Convolutional neural network-based sintering mixture trough material level prediction method and system, and memory

Abstract

The invention discloses a method and a system for predicting the material level of a sintering mixture trough based on a convolutional neural network, and a memory. According to the method, characteristic variables related to the material level of the mixture trough are screened through a random forest algorithm to serve as model input parameters, autocorrelation analysis is conducted on time sequence data of the input variables, and a two-dimensional characteristic variable matrix is constructed; inputting the two-dimensional matrix into a pre-trained mixture trough material level prediction model based on a convolutional neural network to obtain mixture trough material level output; based on the above method, the invention develops a system for predicting the material level of the sintering mixture trough, and the system can predict the material level of the mixture trough in advance, so that an important basis is provided for an operator to timely discover the change of the material level of the mixture trough and accurately regulate and control operation parameters.

Description

Technical Field

[0001] The present invention relates to a method for predicting the material level of a sintering mixing trough, and in particular to a method, system and memory for predicting the material level of a sintering mixing trough based on a convolutional neural network, and belongs to the field of iron ore sintering production. Background Art

[0002] Sintering is an important process in the steel production process. Its stable, efficient and low-consumption operation is the prerequisite for reducing costs and increasing efficiency in the steel production process. After batching, mixing and granulation, the iron ore mixture is transported to the mixing trough, and then unloaded onto the sintering machine trolley through the round rollers and multi-roller distributors under the trough. The mixing trough plays a role of storage and buffering. Maintaining the stability of the material level in the mixing trough is a key link in the stable and smooth sintering production. Too low a material level can easily lead to uneven distribution, and the material surface of the sintering machine trolley will be high in the middle and low on both sides, increasing the load on the fan. At the same time, it may need to be shut down due to insufficient material level. If the material level is too high, it will cause extrusion and damage to the granulated balls of the mixture, affecting the air permeability of the mixture.

[0003] In the actual production process, operators are usually required to pay attention to the changes in the material level of the mixing tank at all times and make operational controls. However, the fluctuations in the material level of the mixing tank are relatively frequent and rapid, and the response time of the material level of the mixing tank to various operations is relatively long. Therefore, it is necessary to establish a prediction model for the material level of the mixing tank, make advance predictions on the material level changes, and make timely and accurate adjustments to the operating parameters to reduce the fluctuations in the material level of the mixing tank. The prediction of the material level of the mixing tank plays an important auxiliary role in improving the stability of the sintering process. Summary of the invention

[0004] In view of the problems existing in the prior art, the first purpose of the present invention is to provide a method for predicting the material level of a sintering mixing tank based on a convolutional neural network. This method selects characteristic parameters related to the material level of the mixing tank, selects data obtained by each characteristic parameter with time as the scale, forms a two-dimensional characteristic variable matrix, and constructs a mixing tank material level prediction model through a convolutional neural network. Based on the correlation between the sintering characteristic parameters and the material level of the mixing tank, the obtained prediction method not only ensures the prediction result rate of the mixing tank material level, but also greatly reduces the influence of the fluctuation of the input variables on the prediction result, thereby greatly improving the prediction accuracy.

[0005] The second object of the present invention is to provide a readable memory including a computer program for implementing the above-mentioned sintering mixing tank material level prediction method.

[0006] The third object of the present invention is to provide a sintering mixing tank material level prediction system based on convolutional neural network. The prediction system realizes accurate prediction of the mixing tank material level based on the synergy between the programmable logic controller, the processor and the readable memory, and can further monitor and control the mixing tank material level in real time, without the need for manual control of the mixing tank material level change, thereby realizing intelligent control of the sintering process.

[0007] In order to solve the above technical problems, the present invention provides a method for predicting the material level of a sintering mixing tank based on a convolutional neural network, comprising:

[0008] Step S1: acquiring sintering process data, dividing the data into a training set and a test set, and then screening i groups of characteristic parameters related to the material level of the mixing tank through the training set;

[0009] Step S2: extracting 1 to T groups of data of each characteristic parameter in each group of characteristic parameters according to the time interval, and stacking them into a two-dimensional characteristic parameter matrix of i*T;

[0010] Step S3: Input the obtained two-dimensional feature variable matrix into the convolutional neural network model, build a mixing tank material level prediction model, and test it through the test set.

