Based on RF-MLP" uses random forest to select features and then uses multilayer perceptron (MLP) to build a prediction model, optimizing the structure through grid search. This method only uses random forest as a feature selection tool, without integrating it with MLP, and fails to leverage the synergistic advantages of the algorithms; in addition, MLP has problems such as slow training speed, easy gradient vanishing / exploding, difficulty in structure optimization, and high hardware resource requirements, resulting in poor practical industrial application. For example, the paper "Optimization and Control of Moisture Content at the Outlet of a Wire Drying Machine Based on the RBF-ARX Model" uses a radial basis function-autoregressive exogenous (RBF-ARX) model to model the wire drying process in stages, and optimizes parameters such as drum temperature and moisture removal based on the model's prediction results. However, the RBF-ARX model in this method has high computational complexity, is difficult to adjust parameters, carries the risk of overfitting, and has insufficient online updating capability. Furthermore, it has limited dynamic adaptability to nonlinear systems and is difficult to cope with complex production scenarios involving multivariate coupling.

[0004] Existing solutions focus on a single prediction model and lack a linkage mechanism between prediction and control, resulting in insufficient accuracy in controlling outlet moisture content. Therefore, there is an urgent need for an intelligent predictive control method that integrates the advantages of multiple models and combines dynamic parameter optimization to improve the accuracy of moisture control at the outlet of the yarn drying machine. Summary of the Invention

[0005] To address the insufficient accuracy of existing methods in controlling the moisture content at the outlet of a wire dryer, a predictive control method for a wire dryer based on multi-model fusion is proposed. This method fully utilizes the advantages of each model, and combines random forest, XGB Optimization, and AdaBoost algorithms through a weighted average algorithm to predict future trends in outlet moisture content. Based on the predicted trends, a genetic algorithm is used to dynamically optimize parameters and make corresponding control decisions, thereby achieving accurate prediction and adaptive control of outlet moisture content.

[0006] To achieve the above objectives, this invention proposes a predictive control method for a yarn drying machine based on multi-model fusion, comprising:

[0007] Step 1: Perform batch screening, moisture meter zero-point value correction, data alignment, and dimensionless processing on the production data to obtain preprocessed data;

[0008] Step 2: Use a combination of correlation analysis and random forest to screen features;

[0009] Step 3: Construct prediction models using three algorithms: Random Forest, XGBoost, and AdaBoost;

[0010] Step 4: Generate training and validation sets using historical production data. Use the training set to train the three prediction models respectively, and use the validation set to validate the three prediction models respectively.

[0011] Step 5: Determine the weights based on the mean square error of the three prediction models, and then use the weighted average to fuse the prediction results of the model to obtain the moisture prediction model of the drying machine outlet based on multi-model fusion.

[0012] Step 6: Based on the predicted trend of moisture change at the outlet, optimize and adjust the control parameters of the drying machine using a genetic algorithm.

[0013] Further, step 1 includes:

[0014] Step 1.1: Remove abnormal production batches and non-experimental batches, retaining batches that pass the process judgment. Remove outliers and fill in missing values ​​according to the brand's process standards; (Retain high-quality samples to avoid noisy data interfering with model training and improve data reliability.)

[0015] Step 1.2: Correct the moisture values ​​of each batch according to the zero point value of the moisture meter to ensure that the data are at the same zero point;

[0016] (Eliminate systematic biases caused by equipment errors to ensure the comparability of data from different batches.)

[0017] Step 1.3: Using the tobacco processing flow point as a reference point, calculate the correlation between parameter values ​​and outlet moisture and outlet temperature using a sliding window method. Find the time difference corresponding to the value with the largest correlation result as the time lag parameter, and use this time lag parameter to align the time series data of the feature variables with the process. (Determining the optimal time lag by calculating the correlation through a sliding window solves the problem of spatiotemporal misalignment between features and moisture values, enabling the model to capture the true causal relationship and improve prediction accuracy.)

