Prediction method for dealkalization amount of float glass

Abstract

The invention relates to a method for predicting the dealkalization amount of float glass, which comprises the following steps of: collecting historical process data of a float glass production line from a tin bath to the tail end of an annealing furnace and the dealkalization amount of each position and orientation of the glass to form a data set D1; performing preprocessing and normalization processing on the data set D1 to form a data set D2; inputting the data set D2 into a plurality of established models for pre-training, and determining an optimization model; performing feature screening on parameters of dealkalization positions of plates in glass production to obtain an optimal feature subset; and inputting the optimal feature subset into the optimization model, optimizing parameters of the model through Bayesian optimization, and predicting the dealkalization amount. According to the method, the historical process data of the float glass production line from the tin bath to the tail end of the annealing furnace and the dealkalization amount of each position and orientation of the glass are collected, the model is selected, and then the model parameters are subjected to Bayesian optimization, so that the accuracy of predicting the dealkalization amount of the production line is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of glass manufacturing prediction, and particularly relates to a method for predicting the dealkalization amount of float glass. BACKGROUND

[0002] In the float glass manufacturing process, dealkalization is performed on the glass surface. Glass surface dealkalization refers to reacting the glass surface with a gas or a solution through gas or solution spraying within the annealing temperature range, reducing the alkali metal ion content of the glass surface, and thus changing the glass properties. Appropriate degree of dealkalization can improve the glass strength, glass chemical stability, prevent glass surface mold and improve the warping of high-aluminum glass after chemical strengthening. There are many factors affecting the dealkalization amount of glass on the float glass manufacturing line, and the interaction between the factors is not clear, resulting in inaccurate control of the degree of glass dealkalization. In the traditional method, engineers adjust the factors that may have an important effect and wait for the manufacturing line to fully respond, and then test in the laboratory to understand whether the dealkalization amount meets the standard. The production line that does not meet the standard needs to repeat the above process, which seriously affects the efficiency and yield of float glass production. Moreover, due to different environments and different glass formulations of different production lines, multiple adjustments are required. In the prior art, engineers analyze the production line data to find the correlation between certain specific factors and the dealkalization amount to obtain the relevant trend of the dealkalization amount, but due to different equipment and complex pipelines, the prediction is still inaccurate, and it is difficult to achieve good prediction effect through simulation experiments. SUMMARY

[0003] The technical problem of the present application is to process the glass manufacturing line process parameters and the glass dealkalization amount data set, and then model through model pre-selection, feature selection and model parameter Bayesian optimization to predict the glass dealkalization amount, which is suitable for different environments and production lines, reduces the production line debugging time, and improves the production efficiency and prediction accuracy of glass.

[0004] The purpose of the present application is to solve the above problems, and a method for predicting the dealkalization amount of float glass is proposed, comprising the following steps: S1: Collecting historical process data of the float glass production line from the tin bath to the end of the annealing furnace, and the dealkalization amount of the glass at each position and orientation to form a data set D1; S2: Preprocessing and normalizing the data set D1 to form a data set D2; S3: Inputting the data set D2 into the established multiple models for pre-training, and determining an optimized model; S4: Feature selection is performed on the parameters of the dealkalization position of the glass production slab to obtain an optimal feature subset; S5: Input the optimal feature subset into the optimization model, optimize the parameters of the model through Bayesian optimization, and predict the dealkalization amount.

[0005] Furthermore, in step S2, the preprocessing includes filling missing values ​​with a sliding window averaging method, setting a sliding window width, sliding the window sequentially and calculating the average value of the data within the window, and then filling the missing data with the average value; Preferably, in step S2, the preprocessing includes determining and processing abnormal values, calculating the median value of data that deviates from the normal value range due to sensor failure, manual input error, or abnormal event, and then determining abnormal values ​​based on the threshold value; Preferably, the determination of outliers includes the following steps: 1) If there are an odd number of data points in the dataset, the median is calculated as: ; Where X represents the data set and N represents the data points; 2) If the number of data points in the dataset is even, the median is calculated as: ; 3) Calculate the absolute deviation of the median using the formula: ; Where, X i Representation dataset X The i values, i Indicates a counting unit; 4) Determine data points X i The anomaly of the data point X i The deviation value is compared with the threshold. If it is greater than the threshold, it is judged as an abnormal value. The calculation formula of the deviation value is: Z= ; Where Z represents the deviation value of the data point.

