KR100328937B1 - Method for predicting blowing of converter using neural network - Google Patents

Abstract

PURPOSE: A method for predicting blowing of converter using neural network is provided to predict molten steel temperature and carbon concentration after blowing by using vector distance between each data during forming of learning data and using neural network method improving generalization performance of the neural network. CONSTITUTION: The method for predicting blowing of converter using neural network comprises first step of storing operation data into database of the processor computer by a processor computer through sensors of converter related facilities and programmable logic controllers during blowing; second step of generating starting signals of neural network first prediction model, and reading the operation data from the database so that the operation data are divided into learning data and test data using vector distance; third step of comparing learning errors of each of the learning data with a set reference error, continuously learning the data by adjusting weight of the data the learning errors of which are the reference error or more, and learning the neural network model by the neural network first prediction model if learning errors of all data are less than the reference error; fourth step of predicting carbon concentration and molten metal temperature by connecting operation data required for predicting carbon concentration that are measured during generation of the starting signals of the neural network first prediction model and stored in the database to an input layer of the neural network first prediction model learned in the above steps; fifth step of learning neural network second prediction model by generating starting signals of neural network second prediction model, and reading the operation data from the database, thereby repeating the second and third steps; and sixth step of predicting carbon concentration and molten metal temperature by connecting operation data required for predicting carbon concentration that is measured during generation of the starting signals of the neural network second prediction model and stored in the database to an input layer of the neural network second prediction model learned in the above steps.

Description

Converter blow prediction method using neural network.

The present invention relates to the prediction of temperature and carbon concentration during QDT (Quick Direct Tapping) during converter operation. In particular, the neural network improves the generalization performance of the neural network by using the vector distance between each data when constructing the training data for learning the neural network. The present invention relates to a converter blow prediction method using a neural network that predicts the molten steel temperature and carbon concentration after the termination of the blow.

In general, whether or not the molten steel temperature and carbon concentration hit the target value in the converter operation is one of the important operation indicators directly related to the product quality. The molten steel temperature and carbon concentration after the current blow is measured by the sublance started after the blow. However, in order to increase molten steel production, some steel grades omit the sub-lance measurement and perform the QDT to pull out, so in this case, it is not known whether the molten steel temperature and carbon concentration after hitting are hit.

Conventional technical methods for calculating molten steel temperature and carbon concentration after blowing are methods that rely on intuition and empirical knowledge of operators and methods using a mathematical model. In addition, the neural network learning used in the present invention requires not only the training data but also test data for generalization performance evaluation. In the conventional method for constructing such learning and data, the learning data is randomly selected from the collected data, and the rest is composed of test data.

However, the above methods had the following problems, respectively. In the case of intuition and experience of the operator, the result of the calculation may not only vary depending on the operator, but also in the same person, the result may be different due to a mistake or a mistake. The formula model method is based on the regression analysis technique. It is difficult to develop the formula model by various influence factors due to the constraint that the input variables are restricted and the linear relationship of the variable values is assumed to reduce the complexity of the analysis process. The mathematical model itself contains inevitable errors due to the linearization of nonlinear systems.

In addition, since the mathematical model includes many statistical formulas based on the experience of operation, it is highly dependent on the characteristics of the operation and equipment. Therefore, in order to improve the forecasting accuracy, appropriate modification and supplementation is required from time to time according to the change of the operation situation. It takes In addition, when randomly configuring the training data and the test data of the neural network, there are problems in that the distribution is not uniform in view of the nature of the training data.

The present invention has been invented to solve the above problems, and relates to a method for predicting molten steel temperature and carbon concentration after the termination of the training using neural network technique, and also to the training data and test data used in neural network learning. By classifying the vector distance between each data in consideration of the classification, the learning data distribution range is made uniform, and the learning error of each learning data is determined by comparing the reference error and the learning error of the learning data, thereby improving generalization performance. The purpose of the present invention is to provide a method for predicting the blowing of the converter using neural networks.

