An industrial process product quality estimation method based on a stacked ensemble learning model
Through stacked ensemble learning models, combined with convolutional neural networks and gradient enhancement decision trees, the problem of estimating the concentration of sodium aluminate solution in the multi-stage evaporation process of alumina production is solved, and accurate estimation and energy savings are achieved.
Patent Information
- Application Number
- CN202411452922.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-17
- Publication Date
- 2025-08-01
- Estimated Expiration
- 2044-10-17
AI Technical Summary
The prior art cannot meet the accurate estimate of the concentration of exported sodium aluminate solution in the multi-stage evaporation process during alumina production process, resulting in inefficient production efficiency and waste of energy.
Using a stacked ensemble learning model method, combined with convolutional neural network, gradient enhancement decision tree and random forest model, accurate estimation of the concentration of exported sodium aluminate solution is achieved through feature extraction and error compensation.
Improves the accuracy of model estimation, reduces the risk of overfitting, improves productivity and reduces energy consumption.
Smart Images
Figure CN119417292B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial process process index estimation, and particularly to an industrial process product quality estimation method based on a stacked ensemble learning model. Background Art
[0002] China is a major country in the production and consumption of the world's aluminum industry. Alumina, as a core industrial raw material, plays a key role in multiple fields. In 2023, the cumulative output of alumina in China reached 82.38 million tons, a year-on-year increase of 3.28%, accounting for 58% of the global output. More than 90% of the production of alumina uses the Bayer process, which has the characteristics of low energy consumption, short process, and high quality. The evaporation process of the Bayer process is a key process in alumina production, maintaining the water balance in the circulation system, enabling the mother liquor to evaporate and concentrate to meet the concentration requirements of bauxite digestion or the preparation of raw meal slurry, and discharging impurity salts. The evaporation process consists of equipment such as multi-stage evaporators and flash evaporators. The raw liquid takes about 2 hours from the inlet to the outlet, while the sodium aluminate solution at the outlet is sampled every 4 hours, and there is a serious lag in the detection of the concentration of the sodium aluminate solution at the outlet. The on-site workers adjust in a rough manner based on the lagging quality of the outlet product concentration, which not only reduces production efficiency but also wastes energy resources.
[0003] Existing measurement means, such as non-contact concentration meters and density meters, have increasingly strong measurement capabilities. However, considering the characteristics of the long process, multiple processes, and multi-variable correlation and coupling of the multi-stage evaporation process, they cannot meet the detection requirements of the concentration of the sodium aluminate solution at the outlet. Therefore, as a key index at the outlet of the evaporation process, accurately estimating the concentration of the sodium aluminate solution at the outlet of the industrial evaporation process is of great significance for ensuring the quality of the outlet product, reducing energy consumption, and stabilizing production operation. Summary of the Invention
[0004] The purpose of the present invention is to provide an industrial process product quality estimation method based on a stacked ensemble learning model to solve the problem of difficult estimation of the concentration of the sodium aluminate solution at the outlet in the multi-stage evaporation process of alumina production under the current industrial background.
[0005] To achieve the above purpose, the present invention provides an industrial process product quality estimation method based on a stacked ensemble learning model, including the following steps:
[0006] S1. Use a convolutional neural network model to preprocess and extract features from industrial evaporation operation data;
[0007] S2. Select and train base learners;
[0008] S3. Select a gradient boosting decision tree as the meta-learner of the stacked ensemble learning model, and use the prediction results of the base learners and the features extracted by the convolutional neural network model as inputs to estimate the concentration of the sodium aluminate solution at the outlet;
[0009] S4. Calculate the error between the estimated value and the actual measured value of the outlet solution concentration, retrain the random forest model for error compensation, and obtain the final estimated value of the outlet product quality.