[0006] Preferably, step S2 further includes converting the categorical features of different categories into binary features by using a hot encoding method for non-numerical categorical feature variables.

[0007] Furthermore, in step S2, normalization processing is performed to calculate the maximum and minimum values ​​of each column of data based on the number of glass defects and various process parameters, and then the data is transformed using the formula: ; Where, Represents the i-th data in the data set, represents the minimum value in the data column, represents the maximum value of the data column summary.

[0008] Further, in step S3, according to the non-uniformity of different positions on the plate surface and different influencing factors in the float glass production, a model is established for characterization and pre-selection of the prediction model according to the model index; the model index includes mean absolute error MAE, mean square error MSE and determination coefficient, and the calculation formula of the model index is: ; ; ; In the formula, y p represents the predicted value, y i represents the actual value, n represents the sample number, represents the mean value of the actual value, and i represents the count unit.

[0009] Further, in step S4, according to the recursive feature elimination method, all features input into the model are used to calculate the feature weight, and the features are sorted according to the absolute value of the weight, and the feature with the smallest weight is removed, and then the optimal feature subset is obtained through the calculation of each iteration.

[0010] Further, in step S5, the model output result is optimized by Bayesian optimization, a probability model is used to proxy the objective function, and a collection function is used to determine the next evaluation point to obtain the optimal solution of the prediction result, including the following steps: 1) initialization, a set of model output results are selected as initial evaluation points for training a Gaussian process model; 2) based on the initial evaluation points and the objective function value, a Gaussian process probability model is constructed; 3) selecting the next evaluation point, using the probability model, the next evaluation point is selected by the collection function; 4) calculate the true function value of the objective function; 5) update the probability model, add the new evaluation point to the probability model, and update the model parameters; 6) return the optimal solution by repeating iteration.

[0011] Preferably, the collection function includes an expected improvement function, a confidence upper bound function and a probability improvement function.

[0012] Compared with the prior art, the beneficial effects of the present application include: 1) The present invention provides a method for predicting the dealkalization amount of float glass. This method collects historical process data from the tin bath to the end of the annealing furnace on the float glass production line, as well as the dealkalization amount of glass at various positions and orientations. It selects a model and then performs Bayesian optimization on the model parameters, thereby improving the accuracy of the dealkalization amount prediction of the production line.

[0013] 2) The present invention provides a method for predicting the dealkalization amount of float glass. By dynamically selecting a model and comprehensively considering the relevant trends of the dealkalization amount, the method overcomes the influence of different environments, pipelines and production lines on the dealkalization amount, thereby improving the accuracy of the dealkalization amount prediction of the production line.

[0014] 3) Correlation between specific factors and dealkalization amount The trend of dealkalization amount is obtained, but due to different equipment and complex pipelines, the BRIEF DESCRIPTION OF THE DRAWINGS The present invention will be further described below with reference to the accompanying drawings and examples.

[0015] Figure 1 Schematic diagram of a flow chart of a method for predicting the dealkalization amount of float glass according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the change of indicators of the dealkalization amount preselection model according to an embodiment of the present invention; Figure 3 This is a schematic diagram showing the relationship between the number of feature deletions and the mean absolute error in an embodiment of the present invention; Figure 4 Schematic diagram for comparing the actual value and predicted value of dealkalization amount in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] like Figure 1 As shown, a method for predicting the dealkalization amount of float glass comprises the following steps: S1: Collect historical process data of the float glass production line from the tin bath to the end of the annealing furnace, as well as the dealkalization amount of the glass at various positions and orientations, to form data set D1; S2: preprocess and normalize the dataset D1 to form the dataset D2; In step S2, the preprocessing includes filling missing values ​​with a sliding window averaging method, setting the sliding window width, sliding the window sequentially and calculating the average value of the data in the window, and then filling the missing data with the average value; Divide the dataset D1 into different data groups. For the missing data at the nth position in the data column, take the average of the 5-20 data before and after the nth position to fill the missing data.

