Method for predicting number of glass defects on glass manufacturing production line

Abstract

The invention discloses a method for predicting the number of glass defects on a glass manufacturing production line. The method comprises the following steps: collecting historical process data of the glass production line and the corresponding number of various glass defects; processing missing values and abnormal values of the collected data, and filling the missing values by adopting a neighbor value method; determining each related process parameter and defect lag duration by adopting a correlation coefficient method; performing n-order differential processing on the glass defect data; carrying out dimension reduction processing on the process parameter data set, and selecting a part of process parameters influencing glass defects; performing normalization processing on all the data; a deep learning model is selected to carry out modeling on the processed data; and carrying out later prediction on glass defects. The method is used for solving the problems that the variation trend of the future glass defects of the production line is difficult to identify and the number of the glass defects is difficult to predict due to numerous influence factors on various glass defects on the glass manufacturing production line and various interaction effects exist among the factors.

Description

Technical Field

[0001] The invention relates to a method for predicting the number of glass defects on a glass manufacturing production line. Background Art

[0002] Common defects in glass production lines include bubbles, glass tendons, stones, and exploded panels. There are many factors that affect various glass defects on the glass manufacturing production line, and there are many interactions between the factors, making it difficult to identify the changing trend of glass defects on the production line in the future. Especially when the production line engineer makes adjustments to the relevant factors, it takes time to wait for the production line to respond (the glass manufacturing process takes at least 48 hours from the raw material end to the finished product stage) to understand whether the factor is effective for a certain defect. Also, due to the complexity of the glass manufacturing process, the floating adjustments of the same factors at different times may have different effects, or even opposite effects. Summary of the invention

[0003] The purpose of the present invention is to provide a method for establishing a glass production line defect model, which is used to solve the problem that there are many factors affecting various glass defects on the glass manufacturing production line, and there are multiple interactions between the factors, making it difficult to identify the changing trend of glass defects in the production line in the future and difficult to predict the number of glass defects.

[0004] In order to solve the above problems, the technical solution of the present invention is: A method for predicting the number of glass defects on a glass manufacturing production line comprises the following steps: Step 1: Collect historical process data of the glass production line and the corresponding number of various glass defects; Step 2: Process missing values ​​and outliers in the collected data, and fill in the missing values ​​using the neighboring value method; Step 3: Use the correlation coefficient method to determine the duration of each relevant process parameter and defect lag; Step 4: Perform n-order difference processing on the glass defect data; Step 5: Perform dimensionality reduction processing on the process parameter data set and select some process parameters that have an impact on glass defects; Step 6: Normalize all data; Step 7: Select a deep learning model to model the processed data; Step 8: Make a late prediction of glass defects.

[0005] In the second step, the outlier determination and processing method includes: for a data set X containing N data points, the median calculation formula is as follows: If N is odd: If N is even: Median absolute deviation MAD = median (|X i-median(X)|), where X i Represents each data point; If the data point X i satisfy The data points identified as outliers are replaced by the median (X).

[0006] The threshold is 1 to 4, and each single process parameter data set is divided into a plurality of N data point sets, and the outliers in each set are judged and replaced respectively, and N is at least selected to be ≥4.

[0007] In the third step, the correlation coefficient is calculated as: In the glass manufacturing process, the maximum lag time is set to a value ≥72h as the maximum lag time to calculate the process parameters and the lag time of glass defects.

[0008] In the fourth step, the first-order difference, that is, the current observation minus the previous observation, constitutes the difference term, and its mathematical expression is: By analogy, the n-order difference processing can be performed with the following mathematical expression:

[0009] In the fifth step, the method of selecting some process parameters that affect glass defects includes: decomposing the original characteristic data set matrix A (m rows and n columns) into three matrices U, Σ, and V, namely Where U and V are both orthogonal matrices, that is, U T U=E m×m 、V T V=E n×n ; The columns of the U matrix are composed of AA T The eigenvectors of , and the eigenvectors are unit column vectors; the diagonal elements of the Σ matrix come from AA T The square root of the eigenvalues ​​of , and they are arranged in descending order. The larger the singular value, the more important the corresponding feature is. The V matrix columns are composed of A T The eigenvector of A is composed of unit column vectors. The first r process parameters are selected, and the sum of the singular values ​​of these r process parameters is required to account for 90% to 99% of all singular values.