[0011] As a preferred solution, the screening process of the characteristic parameters includes:

[0012] Step S1-1: After preprocessing and standardization, the sintering process data is randomly divided into a training set and a test set with a ratio of 3 to 4:1;

[0013] Step S1-2: Use the random Senli algorithm to establish more than 18 characteristic parameters including the moisture content of the mixture and the thickness of the material layer, and calculate the importance index n of each characteristic variable for the material level of the mixture tank. k ;

[0014] Step S1-3: After normalizing the importance index of all feature parameters, select the features that are greater than the threshold. As the number of feature parameters increases, the mean square error of the model built using random forest decreases significantly, and when it tends to be stable, the appropriate number of feature parameters can be determined. Sort the importance index of the feature parameters, and the minimum value of the feature parameter importance index can be determined as the threshold.

[0015] As a preferred solution, the importance index n k The calculation process is:

[0016] Formula 1: n k =w k *G k -w left *G left -w right*G right ;

[0017] Formula 2:

[0018] In formula 1 and formula 2, w k 、w left and w right are the ratio of the number of training samples in node k and its left and right child nodes to the total number of training samples, G k , G left and G right are the Gini index of node k and its left and right child nodes, c i is a characteristic variable, v ij is a cutoff value of the variable, N s is the number of all training samples, y left and right Respectively represent the number of training samples of the left and right child nodes after segmentation.

[0019] As a preferred solution, the process of constructing the two-dimensional feature parameter matrix is:

[0020] Step S2-1: Arrange the selected characteristic parameters in chronological order, denoted as X i,1 ,X i,2 ,…,X i,T , where i is the number of the characteristic parameter and T is the time series;

[0021] Step S2-2: Calculate the autocorrelation coefficient of each variable according to the time intervals of 1 to K, and select the time interval when the autocorrelation coefficient is greater than the threshold (the threshold is usually 0.6), which is the time interval of each characteristic parameter;

[0022] Step S2-3: With the characteristic parameter types as vertical columns and the time intervals as horizontal columns, obtain a two-dimensional characteristic parameter matrix of i*T.

[0023] As a preferred solution, the autocorrelation coefficient r i,k The calculation process is:

[0024] Formula 3:

[0025] In formula 3, is the time series mean of variable i.

[0026] As a preferred solution, the structure of the convolutional neural network model consists of a convolution module and a fully connected regression module.

[0027] As a preferred solution, the convolution module includes three convolution-batch normalization-activation function modules. The first two convolution modules use a convolution kernel with a size of 3x3 and a step size of 2 to perform convolution operations, and the last module uses a convolution kernel with a size of 3x3 and a step size of 1 to perform convolution operations.

[0028] As a preferred solution, the fully connected regression module includes a global average pooling layer and a fully connected-Relu activation function-Dropout layer.

[0029] As a preferred solution, the Dropout refers to randomly selecting 10-20% of neurons to inactivate them during each training process.

[0030] As a preferred solution, the number of neurons in the fully connected regression module is: 64-32-1.

[0031] The present invention also provides a computer-readable memory comprising a computer program, wherein the computer program can implement any of the above-mentioned methods for predicting the material level of a sintering mixing tank based on a convolutional neural network.

[0032] The present invention also provides a sintering mixing tank material level prediction system based on a convolutional neural network, comprising: a programmable logic controller (100), a memory (200) as claimed in claim 9, and a processor (300).

[0033] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are:

[0034] 1) The prediction method provided by the present invention screens the characteristic parameters related to the material level of the mixing trough, selects the data obtained by each characteristic parameter with time as the scale, constitutes a two-dimensional characteristic variable matrix, and constructs a mixing trough material level prediction model through a convolutional neural network. Based on the correlation between the sintering characteristic parameters and the material level of the mixing trough, the obtained prediction method not only ensures the prediction result rate of the mixing trough material level, but also greatly reduces the influence of the fluctuation of the input variables on the prediction result, thereby greatly improving the prediction accuracy.

[0035] 2) The technical solution provided by the present invention uses a convolutional neural network to extract features based on time series reconstruction of variables, deeply explores the correlation between process variables and their time series relationships and the material level of the mixing tank, fully utilizes the time series characteristics of process detection parameters, and improves the accuracy of mixing tank level prediction.

[0036] 3) The prediction system provided by the present invention is based on the synergy between the programmable logic controller, the processor and the readable memory, which realizes the accurate prediction of the material level of the mixing trough, and then can monitor and control the material level of the mixing trough in real time, without the need for manual control of the material level change of the mixing trough, thereby realizing intelligent control of the sintering process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the flow of the method for predicting the material level of the sintering mixing tank;

[0038] Figure 2 This is a graph showing the accuracy of the prediction results in the embodiment;

[0039] Figure 3 This is the composition diagram of the sintering mixing tank material level prediction system;

[0040] Figure 4 It is the computer program interface of the sintering mixing tank material level prediction system. DETAILED DESCRIPTION

[0041] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively and meticulously below in conjunction with the accompanying drawings and preferred embodiments of the present invention, but the protection scope of the present invention is not limited to the following specific embodiments.