[0017] In step S2, pre-processing includes determining and processing abnormal values, calculating the median value of data that deviates from the normal value range due to sensor failure, manual input error, or abnormal event, and then determining abnormal values ​​based on the threshold value; The determination of the outlier includes the following steps: 1) If the number of data points in the data set is odd, the calculation formula of the median is: ; In the formula, M represents the median, X represents the data set, and N represents the number of data points; 2) If the number of data points in the data set is even, the calculation formula of the median is: ; 3) Calculate the absolute deviation of the median, and the calculation formula is: ; In the formula, X i X represents the value in the data set X ; i X represents the value in the data set i ; 4) Determine the data point X i is abnormal, compare the deviation value of the data point X i with the threshold value, if it is greater than the threshold value, it is determined to be an outlier, and the calculation formula of the deviation value is: Z= ; In the formula, Z represents the deviation value of the data point.

[0018] In step S2, the category characteristic variable of the non-numeric value is converted into a binary feature by using the hot encoding method.

[0019] In step S2, the normalization processing is performed, the maximum value and the minimum value of each column of data are calculated based on the number of glass defects and each process parameter, and then the formula is used for transformation, and the calculation formula is: ; In the formula, Xi represents the i-th data in the data set, min represents the minimum value in the data column, max represents the maximum value of the data column.

[0020] S3: input the data set D2 into the established multiple models for pre-training, and determine the optimized model.

[0021] In step S3, considering that the dealkalization process in the actual production line is non-uniform at different positions on the entire panel, the dealkalization amount Y of each position is distinguished and represented, the left air surface Y LA , the middle air surface Y MA , and the right air surface Y RA, left tin surface Y LS , middle tin surface Y MS , and right tin surface Y RS The influencing factors are not completely the same, respectively, a model is established to characterize and preselect the prediction model according to the model index, including a linear model, a random forest model, a multilayer perception neural network model, an XGB model and a support vector regression SVR model; the model index includes a mean absolute error MAE, a mean square error MSE and a determination coefficient, and the calculation formula of the model index is: ; ; ; In the formula, y p represents a predicted value, y i represents an actual value, n represents a sample number, represents a mean value of the actual value, and i represents a counting unit.

[0022] The fitting effect of the model is shown in Table 1 and Figure 2 .

[0023] S4: The parameters of the dealkalization position of the plate block in glass production are subjected to feature screening to obtain an optimal feature subset; As shown in Figure 3 , in step S4, all features are input into a model to calculate feature weights according to a recursive feature elimination method RFE, the features are sorted according to the absolute values of the weights, the feature with the smallest weight is removed, and then the optimal feature subset is obtained through the calculation of each iteration. Figure 3 The negative sign is added to the MAE number to obtain -MAE, and the larger the value is, the better the number of selected features is.

[0024] S5: The optimal feature subset is input into an optimized model, the parameters of the model are optimized through Bayesian optimization, and the dealkalization amount is predicted.

[0025] In step S5, the output results of the model are subjected to Bayesian optimization, a probability model is used to proxy the objective function, and a collection function is used to determine the next evaluation point to obtain an optimal solution of the prediction result, including the following steps: 1) initialization, a group of model output results are selected as initial evaluation points for training a Gaussian process model; 2) based on the initial evaluation points and the value of the objective function formula, a Gaussian process probability model is constructed; 3) the next evaluation point is selected, the probability model is used to select the next evaluation point through the collection function; 4) the true function value of the objective function is calculated; 5) update the probability model, add the new evaluation point to the probability model, update the model parameters; 6) return the optimal solution by repeating iteration.

[0026] The acquisition function includes the expected improvement function, the confidence upper bound function and the probability improvement function, and the calculation formula is as follows: 1) the calculation formula of the expected improvement function is as follows: ; ; In the formula, and denote the cumulative distribution function and the probability density function of the standard normal distribution, denote the current best point; denote the exploration parameter, and increase the exploration when it is larger.

[0027] 2) the calculation formula of the confidence upper bound is as follows: ; In the formula, denote the control of the trade-off between exploration and utilization.

[0028] 3) the calculation formula of the probability improvement is as follows: ; In the formula, denote the predicted value of f(x), denote the variance, denote the current best point, denote the exploration parameter, and increase the exploration when it is larger.

