KR970010980B1 - Temperature and components in furnace bath using neural network - Google Patents

Abstract

The ingot steel temperature and component variation predicting method using an artificial neural network comprises the steps of: inputting initial operation condition and factors of continuous sampling data to the artificial neural network to obtain an ingot steel temperature and component value according to the operation time; comparing past continuous sampling data obtained by the same operation condition and factors with the obtained ingot steel temperature and component value in the artificial neural network to execute learning for weighting value within an allowable error range; setting the weighting value when the addition of each error reaches the allowable error range and completing the learning; and inputting an initial operation condition to the learned artificial neural network to predict the ingot steel temperature and component value.

Description

Prediction Method of Molten Steel Temperature and Component Change Using Artificial Neural Network

1 is a functional block diagram illustrating the principles of the present invention.

2 is an overall configuration diagram showing an example of an artificial neural network applied to the present invention.

3 is a flowchart showing an example of neural network application in the present invention.

Figure 4 shows the results of the operation showing changes in carbon concentration over time in the converter process.

5 is an operation record showing the change in molten steel temperature over time in the converter process.

6 is an operation record showing the change of manganese temperature with time in the converter process.

7 is an operation record showing the change in phosphorus temperature over time in the converter process.

8 is a learning curve diagram of the artificial neural network in the present invention.

9 is a comparison of the carbon concentration and the actual carbon concentration change predicted when the artificial neural network in the present invention.

Figure 10 is a comparison of the actual temperature change and the molten steel temperature predicted when applying the artificial neural network in the present invention.

FIG. 11 is a comparison diagram of manganese concentration and actual manganese concentration change predicted when an artificial neural network is applied in the present invention. FIG.

Figure 12 is a comparison of the phosphorus concentration and the actual phosphorus concentration changes predicted when applying the artificial neural network in the present invention.

* Explanation of symbols for main parts of the drawings

1: normalized input layer 2: first intermediate layer

3: second intermediate layer 4: output layer

5: weight between the input layer and the first intermediate layer

6 j nodes in the first middle layer

7 weight between the first and second intermediate layers 8 k nodes in the second intermediate layer

9: weight between the second intermediate layer and the output layer 10: m × n nodes in the output layer

The present invention relates to a method for predicting the temperature and compositional changes of steel, which are essential for producing high quality steel, using artificial neural networks as a technology for converter operation support in steel mills in steel mills.

Converter operation is one of the essential operations in steel mills to remove impurities (carbon, phosphorus, manganese, etc.) in molten iron below an appropriate content through oxidation. Therefore, many studies have been attempted because the process prediction and guidance are very important for achieving efficient operation.

In the related art, there have been two approaches.

The first method is based on the mathematical model that identifies the reaction conditions in the furnace and establishes the resin formula to predict and guide the operation. The second method is the by-product generated by the reaction with oxygen during the conversion of the converter (the period during which the reaction process takes place). There is an exhaust gas analysis method that measures and predicts the amount and composition of gas (carbon monoxide, carbon dioxide).

Although the above conventional methods contributed to the support of converter operation, the converter process is not suitable for applying conventional methods only because the changes of reactive components do not show a regularity and represent a significant change according to each operation condition and situation. There has been a problem, which will be described below.

The characteristics of the converter process is a black box type process, and it is impossible to grasp the situation in the furnace during operation. However, the information in the furnace is only a sub-lance measuring the molten iron temperature past the middle of the process as a single real-time measuring device. The only continuous measurement is a gas analyzer that measures only off-gas components outside the furnace. In addition, the reaction mechanism in the furnace is very complicated during the reaction, and factors that are difficult to clearly explain, such as the change of slag, the lance height, and the oxygen flow rate, have a great influence on the operation. In the conventional method, it is difficult to cope with the operation error caused by these qualitative factors. As shown in FIGS. 4 to 7 comparing the molten steel temperature and the concentration distribution between each component for about 20 operations using the continuous harvesting data of the past, the component cotton varies greatly according to each operation. see. Therefore, in the case of the carbon of FIG. 4 and the molten steel temperature of FIG. 5, the method by the mathematical model should obtain the model coefficient for each operation, and the coefficient value obtained is not applied to the operation but has similar conditions during the next operation. In the case of manganese of FIG. 6 and phosphorus of FIG. 7, the reduction reaction is dominant in the middle of the process, and thus, a very complex chemical reaction form is applied. Therefore, the mathematical model does not apply to this situation. In addition, the temperature and carbon components do not fit well even if the corrected coefficients are applied, so a sub-lance is put in the middle of the process to measure the temperature and carbon content and recalculate until the end of the operation based on the measured values. The biggest problem in using the mathematical model is that the reactive components in the converter are intricately intertwined with each other when they react. The flue gas analysis method using the gas analyzer, which is the second method of the conventional method, estimates the component change with respect to carbon because only the components of gas (carbon monoxide, carbon dioxide) are predicted, but shows a large deviation with respect to temperature, and other components. There is a problem that cannot be considered at all.