[0018] Step 1.4: Perform dimensionless processing on all process parameters. (This eliminates unit differences between different parameters (such as temperature and wind speed), avoids features with large numerical ranges dominating model training, and ensures fair contribution of each variable to the results.)

[0019] Further, step 2 includes:

[0020] Step 2.1: Calculate the Pearson correlation coefficient between all candidate features and export moisture content. When the absolute value of the correlation coefficient is in the range of 0.2 to 1, it can be determined that there is a certain correlation between the two variables.

[0021] Step 2.2: Use a random forest model to perform preliminary training on the preprocessed data, obtain the Gini importance index of each feature, and retain the features with higher contribution.

[0022] Step 2.3: Perform a union operation on the features selected in Step 2.1 and Step 2.2 to obtain the final feature set.

[0023] A dual screening mechanism is adopted to quantify linear relationships, quickly filter irrelevant features, capture nonlinear relationships, identify features that have a significant impact on water but have weak linear correlation, and merge the two to ensure that the feature set fully covers both linear and nonlinear influencing factors, thereby improving the model's interpretability and generalization ability.

[0024] Furthermore, the random forest model in step 3 includes:

[0025] Step 3.1.1: Random Sampling: From the original training set D = {(x1,y1),(x2,y2),...,(x...} n ,y n In the process of sampling with replacement, T subsets D1, D2, ..., D are generated. T .

[0026] Step 3.1.2: For each subset D T (t = 1, 2, ..., T), construct a decision tree:

[0027] Randomly select features: at each node split, randomly select m features from all M features;

[0028] Select the optimal split point using formula (1):

[0029]

[0030] Repeat the above process until the stopping condition is met;

[0031] Step 3.1.3: Obtain the final result using formula (2):

[0032]

[0033] Among them, h t (x) represents the prediction output of the t-th decision tree for the input sample x.

[0034] Random forests extract multiple subsets of the original training dataset with replacement through bootstrapping, and then build a decision tree for each subset. For regression problems, the predictions from all decision trees are averaged. This effectively reduces model variance and is highly resistant to noise and overfitting.

[0035] Furthermore, the XGBoost model in step 3 includes:

[0036] Step 3.2.1: Initialize predicted values

[0037] Step 3.2.2: Iterative training:

[0038] For each tree from t=1 to T, calculate the first derivative g. i and second derivative h i ;

[0039] Based on g i and h i Construct the loss function:

[0040]

[0041] Where Ω is the regularization term, used to control the t-th decision tree f. (t) This reduces the complexity of the model and prevents it from overfitting.

[0042] Use a greedy algorithm to build a decision tree and update the predicted values:

[0043]

[0044] Among them, f (t) (x i Let t be the decision tree for sample x. i The predicted output;

[0045] Step 3.2.3: Output the final prediction model:

[0046]

[0047] Among them, f (k) (x i ) represents the sample x of the k-th tree. i The predicted output.

[0048] The XGBoost algorithm builds upon a decision tree by adding additional decision trees that contribute to reducing model error, thus theoretically resulting in higher accuracy. Furthermore, by incorporating regularization terms during training, it effectively prevents overfitting. When calculating the optimal split point for the objective function, utilizing second-order derivative information allows for a faster identification of the direction that leads to the steepest descent of the objective function. Compared to methods using only first-order derivatives, this significantly improves the efficiency of node splitting, thereby accelerating the overall model training speed. This rapid model training allows the model to adapt to changes promptly.