[0010] Preferably, in step S1, the steps for feature extraction using a convolutional neural network model are as follows:
[0011] S11. Divide the data set into a training set and a test set in a ratio of 7:3, input the divided data into the trained convolutional neural network model to extract features, convert the two-dimensional data into three-dimensional data through a convolution operation, the convolutional neural network extracts local features from the data, and the convolutional layer is the core part of the convolutional neural network for extracting features from the input data. The formula for the convolution operation is:
[0012]
[0013] where I is the input image, K is the convolution kernel, (x, y) is the coordinate of the output feature map, and k is the size of the convolution kernel;
[0014] S12. Reduce the dimension of the feature map through a pooling layer, and at the same time extract the most obvious part of the features and input them into the lower layer, and select the maximum value from the 2×2 window;
[0015] P(x,y)=max(I(x+i,y+j))for i,j∈{1,2}
[0016] where P(x, y) is the position of the pooled feature map, and I(x + i, y + j) is the input value in the pooling window;
[0017] S13. The flattening layer flattens the multi-dimensional feature map into a one-dimensional vector full connection layer, and maps the flattened features to the final output. The flattening operation:
[0018] Flatten(P)=Vector(P1,P2,K,Pn)
[0019] where P is the pooled feature map, and Vector is the flattening operation that converts it into a one-dimensional vector.
[0020] Preferably, in step S2, the specific steps are as follows:
[0021] S21. Select 7 machine learning and deep learning models, including K-nearest neighbor algorithm, support vector regression, random forest, convolutional neural network, extreme gradient boosting, gated recurrent unit, and recurrent neural network for training; during the training process, optimize the model parameters to improve the model estimation performance;
[0022] S22. Conduct a comparative test on export product quality estimation data, input the data features extracted by the convolutional neural network model into the base learner, and compare the mean absolute percentage error, mean absolute error, root mean square error, and coefficient of determination of the data estimated by the selected algorithms to test the performance of each algorithm in estimating the features extracted by the convolutional neural network;
[0023] S23. Select the three models with the best estimation performance, namely K nearest neighbor algorithm, support vector regression and random forest as the ensemble model base learners; for each sample x in the test set test,j (j=1,2,...,m), estimated using different basis learners;
[0024] The estimation result of the base learner is expressed as:
[0025]
[0026] Among them, f k is the estimation function of the k-th base learner, is the estimation result of the jth test sample obtained by the kth base learner.
[0027] Preferably, in step S3, the estimation results of the three base learner models and the features extracted by the convolutional neural network model are used as input to combine into a new feature matrix X meta,test :
[0028]
[0029] Each row represents the estimation result of a test sample, each column represents the estimation of a base learner, and the result output by the meta-classifier is the estimated value of the integrated model; use the trained gradient boosting decision tree to evaluate the newly generated feature matrix X meta,test To make an estimate, the expression is as follows:
[0030]
[0031] Preferably, in step S4, calculating the error between the estimated value of the outlet solution concentration and the actual measured value specifically includes the following steps:
[0032] S411. Train the meta-learner and initialize the model as a simple estimator:
[0033]
[0034] S412. Calculate the residual between the estimated value and the actual value of the current model, that is, the outlet concentration estimation error:
[0035]
[0036] S413. Train a new weak learner decision tree on the current residuals. The goal of this model is to fit the residuals:
[0037]
[0038] S414. Update the model by weighting and adding the estimation results to the gradient boosting decision tree estimation model; the weighting factor is called the learning rate, which controls the contribution degree of each base learner to the final estimation:
[0039]
[0040] The estimated value of the final model is the weighted sum of the model estimated values in all iterations:
[0041]
[0042] S415. The stacked ensemble learning model screens out the best base learners through prediction experiments and uses the outputs of the base learners as the inputs of the meta-learner; then, the meta-learner is used for estimation to integrate the estimation results generated during the learning process of the base learners;
[0043] S416. Use the estimation results of the base learners and the features extracted by the convolutional neural network model as inputs to train a gradient boosting decision tree as the meta-learner; through training the meta-learner, the estimation result of the concentration of sodium aluminate solution at the industrial evaporation outlet is obtained.
