Building ceramic sintering deformation prediction method based on machine learning

Abstract

The invention discloses an architectural ceramic sintering deformation prediction method based on machine learning, and the method comprises the steps: obtaining the moisture and granularity of powder at the top of a press, the temperature of a sintering kiln and sintering deformation detection data based on the actual production data of a certain enterprise, and carrying out the data matching, abnormal value processing, feature selection and data standardization processing of the collected data. The processed data is divided into a training set and a test set according to 4: 1, and a random forest and a support vector regression algorithm are selected to build a prediction model. The training set is used for model training and hyper-parameter adjustment, and the test set data is used for model verification. According to the method, the sintering deformation can be effectively predicted, and the product quality problem in the sintering process can be found in advance.

Description

Technical Field

[0001] The invention relates to the technical field of building ceramic quality prediction, and in particular to a method for realizing building ceramic firing deformation prediction based on machine learning. Background Art

[0002] Firing is the most important process in the production of building ceramics. It refers to the process in which the green body undergoes high-temperature firing in a kiln and continuously undergoes physical and chemical reactions to finally achieve a completely dense vitrification process. This process involves the transformation of multiple substances and the interaction of reaction kinetics and thermodynamics. Different reaction rates and reaction equilibrium positions will affect each other, and these physical and chemical reactions and interactions often overlap and occur in an interlaced manner without obvious temperature boundaries.

[0003] Firing deformation is the key product quality concern during the firing process. However, due to the complexity of the detection process, all inspections are done manually offline, which is not conducive to real-time monitoring of changes in product quality. For the same batch of products, the formula is fixed, and the firing deformation is greatly affected by the kiln temperature. This is mainly due to the complex chemical reactions of various chemical components at different temperatures. A prediction model for firing deformation can be established through the kiln temperature, but the theoretical model of the reaction mechanism of the ceramic material firing process is incomplete, including the kinetics of high-temperature reactions, system complexity and heterogeneity problems, so there is a certain degree of difficulty in modeling through mechanisms. Data-driven machine learning algorithms can quickly construct complex nonlinear relationships between input data and output targets, and machine learning has been widely used in the field of materials science in recent years. Summary of the invention

[0004] In order to solve the problem that the manual measurement of the sintering deformation of building ceramics takes a long time and thus cannot be monitored online in real time, the present invention proposes a method for predicting the sintering deformation of building ceramics based on machine learning.

[0005] Technical solution: The method for predicting the deformation of architectural ceramics during firing based on machine learning of the present invention comprises the following steps:

[0006] Step S1: Data collection. Collect the kiln temperature distribution data, forming powder particle size and moisture data, and firing deformation detection data at the kiln outlet during the production of a certain product by a certain enterprise, and align the data in time series.

[0007] Step S2: Data preprocessing. Perform outlier detection on the original data set and replace the outliers. Perform min-max normalization on the preprocessed data to obtain the final data set.

[0008] Step S3: Feature screening: Random forest feature importance ranking is used to screen the most critical features that affect firing deformation, achieve dimensionality reduction, and reduce model complexity.

[0009] Step S4: Model training. The reduced-dimensional data is divided into a training set and a test set at a ratio of 4:1. The training set data is used for cross-validation and algorithm hyperparameter tuning to determine the optimal model parameters.

[0010] Step S5: Model verification: The test set data is input into the trained prediction model to obtain the predicted value.

[0011] Furthermore, in step S1, there are 42-dimensional kiln temperature data, 3-dimensional forming powder particle size data and 1-dimensional forming powder moisture data, totaling 46-dimensional input variables. For each brick, the firing deformation data at 8 different positions are detected, and the variance is used in the present invention to represent the overall deformation degree. In order to correspond to the kiln temperature experienced by each brick, the time required for the blank to pass through each temperature sensor is calculated by the number of temperature sensors and the firing cycle, so as to match the kiln temperature experienced by the blank through the detection time of the firing deformation.

[0012] Furthermore, the powder particle size and moisture data in step S2 are detected for outliers using a box plot, and the outliers are replaced by the mean; the kiln temperature data are detected for outliers using linear regression, and the outliers are smoothed over time, and the outliers are corrected by the average of the observations at adjacent moments.

[0013] Furthermore, in step S3, the 9-dimensional most relevant characteristic variables are finally determined, including the temperatures at kilns M90, M29, M39, M41, M66, M51, T5, and M49 and the 60-mesh sieve residue of the powder, where M90 ​​represents the temperature of the kiln M90 sensor, and the same applies to the others.