[0049] Furthermore, the Adaboost model in step 3 includes:

[0050] Step 3.3.1: Initialize the weights for each sample:

[0051]

[0052] Step 3.3.2: Perform iterative training, for each iteration from t=1 to T;

[0053] Using the current sample weight ω i (t) Training weak classifier h t (x);

[0054] The weighted error rate of the weak classifier is calculated using formula (7):

[0055]

[0056] Where γ is the indicator function;

[0057] Calculate the weights of the weak classifiers based on the weighted error rate:

[0058]

[0059] Step 3.3.3: Update the sample weights, increasing the weight of misclassified samples and decreasing the weight of correctly classified samples as shown in formula (9):

[0060] ω i (t+1) =ω i (t) exp(-α t y i ht (x i )) (9);

[0061] The weights are normalized as shown in formula (10):

[0062]

[0063] Step 3.3.3: The weighted combination of all weak classifiers is used to construct the final strong classifier as shown in formula (11):

[0064]

[0065] By iteratively adjusting sample weights, focusing on unpredictable outlier data points, the model's robustness to extreme values ​​is improved. Weak learners (such as single-layer decision trees) are computationally lightweight and suitable for complementing other models.

[0066] Furthermore, step 5 specifically includes:

[0067] Step 5.1: Evaluate the three prediction models using formula (12):

[0068]

[0069] Step 5.2: Calculate the weights of the three prediction models using formula (13):

[0070]

[0071] Among them, MSE i Represents the MSE of the i-th model;

[0072] Step 5.3: Use the weights and the model's prediction results to perform a weighted average, which is the output of the final fusion model;

[0073]

[0074] By combining the stability of random forests, the efficiency of XGBoost, and the outlier resistance of AdaBoost, the overall prediction accuracy is improved (significantly reducing MSE compared to a single model). Models with smaller errors have higher weights, ensuring that the prediction results are dominated by the better-performing model, further optimizing the reliability of the results.

[0075] Furthermore, step 6 specifically includes:

[0076] Step 6.1: Predict future export moisture content trends. When the moisture content deviation exceeds a set threshold, the control model takes effect. (The control model is only activated when the moisture content deviation exceeds the threshold to avoid frequent and ineffective adjustments, thereby reducing equipment wear and energy consumption.)

[0077] Step 6.2: Define the deviation between the predicted future export moisture content and the set value as the objective function. Based on the hot air fan frequency, dehumidification opening degree and thin plate temperature, randomly generate a set of initial values ​​for control parameters as the initial population of the genetic algorithm.

[0078] Step 6.3: Use a multi-model fusion-based moisture prediction model for the outlet of the drying machine to simulate the control parameters of each individual, calculate the objective function value as the fitness, and select the best individuals based on the fitness.

[0079] Step 6.4: Generate new individuals from multiple excellent individuals through crossover operation, increase diversity by mutation operation on some excellent individuals, and repeat fitness evaluation, selection, crossover and mutation operation until the number of iterations or fitness converges.

[0080] Step 6.5: Apply the optimized control parameters to the yarn drying machine and adjust them in real time to maintain stable outlet moisture content.

[0081] By employing a selection-crossover-mutation mechanism based on genetic algorithms, optimal solutions (such as the best match between hot air frequency and exhaust opening) are rapidly searched in a multi-parameter combination space. The predicted results are then used to directly drive equipment parameter adjustments, forming a real-time closed loop of "monitoring-prediction-control" with a fast response speed.

[0082] The beneficial effects of the present invention through the above technical solution are as follows:

[0083] (1) This invention combines three models—Random Forest, XGBoost, and AdaBoost—with the independent training and validation in step 4 and the weighted fusion mechanism in step 5, significantly improving prediction accuracy and stability. Random Forest reduces variance through bootstrapping and has strong anti-overfitting capabilities; XGBoost utilizes second-order derivative optimization to improve training speed and accuracy; and AdaBoost focuses on difficult samples to improve robustness against outliers. After independent training, the three models are weighted and fused using mean squared error, allowing each model to complement its strengths in different data scenarios, resulting in predictions that more closely reflect actual water content trends. Weights are dynamically allocated based on the model's performance on the validation set (higher weight for smaller errors), avoiding a single model dominating the results. This provides a more accurate basis for subsequent control decisions.