[0044] Preferably, in step S4, the random forest model is retrained for error compensation to obtain the final estimated value of the export product quality, which specifically includes the following steps:
[0045] S421. Calculate the error value between the estimated value of the export solution concentration initially obtained by the model and the actual measured value;
[0046] S422. Use the calculated error value as the target value, and use the initial estimation results in the training data as features, that is, the estimated values of the base learners K-nearest neighbor algorithm, support vector machine, and random forest and the features extracted by the convolutional neural network model, and the error between the estimated value of the export solution concentration and the actual measured value, and input them into the random forest to obtain the estimated error value;
[0047] S423. Compensate the estimated error value to the estimated value output by the meta-learner to obtain the final concentration estimation result; through evaluation indicators, verify the feasibility of the estimation model for the concentration of sodium aluminate solution at the outlet.
[0048] Therefore, the beneficial effects of the present invention adopting the above industrial process product quality estimation method based on a stacked ensemble learning model are as follows:
[0049] (1) The quality estimation method of the present invention using a stacked ensemble learning model can maximize the extraction of multi-dimensional features of data variables through a Convolutional Neural Network (CNN) model, thereby improving the accuracy of model estimation.
[0050] (2) The quality estimation method of the present invention using a stacked ensemble learning model uses the K-nearest neighbor algorithm, support vector regression, and random forest as base learners to automatically integrate the advantages of the models, thus improving the performance of the overall estimation model. Using a gradient boosting decision tree model in the meta-learner model can reduce the risk of overfitting.
[0051] (3) The quality estimation method of the present invention using a stacked ensemble learning model applies a random forest error compensation model. The data features extracted by the CNN and the concentration estimated by the base learners are used as inputs. The random forest estimates the error and compensates it to the estimated value output by the meta-learner, and the final estimation result of the sodium aluminate solution at the outlet can be obtained.
[0052] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings
[0053] Figure 1 is a step diagram of an embodiment of the industrial process product quality estimation method based on a stacked ensemble learning model of the present invention;
[0054] Figure 2 is a flow chart of the stacked ensemble learning and error compensation model in an embodiment of the industrial process product quality estimation method based on a stacked ensemble learning model of the present invention;
[0055] Figure 3 is a comparison diagram of the estimated values and actual values of the sodium aluminate solution at the outlet of different models in an embodiment of the industrial process product quality estimation method based on a stacked ensemble learning model of the present invention;
[0056] Figure 4 is an error distribution diagram between the actual value and the compensated estimated value of the sodium aluminate solution at the outlet in an embodiment of the industrial process product quality estimation method based on a stacked ensemble learning model of the present invention. Detailed Embodiments
[0057] The technical solution of the present invention will be further described below with reference to the drawings and embodiments.
[0058] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings as understood by those of ordinary skill in the field to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "comprising" or "including" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.
[0059] Example 1
[0060] As Figure 1 shown, the present invention provides an industrial process product quality estimation method based on a stacked ensemble learning model, comprising the following steps:
[0061] S1. Using a Convolutional Neural Network (CNN) model, preprocess and extract features from industrial evaporation operation data, specifically including the following steps:
[0062] S11. Collect alumina production evaporation process variable data, including the flow rate of the original solution entering the evaporator, the flow rate of fresh steam, the flow rate of secondary steam, the flow rate of condensate, the concentration of the inlet feed liquid, etc. Use CNN to extract multi-dimensional features of the data variables to the maximum extent, construct a hierarchical representation of the data, from low-level details to high-level abstract concepts, enhancing the model's expressive ability and understanding depth.