1 is a flow chart of the steelmaking process.

2 is a flow chart for predicting first and second molten steel temperature and carbon concentration using neural networks.

Figure 3 is the converter structure and neural network input data.

Figure 4a is a first-order prediction neural network model configuration for predicting the molten steel temperature after blowing in the application of the present invention.

Figure 4b is a first predictive neural network model configuration for predicting the carbon concentration after blowing in the application of the present invention.

Figure 4c is a second prediction neural network model configuration for predicting the molten steel temperature after blowing in the application of the present invention.

Figure 4d is a second prediction neural network model configuration for predicting the carbon concentration after blowing in the application of the present invention.

5 is a flowchart for automatically classifying learning and test data in consideration of a vector distance between two data.

6 is a modified sigmoid function to improve generalization performance.

Figure 7a is an efficient learning method using the neural network structure and the reference error.

7B is an efficient learning method using neural network structure and reference error.

The converter blow prediction method using the neural network of the present invention for achieving the above object is a method for predicting carbon concentration and temperature using the neural network,

A method of predicting carbon concentration and temperature using a neural network, the method comprising: a first step of storing, in a database, operation data by a processor computer through sensors and PI of a converter-related facility during drilling;

The second step in which the processor computer generates the neural network first predictive model starting signal at the time of starting sub lances and reads operation data up to the point of sub lance measurement from the database and divides the training data and the test data using the vector distance. ;

The training data and the learning objectives are connected to the input layer and the output layer of the neural network primary prediction model, respectively, and the learning error of each learning data is compared with the set reference error. A third step of training the neural network model so that the neural network first predictive model terminates learning when the learning error of all data is less than the reference error;

Neural network primary prediction model Measured at the start signal and stored in the database, carbon sub iron carbon, iron ore, CBP, HBI, total loading amount, HMR, transport data up to the timing of the blowing sub lance, target carbon In this step, learn the blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, transfer amount up to the timing of the blowing sub lance, old age and target temperature. A fourth step of predicting carbon concentration and molten steel temperature by connecting to the input layer of the first neural network predictive model;

The processor computer generates the neural network 2nd prediction model start signal before the end point sub-branch measurement and reads the operation data up to the point from the database and repeats the above steps 2 and 3 to train the neural network 2nd prediction model. Step 5;

Neural Network Secondary Prediction Model Blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, target carbon, transmission acid amount, calculation difference and temperature prediction Operational data required for the blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, transfer amount, and transmission amount difference (total loss amount-transfer amount up to the time of blowing sub lance), old age, And a sixth step of predicting carbon concentration and molten steel temperature by connecting the target temperature to the input layer of the neural network secondary prediction model learned in the above step.

Hereinafter, with reference to the accompanying drawings will be described in detail the operation and effect of the converter blow prediction method using the present invention neural network.

(1) Starting procedure of steelmaking process flow and neural network model

The overall working situation of the steelmaking process is shown in FIG. The steelmaking process charges main raw materials such as molten iron, scrap metal, and various subsidiary materials into the converter (S1) and blows 99% of pure oxygen through a sub lance (S2) to achieve the molten steel temperature and carbon concentration required for the next process (S3). Generally, the steelmaking process time is about 30 to 45 minutes, and the sub lance measurement is carried out twice, about 80% of the blowing process (S4) and after the start of the next process (S6) before the tapping (S6). To check the molten steel temperature and carbon concentration.

Currently, the converter operation control uses a static equation model (S7) and a dynamic equation model (S8). The static equation model calculates the amount of oxygen and subsidiary materials to meet the target temperature and carbon concentration by using the initial input information, and the dynamic equation model uses the information such as quantitative temperature and carbon concentration through measurement of sub lance during drilling. After the final blow, calculate the amount of acid and coolant to match the molten steel temperature and carbon concentration.