Therefore, the present invention does not depend on the sensor, regardless of the interrelated reaction situations inherent in the converter process as described above, the reaction time of phosphorus, manganese, etc. Method of predicting molten steel temperature and component change using artificial neural network that can batch process the component change according to the system and solve the problems that affected the operation situation and various operating conditions due to factors that cannot be identified in the mathematical model It is intended to provide a purpose.

Hereinafter, the present invention will be described in more detail.

The present invention is a method for predicting the molten steel temperature and component changes using the neural network during the converter operation, the input layer consisting of the operating conditions that can be measured at the beginning of the converter operation, the intermediate layer, and the operation A plurality of output terms representing temperature and component values of molten steel over time are inputted to the neural network having an output layer consisting of a matrix structure to input the initial operating conditions and factors of continuous sampling data to obtain the molten steel temperature and the value of each component according to the operating time. And the results of continuous sampling data for the initial operating conditions and the factors inputted to the neural network are compared with the measured values shown in the respective output terms of the output layer of the artificial neural network in the step, and the error is in the tolerance range. The weights are trained in the usual way, and the above steps are repeated so that the sum of the errors of each output term is within the allowable range. Ending the learning after setting the weight when coming in, and inputting the initial operating conditions of the converter operation in the artificial neural network learned as described above, including the step of predicting the molten steel temperature and component changes The present invention relates to a method for predicting molten steel temperature and component change using neural networks.

The present invention stands out as a new control technology in recent years based on continuous sampling data accumulated in the past, not based on a real-time measurement of an open gas analyzer or a sublance like a conventional prediction method. By applying the Neural Network (Neural Network) technique that is receiving the predicted changes in the major reactive components and molten steel (iron) temperature during operation, it will be described in detail with reference to Figures 1 to 3 as follows.

Figure 1 shows a block diagram learned using the artificial neural network for the converter process in the present invention. Key items (preferably 13 items) such as age, charter dose, and blowing pattern, which represent important operating factors and conditions in the converter operation, are input into the artificial neural network (II) during each operation.

In addition, the main items become initial operating conditions and factors of past operating materials (continuous collection data) in the converter process (IV).

In other words, the molten steel temperature and the cotton values (VI) of the components during the desired blowing are obtained as performance values from the under-low continuous sampling data (VI) according to each operation.

The artificial neural network (II) receives the main item data corresponding to the same operation as the operation when obtaining the performance value from the continuous sampling data (IV) and adjusts the weight to adjust the target molten steel temperature and component change. The weights are trained so that the output value (III) thus obtained falls within the tolerance method compared with the actual industrial data (VI). When learning up to the nth operation, when the sum of the errors falls within the tolerance range, the learning is finished. At this time, the new operation is not applied in the training using the artificial neural network. It is generalized by applying to.

FIG. 2 is a structural diagram specifically showing an example of the neural network (II) part of FIG.