[0063] Divide the data set into a training set and a test set in a ratio of 7:3, input the divided data into the trained convolutional neural network model to extract features, and convert the two-dimensional data (number of samples × number of features) into three-dimensional data (number of samples × number of features × number of channels) through convolution operations. CNN extracts local features in the data. The convolutional layer is the core part of CNN and is used to extract features from the input data. The formula for the convolution operation is:
[0064]
[0065] where I is the input image, K is the convolution kernel (filter), (x, y) are the coordinates of the output feature map, and k is the size of the convolution kernel.
[0066] S12. Reduce the dimension of the feature map through the pooling layer (MaxPooling1D), and at the same time extract the most obvious part of the features and input them into the lower layer, selecting the maximum value from the 2×2 window;
[0067] P(x,y)=max(I(x+i,y+j))for i,j∈{1,2}
[0068] Among them, P(x,y) is the position of the feature map after pooling, and I(x+i,y+j) is the input value in the pooling window;
[0069] S13. The flattening layer (Flatten) flattens the multi-dimensional feature map into a one-dimensional vector, and the fully connected layer (Dense) maps the flattened features to the final output. The flattening operation:
[0070] Flatten(P)=Vector(P1,P2,K,P n )
[0071] Among them, P is the feature map after pooling, and Vector is the flattening operation that converts it into a one-dimensional vector.
[0072] S2. Select and train the base learners, which specifically include the following steps:
[0073] S21. Select 7 machine learning and deep learning models, including K-Nearest Neighbor algorithm, Support Vector Regression, Random Forest, Convolutional Neural Network, Extreme Gradient Boosting (XGBoost), Gated Recurrent Unit (GRU), and Recurrent Neural Network (RNN) for training. During the training process, optimize the model parameters to improve the model estimation performance.
[0074] S22. Conduct a comparative experiment on the export product quality estimation data. Input the data features extracted by CNN into the base learners, and compare the Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R 2 ) of the estimated data of the selected algorithms to test the estimation performance of the above 7 models for the features extracted by CNN. The results are shown in Table 1:
[0075] Table 1 Comparison of error indicators for the estimated values of sodium aluminate solution exports by different base learners
[0076]
[0077] It can be seen from Table 1 that the KNN, SVR, and RF models perform well in terms of the MSE, MAE, and RMSE error indicators.
[0078] S23. Select the three models with the best estimated performance, namely the K-nearest neighbor algorithm, support vector regression, and random forest, as the base learners of the ensemble model. Use the training set to train KNN, SVR, and random forest as part of the ensemble learning, and estimate the concentration of sodium aluminate solution at the outlet respectively.
[0079] The goal of SVR is to find a regression function that minimizes the error between the predicted value and the actual value of the sample points while ensuring the lowest possible complexity of the model. The error of all data points from the regression line is within a preset threshold. The regression function of SVR has the following form:
[0080] f(x) = w T x + b
[0081] where x is the input feature vector, w is the weight vector, and b is the bias term. The goal of SVM is to find the optimal values of w and b to maximize the margin between the two classes in the feature space. When using a kernel function to handle non-linear problems, the regression function of SVR is:
[0082]
[0083] where α i and α i * are Lagrange multipliers, K(x i , x) is the kernel function used to map the input data to a high-dimensional space. Optimizing the following objective function to the minimum is expressed as:
[0084]
[0085] Subject to the following conditions:
[0086] y i - (w T + b) ≤ j + ξ i
[0087] (w T + b) - y i ≤ j + ξ i *
[0088] ξ i , ξ i * ≥ 0
[0089] ξ i and ξ i *is a slack variable used to handle errors outside the ε range. C is a regularization parameter that controls the trade-off between the complexity of the model and the error. j is the allowable range of regression error. By optimizing the above objective function, the optimal values of w and b and the Lagrange multipliers α i and α i * , and use the optimized SVR model to predict the concentration of the exported sodium aluminate solution.