[0084] (2) This invention achieves a highly efficient transformation from "precise prediction" to "intelligent control" through a closed-loop linkage of multi-model fusion prediction and genetic algorithm control, significantly improving the accuracy of outlet moisture control and production adaptability. The multi-model fusion-based outlet moisture prediction model for the yarn drying machine outputs real-time future moisture prediction values. When the deviation exceeds a threshold, the genetic algorithm is immediately activated, using the minimization of deviation as the objective function to perform global optimization in the parameter space of hot air frequency, dehumidification opening, and other parameters. The response speed is extremely fast, allowing for rapid adaptation to fluctuations in production conditions such as raw material humidity changes and equipment wear. This significantly reduces quality fluctuations caused by process changes.

[0085] (3) This invention improves data quality and model input effectiveness from the source through a data preprocessing and feature screening linkage mechanism, which can prevent noise interference. At the same time, the feature union strategy retains multi-dimensional information, providing solid data support for subsequent multi-model fusion and control decision-making. Attached Figure Description

[0086] Figure 1 This is a flowchart illustrating the steps of a predictive control method for a wire drying machine based on multi-model fusion according to the present invention.

[0087] Figure 2 This is a flowchart of the algorithm for a predictive control method for a wire drying machine based on multi-model fusion, as described in this invention. Detailed Implementation

[0088] Example 1

[0089] like Figures 1-2 As shown, a predictive control method for a wire drying machine based on multi-model fusion includes:

[0090] Step 1: Perform batch screening, moisture meter zero-point value correction, data alignment, and dimensionless processing on the production data to obtain preprocessed data;

[0091] Step 2: Use a combination of correlation analysis and random forest to screen features;

[0092] Step 3: Construct prediction models using three algorithms: Random Forest, XGBoost, and AdaBoost;

[0093] Step 4: Generate training and validation sets using historical production data. Use the training set to train the three prediction models respectively, and use the validation set to validate the three prediction models respectively.

[0094] Step 5: Determine the weights based on the mean square error of the three prediction models, and then use the weighted average to fuse the prediction results of the model to obtain the moisture prediction model of the drying machine outlet based on multi-model fusion.

[0095] Step 6: Based on the predicted trend of moisture change at the outlet, optimize and adjust the control parameters of the drying machine using a genetic algorithm.

[0096] Step 1 includes:

[0097] Step 1.1: Remove abnormal production batches (broken or blocked materials) and non-test batches, and retain batches that pass the process judgment. Remove abnormal values ​​and fill in missing values ​​according to the brand process standards.

[0098] Step 1.2: Correct the moisture values ​​of each batch according to the zero point value of the moisture meter to ensure that the data are at the same zero point;

[0099] Step 1.3: Using the tobacco processing flow point as a reference point, calculate the correlation between parameter values ​​and outlet moisture and outlet temperature using a sliding window method, find the time difference corresponding to the value with the largest correlation result as the time delay parameter, and use this time delay parameter to perform process alignment on the time series data of characteristic variables.

[0100] Step 1.4: Dimensionless processing of all process parameters.

[0101] During the data acquisition process related to the moisture content of dried silk, the sensors are spaced a certain distance apart. The material, not sufficiently and evenly dispersed, often causes nonlinear changes in the characteristic data collected by the sensors during the vibrating conveyor belt transport process. Random forests, however, have a good effect on handling nonlinear problems.

[0102] Step 2 includes:

[0103] Step 2.1: Calculate the Pearson correlation coefficient between all candidate features and export moisture content. When the absolute value of the correlation coefficient is in the range of 0.2 to 1, it can be determined that there is a certain correlation between the two variables.

[0104] Step 2.2: Use a random forest model to perform preliminary training on the preprocessed data, obtain the Gini importance index of each feature, and retain the features with higher contribution.

[0105] Step 2.3: Perform a union operation on the features selected in Step 2.1 and Step 2.2 to obtain the final feature set.

[0106] According to Pearson correlation analysis, when the absolute value of the correlation coefficient is between 0.2 and 1, it can be determined that there is a certain correlation between the two variables, and a negative coefficient proves that there is a negative correlation between the variables. The union of the results of two feature selection methods is used to avoid missing key features. These key features include inlet moisture content, outlet moisture setpoint, ambient temperature, ambient humidity, cumulative flow rate, hot air fan frequency, hot air temperature, and steam pressure.