The structure applied for the present invention consists of an input layer 1, two intermediate layers 2, 3 and an output layer 4. The input layer 1 is composed of 13 nodes including operating conditions and bias terms to be considered in the initial operation of the converter.In this case, each item has a different unit and a significantly different value (for example, For example, the molten iron dose is 300 tons, the molten steel temperature is 1300 ° C.), and the input terms are calculated by normalizing them to 0 to 1 using the following equations, respectively. Xi = (Data-min (Data)) / (max (Data)-min (Data)), where Xi represents the normalized value of i term, Data refers to the unnormalized actual input value, max (Data), min (Data) represents the maximum and minimum values of Data, respectively. The artificial network technique applied to the present invention is a back-propagation neural network. The principle of this is briefly described with reference to FIG. 2 as follows. The normalized input term (1) is multiplied by a random weight initially set and summed to sum the nodes N ₁ ¹ , N ₂ ¹ ... Shown in Nj ¹ . That is, Nj ¹ =? Xw. The values calculated at each node of the first intermediate layer are input terms of the second intermediate layer 3 through the activation function.

f (N ² ) = 1 / (1+ exp [-∑xw])

This is called the sigmoid function (sigmoid). In the next step, the values at each node of the second intermediate layer 3 are obtained as follows.

N ² = ∑f (N ¹ ) w, (k = 1,2,3,…)

N ² obtained from the second intermediate layer 3 serves as an input term to the output layer through the activation function.

f (Nj ¹ ) = 1 / (1+ exp [-∑f (N ¹ ) w])

By multiplying the above activation function and the weight between the output layer 4, the following values are calculated at each node of the output layer.

N = ∑f (N ² ) w (m = 1,2, …… m, n = 1,2, …… n)

The final output value is represented by the activation function as follows.

Y = 1 / (1+ exp [-∑f (N ¹ ) w])

After calculating all the above output values to the nth operation applied in learning, the squared error (RMS error) of the sum of the error values at each output node in the error (V) term of Fig. 1 compared with the past continuous data collection results. Continue to adjust the weights W ^ (weight on the pressure layer and the second intermediate layer) (5), W ^ (weight between the first and second intermediate layers) (7) and W ^^ (9) in FIG. Way. Among the methods for adjusting these weights, the present invention is applied to a general delta rule (GDR) including a momentu. GDR is one of the general optimization techniques that finds the optimal point by gradient with the derivative of a given function. Also, to find the optimal point more quickly and accurately, we add the derivative of the function called the inertia coefficient to the derivative term of the function. In neural networks, the coefficient that precedes the slope of the derivative of the function is also called the learning rate.

In Fig. 2, reference numeral 6 denotes j nodes in the first intermediate layer, 8 denotes 4 nodes in the second intermediate layer, and 10 denotes m x n nodes in the output layer.

The error control method is based on the following principle.

1. An input vector (X ₁ = X ₁ , X ₂ , X ₃ ,... X) is applied to 13 processing elements of the input layer.

2. Calculate N ¹ , N ² , the output of the coupling function of each processing element in the intermediate layer.

3. Determine the output of each processing element by calculating the transfer function (activation function) of each processing element in the intermediate layer.

4. Calculate N, the coupling function output of each processing element, in the output layer.

5. Determine the output of each processing element by calculating the transfer function of each processing element in the output layer.

6. The error term is determined by calculating the error between the output and the performance value of each processing element of the output layer.

7. Calculate the error term in the middle layers 1 and 2.

8. Adjust the connection strength of the output layer including the momentum and then the connection strength of the intermediate layer (update weights and bias).

9. Repeat the number of learnings so that the output sum error is within the allowable range.

The continuous sampling data applied as the learning data in FIG. 1 was not applied to the operation prediction Goddel at the time of the steelworks converter operation, and the training process value of the steelworks was the black box type as the learning data that was not previously applied in the artificial neural network. This is a unique method that can be used to estimate the change of process over time. In the converter process, the continuous collection data contains a concept that is different from the general meaning. That is, during operation, operators take up to 6 to 7 samples for a process using sub lances and analyze the components according to each time after the operation is completed. Therefore, until now, continuous sampling data could not play a role in predicting and supporting operations in real time, but as a performance data, it was able to grasp changes in operating temperature and changes of each component during processing according to operating conditions. 4 to 7 show distribution charts for each component as a result of more than 20 operations as an example of continuous operation data of past operations for applying as learning data to the present invention, as shown in FIGS. 4 to 7. Likewise, it shows that it is a very difficult form to implement general stereotypes.

4 shows changes in carbon content according to blowing time, FIG. 5 shows changes in molten steel temperature according to blowing time, FIG. 6 shows changes in manganese content according to blowing time, and FIG. 7 shows changes in phosphorus content according to blowing time. Indicates.