[0042] The workflow of a method for predicting the material level in a sintering mixing tank is as follows: Figure 1 Shown, including

[0043] Step 1: Take the 9 parameters of the mixed material moisture, comprehensive conveying volume, material level of the mixed material tank, material distribution gate opening, roller feeder speed, material layer thickness, trolley speed, ignition furnace negative pressure, and 1# wind box negative pressure during the sintering process as input variables;

[0044] Step 2: Extract 10-dimensional data with a time interval of 1 minute for each input variable and stack them into a two-dimensional feature variable matrix with 9 rows and 10 columns;

[0045] Step 3: Input the two-dimensional feature variables into the pre-trained convolutional neural network model to calculate the material level of the mixing tank at the next moment and realize the advance prediction of the material level.

[0046] Step 1 specifically includes:

[0047] After preprocessing and standardizing the data collected from the sintering production site, the samples were randomly divided into a training set and a test set with a ratio of 4:1. The random forest algorithm was used to establish a relationship model between 18-dimensional feature variables such as mixture moisture, layer thickness, and the material level of the mixture tank, and the importance index of each dimensional feature variable was calculated. The importance index of each feature was normalized, and the index greater than the threshold T was selected. hThe nine characteristics (mixture moisture, comprehensive conveying volume, mixture trough material level, material distribution gate opening, roller feeder speed, material layer thickness, trolley speed, ignition furnace negative pressure, 1# bellows negative pressure) are used as input variables for mixture trough material level prediction.

[0048] Step 2 specifically includes:

[0049] Extract the time series data of 9 input variables respectively, denoted as X i,1 ,X i,2 ,…,X i,T , where i is the input variable number; T is the time series length, which is 20 minutes; calculate the autocorrelation coefficient of each variable with a time interval of 1 to k respectively; select the autocorrelation coefficient r of all input variables i,k The time interval k when all are greater than the threshold is 10; the sequence data of the 9 input variables with a time length of 10 minutes are combined to obtain a two-dimensional feature vector of 9 rows and 10 columns.

[0050] Step three specifically includes:

[0051] The network structure of the model consists of a convolution module, a global average pooling layer, and a fully connected regression module. The convolution module contains three convolution-batch normalization-activation function modules. The first two convolution modules use a convolution kernel of size 3x3 and a step size of 2 to perform convolution operations, and the last module uses a convolution kernel of size 3x3 and a step size of 1 to perform convolution operations; the fully connected regression module contains a global average pooling layer and a fully connected-Relu activation function-Dropout layer. Dropout means that 20% of neurons are randomly selected to be inactivated during each training process. The number of fully connected neurons is: 64-32-1.

[0052] Example 1

[0053] like Figure 1 As shown, the present invention provides a method for predicting the position of a sintering mixture tank, comprising the following steps:

[0054] 1) According to the method in step 1 above, the random forest algorithm is used to select the 18-dimensional process variables with higher correlation as the input variables for the prediction of the material level of the mixing tank. The characteristic contribution rate of the process variables is shown in Table 1. The threshold T is taken as h =0.03, and the 9-dimensional variables of mixture moisture, comprehensive conveying volume, material level in the mixture tank, material gate opening, roller feeder speed, material layer thickness, trolley speed, ignition furnace negative pressure, and 1# wind box negative pressure were determined as input;

[0055] Table 1

[0056]

[0057]

[0058] 2) According to the method of step 2 above, the time series correlation analysis of the sintering process variables selected in Table 1 is performed, and the autocorrelation coefficients of each variable at different time intervals are shown in Table 2. It can be seen that when k=10, the autocorrelation coefficients of all characteristic variables are greater than 0.6. It can be considered that the variables have strong autocorrelation within 10 sampling times, so the time scale k is selected as 10. The data is spliced ​​into a 9*10 sintering process time series feature matrix, and the sample data set is composed of the mixing tank material level at the corresponding time, which is the input and output structure of the mixing tank material level prediction model;

[0059] Table 2

[0060]

[0061]

[0062] 3) According to the method in step 3 above, the network structure of the sintering mixing tank material level prediction model is as follows: Figure 2 As shown, using production process data, after obtaining the model input through steps one and two, the network structure is trained to obtain the final mixing tank material level prediction model. The prediction results of the model are shown in Table 3. The prediction effect of the model over a period of time is shown in Table 4, where the accuracy is based on the ±8t of the measured value as the judgment standard;