[0107] The random forest model in step 3 includes:

[0108] Step 3.1.1: Random Sampling: From the original training set D = {(x1,y1),(x2,y2),...,(x...} n ,y n In the process of sampling with replacement, T subsets D1, D2, ..., D are generated. T .

[0109] Step 3.1.2: For each subset D T (t = 1, 2, ..., T), construct a decision tree:

[0110] Randomly select features: at each node split, randomly select m features from all M features;

[0111] Select the optimal split point using formula (1):

[0112]

[0113] Repeat the above process until the stopping condition is met;

[0114] Step 3.1.3: Obtain the final result using formula (2):

[0115]

[0116] Among them, h t (x) represents the prediction output of the t-th decision tree for the input sample x.

[0117] The XGBoost model in step 3 includes:

[0118] Among them, h t (x) represents the prediction output of the t-th decision tree for the input sample x.

[0119] Step 3.2.1: Initialize predicted values

[0120] Step 3.2.2: Iterative training:

[0121] For each tree from t=1 to T, calculate the first derivative g. i and second derivative h i ;

[0122] Based on g i and h i Construct the loss function:

[0123]

[0124] Where Ω is the regularization term, used to control the t-th decision tree f.(t) To reduce complexity and prevent overfitting, a greedy algorithm is used to construct a decision tree and update the predicted values.

[0125]

[0126] Among them, f (t) (x i Let t be the decision tree for sample x. i The predicted output.

[0127] Step 3.2.3: Output the final prediction model:

[0128]

[0129] Among them, f (k) (x i ) represents the sample x of the k-th tree. i The predicted output.

[0130] The Adaboost model in step 3 includes:

[0131] Step 3.3.1: Initialize the weights for each sample:

[0132]

[0133] Step 3.3.2: Perform iterative training, for each iteration from t=1 to T;

[0134] Using the current sample weight ω i (t) Training weak classifier h t (x);

[0135] The weighted error rate of the weak classifier is calculated using formula (7):

[0136]

[0137] Where γ is the indicator function;

[0138] Calculate the weights of the weak classifiers based on the weighted error rate:

[0139]

[0140] Step 3.3.3: Update the sample weights, increasing the weight of misclassified samples and decreasing the weight of correctly classified samples as shown in formula (9):

[0141] ω i (t+1) =ω i (t) exp(-α t y i ht (x i )) (9);

[0142] The weights are normalized as shown in formula (10):

[0143]

[0144] Step 3.3.3: The weighted combination of all weak classifiers is used to construct the final strong classifier as shown in formula (11):

[0145]

[0146] In this embodiment, step 4: After the parameters of each model are set, they are trained using the same data. Using historical production data, 70% of the data is selected as the training set to train the three models respectively, and 30% of the data is selected as the validation set to validate the three models respectively.

[0147] Step 5 specifically includes:

[0148] Step 5.1: Evaluate the three prediction models using formula (12):

[0149]

[0150] Step 5.2: Calculate the weights of the three prediction models using formula (13):

[0151]

[0152] Among them, MSE i Represents the MSE of the i-th model;

[0153] Step 5.3: Use the weights and the model's prediction results to perform a weighted average, which is the output of the final fusion model;

[0154]

[0155] Step 6 specifically includes:

[0156] In this embodiment, step 6.1: the predicted trend of moisture change at the outlet; when the deviation of moisture change exceeds a set threshold, the control model starts to take effect.

[0157] The control model mainly uses genetic algorithms to optimize parameters, find the optimal combination of control parameters that can achieve the target water content, and then make a decision.

[0158] Step 6.2: Define the deviation between the predicted future export moisture content and the set value as the objective function. Based on the hot air fan frequency, dehumidification opening degree and thin plate temperature, randomly generate a set of initial values ​​for control parameters as the initial population of the genetic algorithm.