[0014] Furthermore, the machine learning models in step S4 are random forest and support vector regression models. The grid search method is used to adjust the hyperparameters of each model. The random forest adjusts the number of decision trees, the maximum depth of the decision tree, the minimum number of leaf node samples, and the minimum number of samples required for node re-division; the support vector regression adjusts the regularization parameter and the kernel function coefficient.

[0015] Furthermore, when evaluating the prediction model in step S5, the determination coefficient, average relative error and root mean square error are used to comprehensively evaluate the prediction performance of the sintering deformation prediction model.

[0016] The technical effects of the present invention are as follows:

[0017] 1. Use machine learning models to predict firing deformation, and quickly build complex nonlinear relationships between input data and output targets, avoiding revealing the complex chemical reaction network in the firing process.

[0018] 2. Through the machine learning model, online real-time monitoring and prediction of the firing deformation of building ceramics are realized, solving the problem of long time cycle of traditional manual measurement methods.

[0019] 3. Through accurate prediction, conditions such as kiln temperature that may cause deformation can be discovered and adjusted in advance, thereby improving production efficiency and product quality and reducing the time and cost consumption caused by manual measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the process framework of the present invention;

[0021] Figure 2 It is the abnormal value result of the moisture and particle size data distribution and box plot detection of the press top powder in the present invention;

[0022] Figure 3 It is a schematic diagram of abnormal samples detected by linear regression of part of the temperature data of the firing kiln in the present invention;

[0023] Figure 4 This is a graph showing the result of ranking the importance of 49 dimensional features that affect sintering deformation in the present invention;

[0024] Figure 5a It is a diagram of the process of adjusting the hyperparameters of the selected SVR machine learning algorithm according to the present invention;

[0025] Figure 5b It is a diagram of the process of adjusting the hyperparameters of the selected RF machine learning algorithm according to the present invention;

[0026] Figure 6a It is the prediction result of the firing deformation by the SVR machine learning algorithm selected by the present invention;

[0027] Figure 6b It is the prediction result of firing deformation by the RF machine learning algorithm selected by the present invention. DETAILED DESCRIPTION

[0028] The present invention is further illustrated below in conjunction with specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0029] like Figure 1 As shown, a method for predicting the deformation of architectural ceramics fired based on machine learning in this embodiment is implemented by the following steps:

[0030] Step S1: Collect the kiln temperature distribution data, forming powder particle size and moisture data, and firing deformation detection data at the kiln outlet during the production process of a certain product by a certain enterprise as the initial sample. Some of the data are shown in Tables 1 and 2 below. Among them, the input variables include powder moisture, powder particle size (3 dimensions) and kiln temperature (42 dimensions); the output variable is firing deformation. In actual production, firing deformation data at 8 different positions were detected. In the present invention, variance is used to represent the overall deformation degree. After sorting, there are 223 firing deformation detection data in total.

[0031] Table 1 Some process variables affecting sintering deformation

[0032]

[0033] Table 2 Detection data of sintering deformation part

[0034]

[0035] At the same time, for other variables in actual production, the present invention makes the following assumptions: the powder particle size and moisture distribution are uniform, all product base materials and glazes are completely consistent, the horizontal temperature of the kiln is consistent, and the firing atmosphere, pressure and exhaust temperature are fixed. The fixed values ​​of other parameters are as follows: 340±20kg / cm 2 Forming pressure; 9.6±0.3mm brick thickness; 5.5±1.5 times / min punching times; 852×1681mm mold size; 200±20℃ maximum drying temperature.

[0036] Step 2: For the powder moisture and particle size data, use a box plot to remove outliers, and replace the outliers with the sample mean. The outlier detection results are as follows: Figure 2 As shown. Calculate the lower quartile Q1 as the number at position = (n+1) / 4 and the upper quartile Q3 as the number at position = 3(n+1) / 4, then the upper limit of the normal value Q max for:

[0037] Q max =Q3+1.5(Q3-Q1)

[0038] Lower limit of normal value Q min for:

[0039] Q min =Q3-1.5(Q3-Q1)

[0040] Linear regression outlier detection is used for kiln temperature data. The Cook distance is used to determine whether the sample is an abnormal sample. The Cook distance threshold is set to 0.04. If the Cook distance is greater than 0.04, it is considered an outlier. Time smoothing is used for the determined outlier, and the average value of the observations at adjacent moments is used to correct the outlier. Some of the outlier detection results are shown in Figure 2. Figure 3 shown.

[0041] The data set is then normalized using min-max standardization, and the sample values ​​are scaled to the interval [0, 1] to obtain the model data set.

[0042] Step S3: further screen the 49 features to reduce the feature dimension. The present invention uses the RF algorithm for feature selection. Figure 4 According to the feature importance ranking, the present invention finally determines 9-dimensional features for modeling, which are M90, M29, M39, M41, M66, powder 60 mesh particle size, M51, T5, and M49.