[0010] In the sixth step, the method of normalizing all data includes: calculating the maximum and minimum values ​​of each column of data, and then using the formula to convert.

[0011] In the seventh step, the method of modeling the processed data includes: dividing the data set into continuous non-overlapping blocks according to a certain time window, using these blocks as basic data elements for data modeling, and establishing a connection between the current state and historical data through a long short-term memory network. The long short-term memory network is controlled by three gates, namely, a forget gate, an update gate, and an output gate.

[0012] The beneficial effects of the present invention are: 1. The data processing, model building and prediction of the present invention do not rely on the previous relevant experience of engineers, and by inputting the process parameters assumed to be modified to adjust the target value, the changing trend of glass defects can be predicted.

[0013] 2. The present invention calculates the correlation between different lag time lengths of process parameters and glass defects, and can obtain the specific lag time lengths of process parameters and glass defects, and can establish a data model with better predictive performance.

[0014] 3. The present invention processes, models and predicts data, eliminating the need for simulation to consider various complex physical and chemical reactions, liquid reflux and other complex environments.

[0015] In summary, by performing certain pre-processing on the glass manufacturing production line process parameters and glass defect data sets, and selecting important process parameters in a special way, and modeling them, the subsequent glass defects are predicted based on the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings: Figure 1 is a flow chart of the present invention, Figure 2 This is a flow chart of the implementation of the forget gate, update gate, and output gate in the seventh step of the present invention. Figure 3 This is a graph of the loss values ​​of the training set and the validation set in the eighth step of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] A method for predicting the number of glass defects on a glass manufacturing production line comprises the following steps: Step 1: Collect historical process data of the glass production line and the corresponding number of various glass defects; Step 2: Process missing values ​​and outliers in the collected data, and fill in the missing values ​​using the neighboring value method; Step 3: Use the correlation coefficient method to determine the duration of each relevant process parameter and defect lag; Step 4: Perform n-order difference processing on the glass defect data; Step 5: Perform dimensionality reduction processing on the process parameter data set and select some process parameters that have an impact on glass defects; Step 6: Normalize all data; Step 7: Select a deep learning model to model the processed data; Step 8: Make a late prediction of glass defects.

[0019] In the second step, the outlier determination and processing method includes: for a data set X containing N data points, the median calculation formula is as follows: If N is odd: If N is even: Median absolute deviation (MAD) = median | X i -median(X)|), where X i Represents each data point; If the data point X i satisfy The data points identified as outliers are replaced by the median (X).

[0020] The threshold is 1 to 4, and each single process parameter data set is divided into a plurality of N data point sets, and the outliers in each set are judged and replaced respectively, and N is at least selected to be ≥4.

[0021] In the third step, since the adjustment of process parameters in the glass manufacturing process will cause different lags in the glass defects at the same time, the present invention uses the correlation coefficient method to determine the lag time between each relevant process parameter and the defect. If the lag is not considered, the subsequent data modeling processing will result in inaccurate prediction results. The principle is: assuming that A is a process parameter that affects the B glass defect, calculate the correlation coefficient between A lag 1 unit time, 2 unit time, 3 unit time... and the glass defect, and calculate its absolute value, and determine the unit time with the largest absolute value as the lag time between A process parameter and B glass defect.

[0022] The formula for calculating the correlation coefficient is: Where R is the correlation coefficient, n is the number of samples, and x i is the value of a point in the X data set (a process parameter), y iis the value of a point in the Y data set (a glass defect), is the average value of the X data set (a process parameter), is the average value of the Y data set (a glass defect).

[0023] In the glass manufacturing process, the maximum lag time is set to a value ≥72h as the maximum lag time to calculate the process parameters and the lag time of glass defects.