[0090] The K-nearest neighbor algorithm is a simple and effective non-parametric classification and regression method. KNN is based on instance learning and uses distance metrics to classify or predict the class or value of new data points. KNN uses distance metrics to calculate the distance between samples. The most commonly used is the Euclidean distance, as shown below:
[0091]
[0092] Random forests improve the accuracy and stability of the model by constructing multiple decision trees and combining the estimation results. In the evaporation process of alumina production, random forests improve the prediction accuracy by integrating multiple decision trees. Since each decision tree is constructed based on a randomly selected subset of features and samples, the entire forest can reduce the risk of overfitting, thereby improving the generalization ability of the model.
[0093] For each sample x in the test set test,j (j = 1, 2,..., m), use different base learners for estimation;
[0094] The estimation results of the base learners are expressed as:
[0095]
[0096] where, f k is the estimation function of the k-th base learner, is the estimation result of the j-th test sample obtained by the k-th base learner.
[0097] S3. Select the gradient boosting decision tree as the meta-learner of the stacking ensemble learning model, and use the prediction results of the base learners and the features extracted by the convolutional neural network model as inputs to estimate the concentration of the exported sodium aluminate solution.
[0098] Combine the estimation results of the three base learner models and the features extracted by the convolutional neural network model as inputs to form a new feature matrix X meta,test :
[0099]
[0100] Each row represents the estimated result of a test sample, each column represents the estimation of a base learner, and the result output by the meta-classifier is the estimated value of the ensemble model; Use the trained gradient boosting decision tree to estimate the newly generated feature matrix X meta,test for estimation, as shown below:
[0101]
[0102] S4. Calculate the error between the estimated value of the outlet solution concentration and the actual measured value, retrain the random forest model for error compensation, and obtain the final estimated value of the outlet product quality.
[0103] Calculate the error between the estimated value of the outlet solution concentration and the actual measured value, which specifically includes the following steps:
[0104] S411. Train the meta-learner, and initialize the model as a simple estimator:
[0105]
[0106] S412. Calculate the residual between the estimated value and the actual value of the current model, that is, the outlet concentration estimation error:
[0107]
[0108] S413. Train a new weak learner decision tree on the current residual, and the goal of this model is to fit the residual:
[0109]
[0110] S414. Update the model by weighting and adding the estimation results to the gradient boosting decision tree estimation model; The weighting factor is called the learning rate, which controls the contribution degree of each base learner to the final estimation:
[0111]
[0112] The estimated value of the final model is the weighted sum of the model estimated values in all iterations:
[0113]
[0114] S415. The stacked ensemble learning model screens out the best base learners through prediction experiments, uses the outputs of the base learners as the inputs of the meta-learner; Then use the meta-learner for estimation and integrate the estimation results generated during the learning process of the base learners.
[0115] S416. Use the estimation results of the base learners and the features extracted by the convolutional neural network model as inputs to train a gradient boosting decision tree as a meta-learner; through training the meta-learner, obtain the estimation result of the sodium aluminate solution concentration at the industrial evaporation outlet.
[0116] Retrain the random forest model for error compensation to obtain the final estimated value of the export product quality, which specifically includes the following steps:
[0117] S421. Calculate the error value between the estimated value of the outlet solution concentration initially obtained by the model and the actual measured value;
[0118] S422. Use the calculated error value as the target value, and use the initial estimation results in the training data as features, that is, the estimated values of the base learners K-nearest neighbor algorithm, support vector machine, and random forest and the features extracted by the convolutional neural network model, and the error between the estimated value of the outlet solution concentration and the actual measured value, and input them into the random forest to obtain the estimated error value;
[0119] S423. Compensate the estimated error value to the estimated value output by the meta-learner to obtain the final concentration estimation result; through evaluation indicators, verify the feasibility of the sodium aluminate solution concentration estimation model at the outlet.
[0120] To further illustrate the effectiveness and feasibility of this embodiment, a comparative experiment was conducted using four models: K-nearest neighbor algorithm, support vector regression, random forest, and stacked ensemble learning. The flowchart is as Figure 2 shown. Using a total of 800 sets of data actually produced during the process operation cycle, the first 560 sets were selected for learning and training, and the last 240 sets were used for prediction verification.