Therefore, the learning about the prediction of the operation that shows the process characteristics of the black box type seems to be very different from the learning data of the artificial neural network which is generally applied in real time.

In addition, in the output (III) part of FIG. 1, the terms that were not applicable in the existing converter process were presented and developed in a structure different from the existing artificial neural network.

Conventional converter operation prediction is a complex physical and chemical phenomena of the converter process, and it is somewhat unreasonable because it depends on a mathematical model based on many assumptions and simplifications. Also, the components other than carbon and temperature are very difficult to form, and therefore, the prediction model is not applied. (See FIGS. 4-7). In the present invention, since the initial conditions and operating conditions which are important for learning are already set in the input term and the components according to each time are obtained in the output term as a result of the learning, the conditions and the mathematical model to be considered in each operation should be assumed. There are no problems to do. In addition, the general neural network structure functions as a continuous input of the sensor according to the real time by putting the time term in the input term when applied in real time, but with the process characteristic that continuous measurement of the sensor cannot be performed in real time. Batch Process) In the converter process, the time term is placed on the output layer, which not only makes it possible to estimate components and temperature changes over time, but also to calculate the optimum operating time.

3 shows an example of a flowchart of an artificial neural network applied to the present invention, which will be described in detail as follows.

In the variable declaration and initial value setting 10, local and global variables used in the program are set, and maximum and minimum values for the input and output terms are set for initial values and normalization. In the data input 20, the past performance operation data (continuous collection data) is input, and at this time, an input term is also input during each operation. As shown in FIGS. 4 to 7, the continuous sampling data are results of carbon, temperature, manganese, and phosphorus components according to each process time, and the more the operation learned, the more accurate the estimation is made. The initial weight value setting 30 sets initial values of weights in the first input term and each layer to be calculated using a random function. The output calculation 40 in the intermediate layer calculates the values at each node of the first intermediate layer and the second intermediate layer using input values, weights, and activation functions according to the principles of the neural network described above. In the output value calculation 50 during output, the values of carbon, temperature, manganese, and phosphorus at three time points are output at a total of 15 nodes, and the time point is also calculated by the artificial neural network. The point of view will be wrong, and if there is enough learning data, you will be able to increase the point of view and print it out (that is, you can make a forecast every minute). The judging 60 of the last operation data determines whether the operation to be learned is mode-learned.

The deviation 70 to be output determines whether the sum of the deviations between the output values and the performance values in the trained operation falls within the 0.0002 error tolerance range. If the deviation 70 does not fall within the error tolerance range, the weight adjustment in the output layer ( 80) and adjust the weights connected to the output layer and the second intermediate layer through an optimization technique (GDR). In the same principle, the weight adjustment 90 in the second middle layer is performed by weights connected between the second middle layer and the first middle layer, and the weight adjustment in the first middle layer is 100 by weights connected between the first middle layer and the input layer. Will be adjusted to fall within range. After this iterative learning, if the output sum deviation (70) falls within the tolerance range, the learning ends.As a result, the fixed weight values at each node are output (110) and applied to a new operation that was not applied to the learning. Generalize

In the present invention, the past continuous data, the operating conditions and the desired target values are learned by using an artificial neural network, and after learning, they are applied to a new untrained operation by using a fixed weight value. do. The artificial neural network technique uses the back-propagation network, which is widely known in supervised learning, and the learning method is a generalized delta rule (GDR) using an inertia coefficient. 2 is a structural diagram of an artificial neural network in the present invention. In order to be able to apply to real industry after learning, the input layer node contains the data and initial operating conditions that can be written by the operator in real time, and the output layer node contains the operating temperature, the content of each component (carbon, phosphorus, manganese) and Time is described.

In addition, the structure of the artificial neural network applied to the present invention is also developed and adopted in a new method that does not exist. In other words, in order to determine the isochronous characteristics of the process, a time term must be entered. Since the prediction over time of the existing general neural network is mainly based on a real-time sensor, the time term is put in the input layer to learn. However, in the present invention, the operation time is also predicted by putting the time term in the output layer in consideration of the characteristics of the batch process.

Hereinafter, the present invention will be described in more detail.