[0024] In the fourth step, the first-order difference, that is, the current observation minus the previous observation, constitutes the difference term, and its mathematical expression is: By analogy, the n-order difference processing can be performed with the following mathematical expression: is the first-order difference result of the X data set, is the n-order difference result of the X data set, X t is the value of the X data set at time t, X t-1 is the value of the X data set at time t-1, is the value of the X data set at the time t of the n-1 order difference, It is the value of the n-1-order difference of the X data set at time t-1.

[0025] The purpose of this step is to process the non-stationary series of glass defect data into a stationary series, so that accurate prediction can be made after the data model is established.

[0026] In the fifth step, because there are many process parameters on the glass manufacturing production line, if all of them are used for data modeling, the complexity of data calculation will increase, the establishment of data models will be complicated, the data model will easily produce overfitting risks, the generalization ability of the algorithm will be poor, and the prediction ability of the later model will be poor. Therefore, the present invention selects some process parameters that affect glass defects, and the method includes: decomposing the original feature data set matrix A (m rows and n columns) into three matrices U, Σ, and V. That is, In this formula, U and V are both orthogonal matrices, that is, U T U=E m×m , V T V=E n×n The columns of the U matrix are composed of AA T The eigenvectors of , and the eigenvectors are unit column vectors. The diagonal elements of the Σ matrix come from AA T The square root of the eigenvalues ​​of is arranged in descending order. The larger the value (named singular value), the more important the corresponding feature is. The V matrix columns are composed of A TThe characteristic vector of A is a unit column vector. The appropriate process parameters can be selected according to the importance of the characteristics. The first r process parameters can be selected, and the sum of the singular values ​​of these r process parameters is required to account for 90% to 99% of all singular values.

[0027] U, Σ, V, and A are matrices with m rows and m columns, m rows and n columns, n rows and n columns, and m rows and n columns, respectively. T is the transposed matrix of U, V T is the transposed matrix of V, A T is the transposed matrix of A. E m×m and E n×n They are the m-row and m-column unit matrix and the n-row and n-column unit matrix respectively. r is the number of selected process parameters.

[0028] In the sixth step, because the number of glass defects and the dimensions of each process parameter are different, the data range is also very different. As a result, when mathematical modeling is performed, the process parameter with larger data will have a greater impact on the defects, which may not be consistent with the actual situation. Therefore, it is necessary to normalize each column of data. The method of normalizing all data includes: calculating the maximum and minimum values ​​of each column of data, and then using the formula to convert.

[0029] In the seventh step, the method of modeling the processed data includes: first, according to a certain time window (q unit time, q can be 6 to 72 hours), the data set is divided into continuous non-overlapping blocks, and these blocks are used as basic data elements for data modeling (such as [X1, X2, X3...X6], [X2, X3, X4...X7]...). The current state is connected with the historical data through the long short-term memory network. The long short-term memory network is controlled by three gates, namely the forget gate, the update gate, and the output gate. Figure 2 shown.

[0030] Where: f , Γ u and Γ o They represent the parameters of the forget gate, update gate, and output gate respectively; σ(·) is a nonlinear activation function, which generally uses the Sigmoid function, but can also use the Relu function, Tanh function, etc.; W fa , W fx , W ua 、V ux , W oa , W ox , W ca , W cx are the weight matrices of each part; b f , b u , bo , b c is the bias matrix; x (t) is the input of the network unit at the current moment; a (t-1) 、a (t) are the outputs of the network unit at the previous moment and the current moment respectively. Tanh(·) is the hyperbolic tangent function, which is the default state activation function of the network.

[0031] The nonlinear activation function can be Sigmoid function, Relu function, Tanh function, etc.

[0032] Sigmoid function: Relu function: f(x)=max(0,x); Tanh function: Then, 60% to 80% of the previous non-overlapping block data set is used as the training set, and the remaining part is used as the validation set. The feature data column in the training set is used as the input of the model, and the glass defect data column is used as the output of the model.

[0033] The parameters that can be set for this model include the number of training times (repeatedly using the model for training), the choice of activation function, the number of LSTM network layers (multi-layer LSTM can increase the predictive ability of the model), and the number of batches (the amount of data given to the model for training each time). The quality of the model can be evaluated by the size of the loss value and the goodness of fit. The loss value can be evaluated using the formula value of MSE or MAE.