[0159] Step 6.3: Use a multi-model fusion-based moisture prediction model for the outlet of the drying machine to simulate the control parameters of each individual, calculate the objective function value as the fitness, and select the best individuals based on the fitness.

[0160] Step 6.4: Generate new individuals from multiple excellent individuals through crossover operation, increase diversity by mutation operation on some excellent individuals, and repeat fitness evaluation, selection, crossover and mutation operation until the number of iterations or fitness converges.

[0161] Step 6.5: Apply the optimized control parameters to the yarn drying machine and adjust them in real time to maintain stable outlet moisture content.

[0162] The method of the present invention is illustrated by the following example:

[0163] (1) Control objective: To stabilize the moisture content at the outlet of the drying machine at a set value of 13.0%, with an allowable fluctuation range of [missing value].

[0164] ±0.05%.

[0165] (2) Control parameters: hot air fan frequency (f, range: 10~50Hz), dehumidification opening (α, range: 20~50%), thin plate temperature (T, range: 129~131℃).

[0166] Objective function definition:

[0167] The cumulative absolute deviation (CAE) between the predicted export moisture value and the set value over the next 3 minutes is used as the objective function: CAE=∑∣cpred(t)-cset∣, where cpred(t) is the predicted export moisture value at minute t output by the prediction model.

[0168] Initial population generation:

[0169] An initial population of N = 20 individuals is randomly generated, with each individual representing a set of control parameters (f, α, T).

[0170] Simulation prediction and objective function calculation:

[0171] For each individual, a multi-model fusion model for predicting the moisture content at the outlet of the drying machine is used to simulate the trend of the outlet moisture content under its control parameters.

[0172] Forecast results: The export moisture content will fluctuate between 12.8% and 13.2% in the next 3 minutes, with a cumulative absolute deviation (CAE) of 1.8.

[0173] Calculate the CAE value for all individuals as fitness (the smaller the value, the better).

[0174] Selection, crossover, and mutation operations:

[0175] A roulette wheel selection algorithm is used to select superior individuals based on their fitness ratio to enter the next generation. Crossover and mutation operations are then performed on different individuals to generate new individuals.

[0176] Iteration termination condition:

[0177] The maximum number of iterations is set to 50, or convergence is achieved when the optimal CAE value changes by less than 0.05 over 10 consecutive generations. The optimal individual is defined as follows: (f = 35 Hz), tidal opening (α = 40%), and plate temperature (T = 130℃), corresponding to a CAE of 0.6.

[0178] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included within the scope of the present invention.

Claims

1. A predictive control method for a wire drying machine based on multi-model fusion, characterized in that, include: Step 1: Perform batch screening, moisture meter zero-point value correction, data alignment, and dimensionless processing on the production data to obtain preprocessed data; Step 2: Use a combination of correlation analysis and random forest to screen features; Step 3: Construct prediction models using three algorithms: Random Forest, XGBoost, and AdaBoost; Step 4: Generate training and validation sets using historical production data. Use the training set to train the three prediction models respectively, and use the validation set to validate the three prediction models respectively. Step 5: Determine the weights based on the mean square error of the three prediction models, and then use the weighted average to fuse the prediction results of the model to obtain the moisture prediction model of the drying machine outlet based on multi-model fusion. Step 6: Based on the predicted trend of export moisture content, optimize and adjust the control parameters of the drying machine using a genetic algorithm.

2. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, Step 1 includes: Step 1.1: Remove abnormal production batches and non-test batches, and retain batches that pass the process judgment. Remove abnormal values ​​and fill in missing values ​​according to the brand process standards. Step 1.2: Correct the moisture values ​​of each batch according to the zero point value of the moisture meter to ensure that the data are at the same zero point; Step 1.3: Using the tobacco processing flow point as a reference point, calculate the correlation between parameter values ​​and outlet moisture and outlet temperature using a sliding window method, find the time difference corresponding to the value with the largest correlation result as the time delay parameter, and use this time delay parameter to perform process alignment on the time series data of characteristic variables. Step 1.4: Perform dimensionless processing on all process parameters.

3. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, Step 2 includes: Step 2.1: Calculate the Pearson correlation coefficient between all candidate features and export moisture content. When the absolute value of the correlation coefficient is in the range of 0.2 to 1, it can be determined that there is a certain correlation between the two variables. Step 2.2: Use a random forest model to perform preliminary training on the preprocessed data, obtain the Gini importance index of each feature, and retain the features with higher contribution. Step 2.3: Perform a union operation on the features selected in Step 2.1 and Step 2.2 to obtain the final feature set.

4. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, The random forest model in step 3 includes: Step 3.1.1: Random Sampling: From the original training set D = {(x1,y1),(x2,y2),...,(x...} n ,y n In the process of sampling with replacement, T subsets D1, D2, ..., D are generated. T . Step 3.1.2: For each subset D T (t = 1, 2, ..., T), construct a decision tree: Randomly select features: at each node split, randomly select m features from all M features; Select the optimal split point using formula (1): Repeat the above process until the stopping condition is met; Step 3.1.3: Obtain the final result using formula (2): Among them, h t (x) represents the prediction output of the t-th decision tree for the input sample x.

5. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, The XGBoost model in step 3 includes: Step 3.2.1: Initialize predicted values Step 3.2.2: Iterative Training: For each tree from t=1 to T, calculate the first derivative g. i and second derivative h i ; Based on g i and h i Construct the loss function: Where Ω is the regularization term, used to control the t-th decision tree f. (t) This reduces the complexity of the model and prevents it from overfitting. Use a greedy algorithm to build a decision tree and update the predicted values: f (t) (x i Let t be the decision tree for sample x. i The predicted output. Step 3.2.3: Output the final prediction model: f (k) (x i ) represents the sample x of the k-th tree. i The predicted output.

6. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, The Adaboost model in step 3 includes: Step 3.3.1: Initialize the weights for each sample: Step 3.3.2: Perform iterative training, for each iteration from t=1 to T; Use the current sample weights Training weak classifier h t (x); The weighted error rate of the weak classifier is calculated using formula (7): Where γ is the indicator function; Calculate the weights of the weak classifiers based on the weighted error rate: Step 3.3.3: Update the sample weights, increasing the weight of misclassified samples and decreasing the weight of correctly classified samples as shown in formula (9): The weights are normalized as shown in formula (10): Step 3.3.3: The weighted combination of all weak classifiers is used to construct the final strong classifier as shown in formula (11):

7. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1: Evaluate the three prediction models using formula (12): Step 5.2: Calculate the weights of the three prediction models using formula (13): Among them, MSE i Represents the MSE of the i-th model; Step 5.3: Use the weights and the model's prediction results to perform a weighted average, which is the output of the final fusion model; 8. The predictive control method for a wire drying machine based on multi-model fusion according to claim 1, characterized in that, Step 6 specifically includes: Step 6.1: Predict the future trend of export moisture content. When the deviation of moisture content change exceeds the set threshold, the control model will take effect. Step 6.2: Define the deviation between the predicted future export moisture content and the set value as the objective function. Based on the hot air fan frequency, dehumidification opening degree and thin plate temperature, randomly generate a set of initial values ​​for control parameters as the initial population of the genetic algorithm. Step 6.3: Use a multi-model fusion-based moisture prediction model for the outlet of the drying machine to simulate the control parameters of each individual, calculate the objective function value as the fitness, and select the best individuals based on the fitness. Step 6.4: Generate new individuals from multiple excellent individuals through crossover operation, increase diversity by mutation operation on some excellent individuals, and repeat fitness evaluation, selection, crossover and mutation operation until the number of iterations or fitness converges. Step 6.5: Apply the optimized control parameters to the yarn drying machine and adjust them in real time to maintain stable outlet moisture content.

Images