However, at present, there are cases where the end temperature measurement is impossible after the purpose of the installation of the measuring probe and the dropping out of the probe. In such a case, there are problems such as waste of the probe, delay in working time, and inherent risk due to manual measurement. have. The prediction of the molten steel temperature and the carbon concentration of the neural network is performed at the time of measuring the sub lance during the measurement (S9) and before the end point sub lance measurement after the measurement (S10).

First and second molten steel temperature and carbon concentration prediction method using a neural network is shown in FIG. The operation information of the steelmaking process is collected by the PLC (Programmable Logic Controller) (hereinafter referred to as PLC) from the sensor A1 of the converter-related equipment and stored in the computer database A4 using the process progress information in the PLC and the processor computer. Becomes In addition, the processor computer generates the first predictive model starting signal A5 so that the neural network model is calculated at the initial starting point of the blowdown sub-branch using the process progress information.

When the first predictive model start signal is generated, the processor computer preprocesses (A6) various types of operation data read from the field sensor via the PLC into data that can be input to the neural network using average values and standard deviations.

The neural network calculates the molten steel temperature and carbon concentration after the input using the preprocessed input data and outputs the result to the driver screen (A7). The neural network model waits after starting the first predictive model and ending the training.

At the end of the blow (A8), the processor computer determines whether or not the secondary prediction for the molten steel temperature and the carbon concentration, which are process progress information, has been performed (A9), and if the secondary prediction is not made, the secondary prediction model start signal ( A10) is generated.

When the 2nd prediction model start signal is generated, the processor computer preprocesses (A6) the various operation data collected from the field sensor into data that can be input to the neural network, as in the case of starting the 1st prediction model, and uses molten steel using the neural network. The temperature and carbon concentration are calculated and output to the driver's screen (A7).

When the second prediction is executed by starting the second prediction model, the temperature and carbon concentration prediction for the neural network model is finished (A11). In each prediction model, the neural network model calculation results are stored in the shared file system A3 and the database A4 of the processor computer, respectively.

The neural network prediction model of the present invention is trained using the error back propagation learning algorithm, and its structure is constructed based on the multilayer perceptron. The error back propagation learning learning algorithm and the multi-layer perceptron are well known to those skilled in the art having ordinary knowledge in the neural network technology. Therefore, the formulas according to the detailed configuration and complex functional relations are omitted.

(2) Application point of the present invention to the steelmaking process

The present model is applied twice. The application point of the first predictive model is when starting the sub lance during the drilling, and shows the estimated molten steel temperature and carbon concentration after the drilling using the operation information up to the sub lance during the drilling. The point of application of the 2nd prediction model is after the purpose, and all the operation information up to the purpose is used to calculate the prediction result of the molten steel temperature and carbon concentration after the exact blow. Therefore, it is possible to obtain the molten steel information such as the molten steel temperature and carbon concentration after the blow, which is highly reliable when the QDT is implemented or when an operation situation such as a sub lance is dropped.

(3) Input item selection method of the present invention

The neural network applied to the present invention is divided into molten steel temperature and carbon concentration, and a first-order prediction model and a second-order prediction model, and each model is divided by a converter. At present, there are three converters in the steel mill to which the present invention is applied, and since each converter operation situation is different, data is divided by converters.

Figure 3 shows the input data selection method of the neural network, in order to determine the input data of the neural network 63 types of data currently used in the operation was collected and analyzed (S1). The analysis method was first reduced to about 10 types (S2) in accordance with the agreement between the operator and the process control system manager about the degree of process control effect of each operation data.

Secondly, after training the neural network using the operation data adjusted by negotiation, only one arbitrary input data is changed and the other input data is fixed to the average value to observe the change of output according to the change of input and then change. The input data with large degree was extracted and confirmed as neural network learning data (S3) and (S4).