[0034] Among them, y p is the predicted value, y i is the actual value, and n is the number of samples.

[0035] By optimizing the above model parameters, the model's goodness of fit R in the validation set is 2 Reach more than 80%. The goodness of fit formula is as follows: Among them, y p is the predicted value, yi is the actual value, is the mean of the actual value, and n is the number of samples. 2 The larger it is, the better the model fit is.

[0036] In the eighth step, the short-term glass defect prediction results can be generated based on the data model established above. Figure 3 As shown, it is used as a reference for production line to make relevant decisions.

[0037] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for predicting the number of glass defects on a glass manufacturing production line, characterized in that: The process includes the following steps: Step 1: Collect historical process data of the glass production line and the corresponding number of various glass defects; Step 2: Process missing values ​​and outliers in the collected data, and fill in the missing values ​​using the neighboring value method; Step 3: Use the correlation coefficient method to determine the duration of each relevant process parameter and defect lag; Step 4: Perform n-order difference processing on the glass defect data; Step 5: Perform dimensionality reduction processing on the process parameter data set and select some process parameters that have an impact on glass defects; Step 6: Normalize all data; Step 7: Select a deep learning model to model the processed data; Step 8: Make a late prediction of glass defects.

2. The method for predicting the number of glass defects on a glass manufacturing production line according to claim 1, characterized in that: In the second step, the outlier determination and processing method includes: for a data set X containing N data points, the median calculation formula is as follows: If N is odd: If N is even: Median absolute deviation MAD = median (|X i -median(X)|), where X i Represents each data point; If the data point X i satisfy Data points identified as outliers are replaced by the median value (X).

3. The method for predicting the number of glass defects on a glass manufacturing line according to claim 2, characterized in that: The threshold is 1 to 4, and each single process parameter data set is divided into a plurality of N data point sets, and the outliers in each set are judged and replaced respectively, and N is at least selected to be ≥4.

4. The method for predicting the number of glass defects on a glass manufacturing line according to claim 1, characterized in that: In the third step, the correlation coefficient is calculated as: In the glass manufacturing process, the maximum lag time is set to a value ≥72h as the maximum lag time to calculate the process parameters and the lag time of glass defects.

5. The method for predicting the number of glass defects on a glass manufacturing line according to claim 1, characterized in that: In the fourth step, the first-order difference, that is, the current observation minus the previous observation, constitutes the difference term, and its mathematical expression is: By analogy, the n-order difference processing can be performed with the following mathematical expression:

6. The method for predicting the number of glass defects on a glass manufacturing line according to claim 1, characterized in that: In the fifth step, the method of selecting some process parameters that affect glass defects includes: decomposing the original characteristic data set matrix A (m rows and n columns) into three matrices U, Σ, and V, namely Where U and V are both orthogonal matrices, that is, U T U=E m×m 、V T V=E n×n ; The columns of the U matrix are composed of AA T The eigenvectors of , and the eigenvectors are unit column vectors; the diagonal elements of the Σ matrix come from AA T The square root of the eigenvalues ​​of , and they are arranged in descending order. The larger the singular value, the more important the corresponding feature is. The V matrix columns are composed of A T The eigenvector of A is composed of unit column vectors. The first r process parameters are selected, and the sum of the singular values ​​of these r process parameters is required to account for 90% to 99% of all singular values.

7. The method for predicting the number of glass defects on a glass manufacturing line according to claim 1, characterized in that: In the sixth step, the method of normalizing all data includes: calculating the maximum and minimum values ​​of each column of data, and then using the formula to convert.

8. The method for predicting the number of glass defects on a glass manufacturing line according to claim 1, characterized in that: In the seventh step, the method of modeling the processed data includes: dividing the data set into continuous non-overlapping blocks according to a certain time window, using these blocks as basic data elements for data modeling, and establishing a connection between the current state and historical data through a long short-term memory network. The long short-term memory network is controlled by three gates, namely, a forget gate, an update gate, and an output gate.

Images