EXAMPLE

The performance data of the past eight operations were applied to the two operations that were not learned after the training using the artificial neural network. As shown in FIG. 2, which is a structural diagram of an artificial neural network, a total of 15 output nodes are set by outputting temperature and carbon, manganese, and phosphorus component values at three time points to two output layers of an input layer and an intermediate layer. The input layer consists only of the operating conditions that can be described at the beginning of the operation, which includes the number of converters used, oxygen blowing pattern, main raw material amount (melting amount), molten iron ratio (HMR), total oxygen amount, scrap, secondary raw material amount, initial temperature, coolant amount. Thirteen input terms were set including the initial carbon content, initial manganese content, initial phosphorus content and bias terms. Thirty nodes were set in the first middle layer and 25 nodes were set in the second middle layer. The output layer was trained by setting the temperature, carbon, phosphorus and manganese values collected at three time points during each operation based on the continuous sampling data of the previous eight operations. Therefore, the output layer shows the result value at three time points, but the sampling time of each operation is different, so the sampling time of eight operations will not appear at the same time. Also, the time term was outputted by learning the neural network. The initial weights were randomized, the learning rate was 0.7, and the moment of inertia was 0.5. As shown in FIG. 8, the learning curve, up to 8000 lessons were made so that the total error sum was less than 0.0002.

After learning about the values of eight operations, the results were generalized by calculating the results for the two new operations. The change values of the components to be predicted in one operation are simultaneously output at three time points, and the results for each component are shown in FIGS. 9 to 12.

FIG. 9 shows the carbon content change according to the blowing time, FIG. 10 shows the molten steel temperature according to the blowing time, FIG. 11 shows the change of the manganese content according to the blowing time, and FIG. 12 shows the change of the phosphorus content according to the blowing time.

In FIG. 9, C1-sample represents actual data from the first applied operation and C2-sample represents actual data from the second applied, solid line is a curve drawn based on actual data, and C1-neural and C2-neural are The results are calculated by the neural network for the first and second operations, respectively. In addition, the indications shown in FIGS. 10 to 12 have the same meaning. As shown in Figs. 9 to 12, when the artificial neural network is applied, the characteristics of each operation can be achieved with only one initial input term (which can be easily written by the operators) without the need to correct the model coefficient or use the sublance. According to the experiments, the neural network is automatically processed according to each operation.

As described above, the present invention can predict the carbon content and the temperature change of carbon in the molten iron without having to rely on the sublance during the process, even with continuous data of historical data. Because of the factors that are difficult to identify clearly, the components with large deviations can be predicted, and because of the characteristics of parallel processing, even if several target values can be output at the same time, the operator can easily inform the operating trend. In addition, it is possible not only to predict the operating time as a batch operation, but also to have the effect of sufficiently predicting the values of components (silicon, sulfur, etc.) to be known during the conversion process.

Claims (2)

A method for estimating molten steel temperature and component change using neural networks in converter operation, comprising an input layer, an intermediate layer, and an operation time depending on the operating time which can be measured at the beginning of the converter operation. Calculating the molten steel temperature and the values of the components according to the operation time by inputting the initial operating conditions and the factors of the continuous sampling data into a neural network having a plurality of output terms representing the temperature and the component values of the molten steel in a matrix structure; In addition, the performance values of the continuous operation data for the initial operating conditions and the factors inputted to the neural network are compared with the measurements shown in the respective output terms of the output layer of the artificial neural network in the step, and the weights are adjusted so that the error is within the tolerance range. When learning by the normal method and repeating the above steps, when the sum of the errors of each output term is within the allowable range Step of terminating the learning after setting the weight of the artificial nerve, characterized in that the step of predicting the molten steel temperature and component changes by inputting the initial operating conditions during the converter operation in the artificial neural network learned as described above Method of predicting molten steel temperature and component change using a network. The input port of the input layer of the artificial neural network according to claim 1, wherein the input term of the artificial neural network includes the frequency of converter use, oxygen blowing pattern, main raw material amount (main dose), molten iron ratio (HMR), total oxygen amount, scrap, secondary raw material amount, initial temperature, coolant amount. A method for predicting molten steel temperature and composition change using an artificial neural network, comprising 13 input terms including an initial carbon content, an initial manganese content, an initial phosphorus content and a bias term.