In addition, a new influence factor was created and used by the combination of the operation data with close correlation.The new influence factor was obtained by combining the amount of transport oxygen measured after drilling and the amount of transport up to the time of blowing sub lance during drilling. Used. Figure 4 shows the influence factors related to the prediction of the first and second molten steel temperature and carbon concentration using a neural network as follows.

The eleven input items (1) of the first-order prediction model for calculating the molten steel temperature are as follows.

Blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, delivery amount to the time of blowing sub lance, old age, target temperature.

The 12 input items (3) of the 2nd prediction model for calculating the molten steel temperature are as follows.

Blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, transmission amount, transmission amount difference (transmission amount to the transmission sub-singing point), old age, target temperature.

The eight input items (2) of the first-order prediction model for calculating the molten steel carbon concentration are as follows.

Blowing Sub-Lance Carbon, Iron Ore, CBP, HBI, Total Charge, HMR, Target Carbon, Delivery to the time of Blowing Sub-Lance.

The nine input items (4) of the secondary prediction model for calculating the molten steel carbon concentration are as follows.

Blowing Sub-Lance Carbon, Iron Ore, CBP, HBI, Total Charge, HMR, Target Carbon, Transfer Acid, Transmitted Vehicle Difference.

(4) Configuration and data preprocessing method of the present invention

① Composition of neural network and method of data normalization

The neural network used in the present invention uses an error back propagation learning algorithm, has a multi-layer perceptron structure, and uses a sigmoid function as the activation function of the hidden layer. In order to remove the influence of the size of each input data, it is preferable to normalize the data by using a Zero Mean Unit Variance method.

The normalization method calculates the average value and standard deviation of the collected data items, subtracts the average value from the input data to be normalized, and divides it by the standard deviation. If there is a unit difference in input values between learning data items, the problem of local minimization is caused by one-sided learning by data having a large unit in neural network learning. The above method helps to solve this problem.

② How to separate the collected data into learning and test data

The collected data is divided into learning data used for learning neural network and test data used for performance analysis of neural network developed after completion of learning. In the present invention, the Euclidean distance method is used to divide the operation data into learning and test data.

5 illustrates a process of automatically classifying learning and test data in consideration of the vector distance between the two data. First, the data to be classified are read and a desired vector distance is set (S1). The setting distance is a value between 0 and 1, and the smaller it is, the finer the classification. Any one of the read N operation data is defined as reference data (S2), and each vector distance between this and the remaining (N-1) data is obtained. The reference data Xi and the data vector distance are calculated (S4), and the vector distance calculation formula is shown in Equation 1 below.

When the calculated vector distance is compared with the predetermined distance value (S5), Yi is classified as data S7 when the value is within the set range, and learning data S6 when the value is larger than the set range. When the comparison between the reference data and the (N-1) vector distances is completed by the above method (S8), the vector distance calculation is terminated (S9), and the neural network training and performance evaluation data using the files classified into the training data and the test data. Take advantage of.

By using the above method, it is possible to eliminate the possibility of local learning that may occur when there is a large amount of operation data that is partly identical or similar to the collected data, and to learn neural networks evenly over the entire control range. Therefore, it is a more scientific and efficient data separation method than the conventional method of randomly dividing the training and test data.

(5) Learning and application method of the present invention

In case of learning by error backpropagation algorithm, if input data is applied to neural network, output is obtained based on current weight. The error between this output and the performance value is detected and the weight of each layer is calculated in the direction of reducing this error. This process is repeated until the error falls below the allowable range to learn the nonlinear correlation between input and output data. Learning ends when the error is within tolerance and obtains the final weights, which are stored in the database of the processor computer and used in the first and second application phases in actual operation.

① Variation of activation function to improve generalization performance

In the present invention, the error backpropagation learning algorithm is used and the sigmoid function most commonly used as an activation function is used. The sigmoid function used is shown in Equation 2 below.