[0121] Compare the mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R 2 ) of the estimation model in this embodiment with three estimation models. The results are shown in Table 2.
[0122] Table 2 Comparison of prediction error indicators for four methods
[0123]
[0124] As can be seen from Table 2, the stacked ensemble learning model has the best performance in terms of RMSE, MAE, MSE, and R 2It is the best among all models in terms of four indicators, with the smallest prediction error and more in line with the changing trend of the true value. For the RMSE indicator, the model proposed in this embodiment is 62.6% lower than KNN, SVR, and Random Forest. In the MAE indicator, compared with KNN, SVR, and Random Forest, the model proposed in this embodiment is reduced by 62.7%, 61.9%, and 62.1% respectively. In the MSE indicator, the model proposed in this embodiment is reduced by 86.3%, 84.2%, and 84.2% compared with KNN, SVR, and Random Forest respectively. For the R 2 indicator, the model proposed in this embodiment reaches 0.83, which is 82.0% higher than other models, indicating that this stacked ensemble learning model has better prediction performance than single models and can better solve the problem of estimating the concentration of sodium aluminate solution at the outlet of the multi-stage evaporation process in alumina production.
[0125] As Figure 3 shown, the estimated values of the concentration of sodium aluminate solution at the outlet obtained by four models are compared with the actual values. It can be seen from the figure that the diamond-shaped broken line of the estimation model in this embodiment is closer to the actual value represented by the circular broken line than other broken lines, indicating that the estimation error of the estimation model in this embodiment is smaller, the accuracy is higher, it can well track the changes of the circular broken line, and has strong stability.
[0126] As Figure 4 shown, it shows the central tendency and dispersion degree of the errors. The error distribution is close to the normal distribution, which indicates that the estimation errors of the model are evenly distributed without obvious bias, proving that the estimation ability of the model is relatively balanced as a whole. In addition, there are no obvious outliers in the figure, indicating that the model performs stably in most cases and is not significantly affected by individual data points. This result provides strong support for the reliability of the model, indicating that it can maintain good estimation quality in practical applications.
[0127] Therefore, the present invention adopts the above-mentioned method for estimating the quality of industrial process products based on a stacked ensemble learning model, and the estimation result has high accuracy, which can solve the problem of difficult estimation of the concentration of sodium aluminate solution at the outlet in the multi-stage evaporation process of alumina production under the current industrial background.
[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. An industrial process product quality estimation method based on a stacked ensemble learning model, characterized in that It includes the following steps: S1. Use a convolutional neural network model to preprocess industrial evaporation operation data and extract features; In step S1, the steps of using the convolutional neural network model to extract features are as follows: S11. Divide the data set into a training set and a test set according to a ratio of 7:3, input the divided data into the trained convolutional neural network model to extract features, convert two-dimensional data into three-dimensional data through convolution operations, and the convolutional neural network extracts local features from the data. The convolutional layer is the core part of the convolutional neural network and is used to extract features from the input data. The formula for the convolution operation is: where I is the input image, K is the convolution kernel, (x, y) is the coordinate of the output feature map, and k is the size of the convolution kernel; S12. Reduce the dimension of the feature map through the pooling layer, and at the same time extract the most obvious part of the features and input them into the lower layer, and select the maximum value from the 2×2 window; P(x,y)=max(I(x+i,y+j))for i,j∈{1,2} where P(x, y) is the position of the pooled feature map, and I(x + i, y + j) is the input value in the pooling window; S13. The flattening layer flattens the multi-dimensional feature map into a one-dimensional vector full connection layer, and maps the flattened features to the final output. The flattening operation: Flatten(P) = Vector(P1, P2,..., P n ) where P is the pooled feature map, and Vector is the one-dimensional vector converted by the flattening operation; S2. Select and train the base learner; S3. Select the gradient boosting decision tree as the meta-learner of the stacking ensemble learning model, use the prediction results of the base learner and the features extracted by the convolutional neural network model as inputs, and estimate the concentration of the sodium aluminate