[0043] Step S4: Divide the reduced-dimensional data into training set and test set according to a ratio of 4:1. Use grid search and 3-fold cross validation to select the best parameters of each model. The parameter adjustment process is as follows: Figure 5a and Figure 5b By comparing the R 2 Finally, the optimal parameters of the RF algorithm are determined as follows: the number of decision trees is 15, the maximum depth of the decision tree is 9, the minimum number of samples for leaf nodes is 1, the minimum number of samples required for node re-division is 3, and the test set R 2 is 0.82; the optimal parameters of the SVR algorithm are: regularization parameter 1, kernel function coefficient 2, test set R 2 is 0.8.

[0044] Step S5: Based on the trained firing performance prediction model, the firing deformation is predicted using the test set data. The prediction results are as follows: Figure 6a and Figure 6b RF for the prediction of firing deformation of the test set R 2 is 0.885, MRE is 0.145, and RMSE is 0.066; the R 2 The RF algorithm is 0.808, the MRE is 0.194, and the RMSE is 0.085. Both of them can make relatively accurate predictions of the sintering deformation. In comparison, the prediction accuracy of RF is better than that of SVR algorithm.

[0045] The above-mentioned specific implementation methods are provided to more clearly introduce the examples of the present invention, rather than to restrict the implementation methods of the present invention. For technicians in the relevant fields, other different forms of changes can be made on the basis of the above description. For example, detection data of different periods, production of different products, selection of different algorithm parameters, etc., all implementation methods cannot be exhaustively listed here. However, the obvious changes derived from the spirit of the present invention are still within the protection scope of the present invention.

Claims

1. A method for predicting the deformation of architectural ceramics during firing based on machine learning, characterized in that: The following steps are involved: Step S1: data collection: collect kiln temperature distribution data, forming powder particle size and moisture data, firing deformation detection data at the kiln outlet, and align the data in time series; Step S2: Data preprocessing; performing outlier detection on the original data set and replacing the outliers; performing min-max normalization on the preprocessed data to obtain the final data set; Step S3: feature screening: using random forest feature importance ranking to screen the most critical features that affect firing deformation, achieve dimensionality reduction, and reduce model complexity; Step S4: model training; the reduced-dimensional data is divided into a training set and a test set according to a 4:1 ratio, and the training set data is used for cross-validation and algorithm hyperparameter tuning to determine the optimal model parameters; Step S5: Model verification: the test set data is input into the trained prediction model to obtain the predicted value.

2. The method for predicting the deformation of architectural ceramics during firing based on machine learning according to claim 1, characterized in that: In step S1, there are 42-dimensional kiln temperature data, 3-dimensional forming powder particle size data and 1-dimensional forming powder moisture data, totaling 46-dimensional input variables; for each brick, the firing deformation data at 8 different positions are detected, and the variance is used to represent the overall deformation degree; in order to correspond to the kiln temperature experienced by each brick, the time required for the blank to pass through each temperature sensor is calculated through the number of temperature sensors and the firing cycle, and the kiln temperature experienced is matched through the detection moment of the firing deformation of the blank.

3. The method for predicting the deformation of architectural ceramics during firing based on machine learning according to claim 1, characterized in that: The powder particle size and moisture data in step S2 are detected for outliers using a box plot, and the outliers are replaced by the mean; the kiln temperature data are detected for outliers using linear regression, and the outliers are smoothed over time, and the outliers are corrected by the average of the observations at adjacent moments.

4. The method for predicting the deformation of architectural ceramics during firing based on machine learning according to claim 1, characterized in that: In the step S3, the 9 most relevant characteristic variables are finally determined, including the temperatures at the kilns M90, M29, M39, M41, M66, M51, T5, and M49 and the 60-mesh sieve residue of the powder, where M90 ​​represents the temperature of the kiln M90 sensor, and the same applies to the others.

5. The method for predicting the deformation of architectural ceramics during firing based on machine learning according to claim 1, characterized in that: The machine learning models in step S4 are random forest and support vector regression models; the grid search method is used to adjust the hyperparameters of each model. The random forest adjusts the number of decision trees, the maximum depth of the decision tree, the minimum number of samples of leaf nodes, and the minimum number of samples required for node re-division; the support vector regression adjusts the regularization parameter and the kernel function coefficient.

6. The method for predicting the deformation of architectural ceramics during firing based on machine learning according to claim 1, characterized in that: When evaluating the prediction model in step S5, the determination coefficient, average relative error and root mean square error are used to comprehensively evaluate the prediction performance of the sintering deformation prediction model.

Images