In the above function, X represents the sum of the product of the input data and the weight, and H (x) represents the result of the sigmoid function. In the neural network learning, the weight is adjusted to reduce the difference between the desired target value and the output value of the neural network. If the arbitrary weight becomes too large, it is known that the generalization performance of the neural network is degraded.

In order to solve this problem, as shown in FIG. 6, it is preferable to adjust the value of the activation function as follows to have a smaller weight for the same output value. The activation function result of the hidden layer has a value between 0 and 1.

The present invention has been performed by adding an alpha value, that is, a value between 0.01 and 0.1, to the sigmoid function. That is, when the sigmoid function graph is shifted by an alpha value, the hidden layer is modified by the modified sigmoid function. Has a value ranging from alpha to (1 + alpha), and the generalization performance of the neural network model can be improved by preventing the weight value of an arbitrary node from being excessively increased by the alpha value. As a result, comparing the generalization performance before and after the adjustment of the activation function, the improvement was confirmed.

② Efficient learning method using standard error

7 shows an efficient learning method using the developed neural network structure and reference error. The difference between the learning target value and the neural network output value for one input data during neural network learning is called a learning error. The learning error value for each of the N input data is different, and some input data may have an error below the learning termination condition. Conventionally, neural networks have been trained in a direction in which learning errors are reduced for all input data.

In this case, learning is performed even for input data that is poor in learning, that is, a learning error is large, and continuous learning is performed on input data in which the learning completion condition is satisfied, that is, input data in which the learning error is less than or equal to the learning completion condition error. . Therefore, in the present invention, in order to improve the efficiency of neural network learning, a reference error is set and input data having a learning error less than or equal to the reference error is not learned. In this way, the error of the input data having a large learning error can be effectively reduced, and the learning time can be reduced.

As described above, according to the method of the present invention, by accurately calculating the molten steel temperature and carbon concentration after the blow twice, the molten steel information after the blowing can be obtained even during QDT execution. Cost savings include improved error rate and corrosion protection of the furnace. In addition, when the training data is constructed, the distribution of training data is made by considering the vector distance between the data, and the generalization performance of the neural network model is improved when the neural network model is constructed by moving the sigmoid function by 0.01 to 0.1. There is this.

Claims (1)

A method of predicting carbon concentration and temperature using a neural network, the method comprising: a first step of storing, in a database, operation data by a processor computer through sensors and PI of a converter-related facility during drilling; The second step in which the processor computer generates the neural network first predictive model starting signal at the time of starting sub lances and reads operation data up to the point of sub lance measurement from the database and divides the training data and the test data using the vector distance. ; The training data and the learning objectives are connected to the input layer and the output layer of the neural network primary prediction model, respectively, and the learning error of each learning data is compared with the set reference error. A third step of training the neural network model so that the neural network first predictive model terminates learning when the learning error of all data is less than the reference error; Neural network primary prediction model Measured at the start signal and stored in the database, carbon sub iron carbon, iron ore, CBP, HBI, total loading amount, HMR, transport data up to the timing of the blowing sub lance, target carbon In this step, learn the blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, transfer amount up to the timing of the blowing sub lance, old age and target temperature. A fourth step of predicting carbon concentration and molten steel temperature by connecting to the input layer of the first neural network predictive model; The processor computer generates the neural network 2nd prediction model start signal before the end point sub-branch measurement and reads the operation data up to the point from the database and repeats the above steps 2 and 3 to train the neural network 2nd prediction model. Step 5; Neural Network Secondary Prediction Model Blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, target carbon, transmission acid amount, calculation difference and temperature prediction Operational data required for the blowing sub lance temperature, blowing sub lance carbon, iron ore, CBP, HBI, total loading amount, HMR, slag amount, transfer amount, and transmission amount difference (total loss amount-transfer amount up to the time of blowing sub lance), old age, And a sixth step of predicting carbon concentration and molten steel temperature by connecting the target temperature to the input layer of the neural network secondary prediction model learned in the above step.

Images