[0029] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this, any change or replacement that can be easily thought of by any person skilled in the art within the technical range disclosed by the present application should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the dealkalization amount of float glass, characterized in that: The following steps are involved: S1: Collect historical process data of the float glass production line from the tin bath to the end of the annealing furnace, as well as the dealkalization amount of the glass at various positions and orientations, to form data set D1; S2: preprocess and normalize the dataset D1 to form the dataset D2; S3: Input the dataset D2 into the established multiple models for pre-training, and determine the optimized model; S4: feature screening of parameters of dealkalization positions of glass plates during production to obtain an optimal feature subset; S5: Input the optimal feature subset into the optimization model, optimize the parameters of the model through Bayesian optimization, and predict the dealkalization amount.

2. The method for predicting the dealkalization amount of float glass according to claim 1, wherein: In step S2, the preprocessing includes filling the missing values ​​with a sliding window averaging method, setting the sliding window width, sliding the window in sequence and calculating the average value of the data in the window, and then filling the missing data with the average value.

3. The method for predicting the dealkalization amount of float glass according to claim 2, wherein: In step S2, the preprocessing includes determining and processing abnormal values, calculating the median of data that deviates from the normal value range due to sensor failure, manual input errors or abnormal events to obtain a threshold, and then determining abnormal values ​​based on the threshold.

4. The method for predicting the dealkalization amount of float glass according to claim 3, wherein: The determination of outliers includes the following steps: 1) If there are an odd number of data points in the dataset, the median is calculated as: ； Where X represents the data set and N represents the data points; 2) If the number of data points in the dataset is even, the median is calculated as: ； 3) Calculate the absolute deviation of the median using the formula: ； Where, X i Representing a dataset X The i values, i Indicates a counting unit; 4) Determine data points X i The anomaly of the data point X i The deviation value is compared with the threshold. If it is greater than the threshold, it is judged as an abnormal value. The calculation formula of the deviation value is: Z= ； Where Z represents the deviation value of the data point.

5. The method for predicting dealkalization amount of float glass according to claim 1, wherein: Step S2 also includes using a hot encoding method for non-numerical categorical feature variables to convert categorical features of different categories into binary features.

6. The method for predicting the dealkalization amount of float glass according to claim 1, wherein: In step S2, the normalization process is to calculate the maximum and minimum values ​​of each column of data based on the number of glass defects and various process parameters, and then convert them using the formula: ； Where, Represents the i-th data in the data set, Indicates the minimum value in the data column. Indicates the maximum value of the data column summary.

7. The method for predicting dealkalization amount of float glass according to claim 1, wherein: In step S3, according to the non-uniformity of different positions on the float glass surface and different influencing factors in the production of float glass, models are respectively established for characterization and the prediction model is pre-selected according to model indicators; the model indicators include mean absolute error (MAE), mean square error (MSE) and coefficient of determination, and the calculation formula of the model indicators is: ； ； ； Where y p Represents the predicted value, y i represents the actual value, n represents the number of samples, represents the mean of the actual value, and i represents the counting unit.

8. The method for predicting dealkalization amount of float glass according to claim 1, wherein: In step S4, all features are input into the model to calculate feature weights according to the recursive feature elimination method, and the features are sorted according to the absolute values ​​of the weights, and the features with the smallest weights are removed, and then the optimal feature subset is obtained through calculation in each iteration.

9. The method for predicting dealkalization amount of float glass according to claim 1, wherein: In step S5, the model output is subjected to Bayesian optimization, the probability model is used as a proxy for the objective function, and the acquisition function is used to determine the next evaluation point to obtain the optimal solution of the prediction result, including the following steps: 1) Initialization: Select a set of model output results as initial evaluation points for training the Gaussian process model; 2) Based on the initial evaluation point and the objective function value, a Gaussian process probability model is constructed; 3) Select the next evaluation point. Use the probability model to select the next evaluation point through the acquisition function. 4) Calculate the true function value of the objective function; 5) Update the probability model, add the new evaluation points to the probability model, and update the model parameters; 6) Return the optimal solution through repeated iterations.

10. The method for predicting the dealkalization amount of float glass according to claim 9, wherein: The acquisition function includes an expectation improvement function, a confidence upper bound function and a probability improvement function.

Images