[0063] Table 3

[0064] Serial number 1 2 3 4 …… 82 83 84 Measurements 76.41 77.70 82.50 82.90 …… 97.99 100.18 101.38 Predicted value 76.01 80.47 77.83 80.16 …… 99.64 99.79 102.10

[0065] Table 4

[0066] Root mean square error Average relative error (%) Accuracy (%) 11.24 3.63 96.60

[0067] The present invention also provides a sintering mixture slot prediction system based on convolutional neural network, which is composed of Figure 3 As shown: a programmable logic controller (100), a memory (200), a processor (300), and a computer program stored in the memory (200) and executable on the processor (300). The computer program is based on the Vue+SpringBoot framework and is mainly compiled in the Java programming language. The program interface is as shown in FIG. Figure 4 shown.

Claims

1. A method for predicting the material level of a sintering mixing tank based on a convolutional neural network, characterized in that: include: Step S1: acquiring sintering process data, dividing the data into a training set and a test set, and then screening i groups of characteristic parameters related to the material level of the mixing tank through the training set; Step S2: extracting 1 to T groups of data of each characteristic parameter in each group of characteristic parameters according to the time interval, and stacking them into a two-dimensional characteristic parameter matrix of i*T; Step S3: Input the obtained two-dimensional feature variable matrix into the convolutional neural network model, build a mixing tank material level prediction model, and test it through the test set.

2. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 1, characterized in that: The screening process of the characteristic parameters includes: Step S1-1: After preprocessing and standardization, the sintering process data is randomly divided into a training set and a test set with a ratio of 3 to 4:1; Step S1-2: Use the random forest algorithm to establish more than 18 characteristic parameters including the moisture content of the mixture and the thickness of the material layer, and calculate the importance index n of each characteristic variable for the material level of the mixture tank. k ; Step S1-3: After normalizing the importance index of all feature parameters, select the features that are greater than the threshold value. As the number of feature parameters increases, the mean square error of the model built using random forest decreases significantly, and when it tends to be stable, the appropriate number of feature parameters can be determined. Sort the importance index of the feature parameters, and the minimum value of the feature parameter importance index can be determined as the threshold.

3. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 2, characterized in that: The importance index n k The calculation process is: Formula 1: n k = w k * G k - w left * G left - w right * G right ; Formula 2: In formula 1 and formula 2, w k 、w left and w right are the ratio of the number of training samples in node k and its left and right child nodes to the total number of training samples, G k , G left and G right are the Gini index of node k and its left and right child nodes, respectively, i is a characteristic variable, v ij is a cutoff value of the variable, N s is the number of all training samples, y left and right Respectively represent the number of training samples of the left and right child nodes after segmentation.

4. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 1, characterized in that: The construction process of the two-dimensional feature parameter matrix is: Step S2-1: Arrange the selected characteristic parameters in chronological order, denoted as X i,1 ,X i,2 ,…,X i,T , where i is the number of the characteristic parameter and T is the time series; Step S2-2: Calculate the autocorrelation coefficient of each variable according to the time intervals of 1 to K, and select the time interval when the autocorrelation coefficient is greater than the threshold (the threshold is usually 0.6), which is the time interval of each characteristic parameter; Step S2-3: With the characteristic parameter types as vertical columns and the time intervals as horizontal columns, obtain a two-dimensional characteristic parameter matrix of i*T.

5. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 4, characterized in that: The autocorrelation coefficient r i,k The calculation process is: Formula 3: In formula 3, is the time series mean of variable i.

6. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 1, characterized in that: The structure of the convolutional neural network model consists of a convolution module and a fully connected regression module.

7. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 6, characterized in that: The convolution module includes three convolution-batch normalization-activation function modules. The first two convolution modules use a convolution kernel with a size of 3x3 and a step size of 2 to perform the convolution operation, and the last module uses a convolution kernel with a size of 3x3 and a step size of 1 to perform the convolution operation.

8. The method for predicting the material level of a sintering mixing tank based on a convolutional neural network according to claim 6, characterized in that: The fully connected regression module includes a global average pooling layer and a fully connected-Relu activation function-Dropout layer; the Dropout refers to randomly selecting 10-20% of neurons to inactivate them during each training process; the number of neurons in the fully connected regression module is: 64-32-1.

9. A computer readable memory comprising a computer program, characterized in that: The computer program can implement the sintering mixing tank material level prediction method based on convolutional neural network as described in any one of claims 1 to 8.

10. A sintering mixing tank material level prediction system based on convolutional neural network, characterized in that: include: A programmable logic controller (100), a memory (200) as claimed in claim 9, and a processor (300).

Images