solution at the outlet; S4. Calculate the error between the estimated value of the outlet solution concentration and the actual measured value, retrain the random forest model for error compensation, and obtain the final estimated value of the outlet product quality; In step S4, calculating the error between the estimated value of the outlet solution concentration and the actual measured value specifically includes the following steps: S411. Train the meta-learner, and initialize the model as a simple estimator: S412. Calculate the residual between the estimated value and the actual value of the current model, that is, the outlet concentration estimation error: S413. Train a new model to train a new weak learner decision tree on the current residual, and the goal of this model is to fit the residual: [[ID= 17]]S414. Update the model to weight and add the estimation result to the gradient boosting decision tree estimation model; the weighting factor is called the learning rate, which controls the contribution degree of each base learner to the final estimation: The estimated value of the final model is the weighted sum of the model estimated values in all iterations: S415. The stacking ensemble learning model screens out the best base learner through prediction experiments, and uses the output of the base learner as the input of the meta-learner; then uses the meta-learner for estimation and integrates the estimation results generated during the learning process of the base learner; S416. Use the estimation results of the base learner and the features extracted by the convolutional neural network model as inputs, and train the gradient boosting decision tree as the meta-learner; through training the meta-learner, obtain the estimation result of the sodium aluminate solution concentration at the industrial evaporation outlet; In step S4, the random forest model is retrained for error compensation to obtain the final estimated value of the quality of the exported product, which specifically includes the following steps: S421. Calculate the error value between the estimated value of the concentration of the exported solution initially obtained by the model and the actually measured value; S422. Use the calculated error value as the target value, and use the initial estimation results in the training data as features, that is, the estimated values of the base learners K-nearest neighbor algorithm, support vector machine, and random forest and the features extracted by the convolutional neural network model, and the error between the estimated value of the concentration of the exported solution and the actually measured value, and input them into the random forest to obtain the estimated error value; S423. Compensate the estimated error value to the estimated value output by the meta-learner to obtain the final concentration estimation result; verify the feasibility of the estimated model of the concentration of the exported sodium aluminate solution through evaluation indicators.
2. The industrial process product quality estimation method based on a stacked ensemble learning model according to claim 1, wherein: In step S2, the specific steps are as follows: S21. Select 7 machine learning and deep learning models, including K-nearest neighbor algorithm, support vector regression, random forest, convolutional neural network, extreme gradient boosting, gated recurrent unit, and recurrent neural network for training; During the training process, optimize the model parameters to improve the model estimation performance; S22. Conduct a comparison experiment on the estimated data of the quality of the exported product. Input the data features extracted by the convolutional neural network model into the base learner, and compare the mean absolute percentage error, mean absolute error, root mean square error, and coefficient of determination of the estimated data of the selected algorithms to test the estimation performance of each algorithm for the features extracted by the convolutional neural network; S23. Select the three models with the best estimated performance, namely the K-nearest neighbor algorithm, support vector regression, and random forest, as the base learners of the ensemble model; for each sample x in the test set test,j (j = 1, 2,..., m), estimate using different base learners; The estimated result of the base learner is expressed as: where, f k is the estimation function of the k-th base learner, is the estimation result of the j-th test sample obtained by the k-th base learner.
3. The industrial process product quality estimation method based on a stacked ensemble learning model according to claim 1, wherein: In step S3, the estimation results of the three base learner models and the features extracted by the convolutional neural network model are used as inputs to form a new feature matrix X meta,test : Each row represents the estimated result of a test sample, and each column represents the estimate of a base learner. The result output by the meta-classifier is the estimated value of the ensemble model; the trained gradient boosting decision tree is used to estimate the newly generated feature matrix X meta,test as shown below: