JPH06200312A - Method for controlling static blowing in steelmaking of converter - Google Patents

Abstract

PURPOSE:To control molten steel temp. and molten steel component concns. at the time of stopping the blowing in high precision by executing the learning of neural network using the actual operational information and actual unknown heat quantity and actual unknown oxygen quantity and executing the learning between units. CONSTITUTION:The neural network is constructed by using the operational information as the input and the unknown heat quantity of the difference between the theoretical input heat quantity and the theoretical output heat quantity and the unknown oxygen quantity of the difference between the theoretical generating oxygen quantity and the theoretical consuming oxygen quantity as the output. A combination factor between the units constituting the neural network is corrected by using the actual operational information, the actual unknown heat quantity and the actual unknown oxygen quantity at the time of operating to execute the learning of the neural network. This learning can precisely be followed up to the variation in the constitution of an error item contained in a model formula and the variation in sequential time, thus the control precision can easily be maintained.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a static blowing control method in converter steelmaking which enables highly precise control of the molten steel temperature and the concentration of molten steel components at the time of blowing during converter steelmaking.

[0002]

2. Description of the Related Art Generally, in static converter steelmaking, static blowing control is carried out in order to bring molten steel components such as molten steel temperature and molten carbon concentration at the time of blowing to a target value. As the static blowing control, (1) static control for determining the blown oxygen amount and auxiliary raw material input amount in the converter blowing based on the heat balance and the mass balance disclosed in Japanese Patent Publication No. 58-58405. In the dynamic control method, in order to improve the accuracy of the mass balance formula, a method of using the amount of oxygen accumulated in slag as a parameter of the multiple regression model formula, and (2) Japanese Patent Laid-Open No. 63-171821, Obtain the temperature distribution in the thickness direction of the refractory lining refractory in the processing furnace immediately before charging the molten iron, and calculate the heat removal amount from the molten iron by the refractory lining refractory during the treatment based on the above temperature distribution. A method is known in which the accuracy of the heat balance equation is increased by incorporating it into a static temperature control element and refining it.

Further, as a method of considering a time series tendency of past results in obtaining a predicted value for an error term (learning term) of a model formula for calculating a material balance and a heat balance during a converter operation, The techniques disclosed in (3) Japanese Patent Application Laid-Open No. 2-190413 and (4) Japanese Patent Application Laid-Open No. 1-230710 are known below.

(3) In the refining control method for a melting and refining furnace that operates in a batch system, the formula for predicting the refining calculation error includes the past operation error, the past operation factor, and the operation factor. A method that uses a time-series model formula to predict operating errors for the time being and reflect them in refining calculations.

(4) Control model equation, y = F (x ₁ ,
x ₂ , ・ ・ ・ x _m ) ＋ Δ a ₀ (y, x ₁ , ・ ・ ・, x _m ：
In the control method using exponential smoothing as a learning procedure for updating Δ a ₀ for each charge based on the blowing result data using a variable, Δ a ₀ : learning term), if the error continues for several charges in the same direction, , If the absolute value of the error is within a predetermined range so that the estimation accuracy of the model formula is improved by increasing the learning term Δ a ₀ update rate and learning stepwise, A converter blowing end point control method for improving the estimation accuracy of the model formula by setting the absolute value of the learning term update amount to a non-zero constant value.

Further, during static calculation of converter blowing, the decarburization reaction behavior in the converter that cannot be accurately predicted is obtained by actually measuring converter exhaust gas information during blowing and using the measured information. Control method based on the idea of maintaining or improving the control accuracy by reconfiguring the static calculation or adjusting the oxygen supply pattern to the furnace during blowing so as to match the premise of the static calculation, The following (5) JP-A-2-185
No. 910, and (6) Japanese Patent Application Laid-Open No. 2-185911.

(5) The heat balance calculation and the oxygen balance calculation are performed with the blow-off C and the blow-off temperature as the target values, respectively, and the static control is performed based on the calculation results. The progress of the refining reaction in the furnace is grasped by integrating the flow rate, the heat balance calculation and the oxygen balance calculation are reconfigured by adding the progress situation data of the refining reaction, and thermal control is performed based on the reconfiguration result. And a blowing control method for a converter for correcting oxygen supply control.

(6) In static control, the change in the ([co] / ([co] + [co ₂ ])) ratio of the converter is measured during the converter operation, and the preset blowing In the furnace ([c
o] / ([co] + [co ₂ ])) Deviations from the ratio change pattern are sequentially detected, and the blowing control of the converter is performed to adjust the oxygen supply pattern into the furnace to minimize the deviation. Method.

[0009]

However, in the method disclosed in (1) Japanese Patent Publication No. 58-58405, the parameters required for the material balance are calculated by statistical processing (multiple regression) of past operation data. I'm looking for
There is a problem that the accuracy is low because the theoretical background is not always sufficient, and the versatility (flexibility) of the model formula is lacking.

In the methods disclosed in (3) Japanese Patent Laid-Open No. 2-190413 and (4) Japanese Patent Laid-Open No. 1-230710, time series data is used as a parameter of a model formula for error prediction, and the learning term update amount is set. Is adjusted according to the tendency of the error, but since the time series data is handled by a statistical method, not the formula or way of thinking that theoretically expresses the temporal change of operating conditions, the construction of the model formula itself is particularly important. There is a problem in that it includes a factor that causes an error because it cannot follow a change such that

In the methods disclosed in (1) Japanese Patent Publication No. 58-58405 and (2) Japanese Patent Laid-Open No. 63-171821, the coefficient of the parameter in the model formula is fixed. Therefore, the model formula becomes obsolete as time passes from the time when the coefficient is determined, and it takes a lot of time and load to try to calculate the coefficient again. There is a problem that it is difficult to maintain the reflected model formula.

Further, the above (2) JP-A-63-171821
The method disclosed in No. 1 has not been theoretically incorporated as a parameter in the heat balance equation used in the conventional static calculation, but the heat absorption amount from the molten iron by the lining refractory during the treatment, By obtaining the temperature distribution in the thickness direction of the refractory lining of the molten iron treatment furnace at the stage immediately before the molten iron charging,
It is excellent in that it can be incorporated as a static control element and the error in heat balance can be reduced. However, in order to obtain the temperature distribution in the thickness direction of the refractory lining, it is necessary to provide a number of temperature sensors in the thickness direction of the refractory and maintain them in a good condition at all times. In the middle to the end of the furnace cost of refractory, there is a large difference in the thickness of the refractory lining depending on the position, so in order to reflect the difference and reduce the calculation error, In addition to this, temperature sensors must be provided at several places, and the calculation of the amount of heat taken is very complicated, which poses a practical problem. Further, in this method, as described above, although consideration is given to the amount of heat removal by the refractory lining, the error terms other than the amount of heat removal are described in (1), (3) and (4) above. The same problems as in the methods disclosed in the respective publications have not been solved yet.

Further, the methods disclosed in the above (5) Japanese Patent Application Laid-Open No. 2-185910 and (6) Japanese Patent Application Laid-Open No. 2-185911 are accurate predictions at the time of static blowing calculation of the heat. The decarburization behavior in the furnace, which cannot be determined, is grasped from the measurement result of the exhaust gas information, and based on the result, the model formula is reconstructed and recalculated, or the oxygen supply pattern is controlled so as to meet the premise of static calculation. Therefore, although the static control accuracy can be improved, the configuration of the error term of the model equation of the mass balance and the heat balance is not corrected. Similarly, necessary changes and changes in time series cannot be followed, and in terms of maintaining control accuracy, the same problem as in the case of the methods disclosed in the publications (1) to (4) is the same. There is.

Further, the techniques disclosed in the above publications have the problems described in detail below. The blowing calculation method based on the regression equation and the learning model equation disclosed in each of the above publications merely outputs the estimated value of the control amount as a continuous amount up to the number of digits that can be calculated.

However, even if an estimated value is obtained as such a continuous amount, in the case of 230 tons converter steel making, for example, the operator sets the oxygen gas injection amount to 50 N, for example.
Change in increments of m ³
We are operating with the idea of selecting options, such as changing it in steps of 0 kg.

The reason why the operator thinks as above is as follows.

(1) In 230 tons bottom-blown converter steelmaking, if the difference in the amount of oxygen blown is 50 Nm ³ , the deviation of the carbon concentration at the time of blow-off is ± 0 of the target value in the case of low-carbon steel. It can be kept within 0.01%, and if the difference in the amount of coolant (iron ore) input is 500 Kg, the temperature of molten steel at the time of blowing is ± the target value.
It can be stored within 5 ℃. That is, the operation with a certain width is allowed within the range of the target molten steel temperature at the end point and the allowable error with respect to the target molten steel temperature.

(2) Since the oxygen gas injection control device and the coolant metering / injecting device originally have a control error and a metering error, the oxygen gas amount actually injected and the coolant injection amount are Since there is an allowable width within the error range, the estimated value up to the number of digits smaller than the error range is not always necessary.

Explaining the reason concretely, for example, the amount of oxygen gas blown at the time of steel making of 230 ton converter is as follows. Is switched to nitrogen gas. At that time, the maximum amount of oxygen gas is 20 Nm ³ in standard deviation (σ) from the start to the end of switching.
Some switching error occurs.

Also, a signal exchange between the process computer, the low-order control microcomputer (DDC), and the control equipment on the converter side causes a delay of up to about 1 second, and this delay is 10 to 20N. Equivalent to the amount of oxygen blown in m ³ .

Regarding the amount of the coolant input, a measurement error of the coolant measurement / input device for the 230 ton converter is ± 50.
Kg, which corresponds to an error of ± 0.75 ° C. at the blow-off temperature.

Further, considering the allowable error in estimating the required oxygen amount and the required coolant amount and the control / measurement error occurring in the equipment of the converter as described above, the unknown heat amount and the unknown oxygen amount are continuously calculated. Even if the number of digits is estimated more than necessary, the accuracy of blowing control is not necessarily improved.

As described in detail above, the method of work by the operator who selects the above-mentioned options has its own grounds, and it is necessary to make a quick decision without lowering the end point (blowing) hit rate. Despite being an efficient method,
Such a work cannot be dealt with by the conventional technique.

The present invention has been made to solve the above-mentioned conventional problems, and can accurately follow changes in the structure of the error term included in the model formula and changes in time series.
A first object of the present invention is to provide a static blowing control method in converter steelmaking that can easily maintain control accuracy.

The present invention solves the above-mentioned first problem, and realizes a blowing calculation method suitable for an actual converter equipment and operation method by a computer by approaching a judgment method by which an operator selects an option. However, it is possible to provide a static blowing control method in converter steelmaking that enables an operator to make a quick judgment while improving the target medium ratio of molten steel carbon concentration and molten steel temperature at the time of blowout. This is the second issue.

[0026]

According to the first aspect of the present invention, the amount of injected oxygen and the amount of coolant input are determined based on the heat balance and the material balance in converter steelmaking, and the molten steel temperature and molten steel component concentration at the time of blowing are determined. In the static blowing control method for BOF steelmaking in which the target value is adjusted to the target value, the operational information in BOF steelmaking is used as input, and the unknown heat quantity, which is the difference between the theoretical heat input and the theoretical heat output, and the theoretical oxygen generation and theoretical consumption. A neural network that outputs the unknown oxygen amount that is the difference from the oxygen amount is constructed, and when estimating the unknown heat amount and the unknown oxygen amount using the neural network, the actual operation information when actually operating and the unknown result The above-mentioned problem is solved by learning the above-mentioned neural network using the amount of heat and the amount of oxygen whose performance is unknown, and correcting the coupling coefficient between the units forming the neural network. Resolve those were.

In the static blowing control method for converter steelmaking according to the first aspect of the present invention, when the neural network is learned, the unknown heat quantity and the unknown oxygen quantity are each set with a predetermined difference width as a pretreatment. If there is a low frequency of performance of a class that has an uncertain amount of heat and an amount of oxygen that is unknown, the operation result data that corresponds to that class is duplicated by copying to increase the number of classes. The learning data frequency distribution equalizing process is performed in advance so that the frequency distribution of the learning data is substantially uniform.

According to the second aspect of the present invention, in the converter steelmaking, the blown oxygen amount and the coolant input amount are determined on the basis of the heat balance and the mass balance, and the molten steel temperature and the molten steel component concentration at the time of blowing are appropriately adjusted to the target values. In the static blowing control method in furnace steelmaking,
A neural network that inputs operation information in converter steelmaking and outputs unknown heat quantity that is the difference between theoretical heat input and theoretical heat output and unknown oxygen quantity that is the difference between theoretical generated oxygen quantity and theoretical consumed oxygen quantity When estimating the unknown heat quantity and the unknown oxygen quantity using the neural network, at least one of the unknown heat quantity and the unknown oxygen quantity is divided into a plurality of classes set with the same difference width, and the unknown heat quantity and the unknown oxygen quantity are unknown. The output layer of the neural network that estimates at least one of the oxygen content is composed of the same number of units as the number of the above classes, and from each unit that constitutes the output layer, the priority selection index that is the criterion for selecting the corresponding class In order to output the number, the above-mentioned neural network is used by using the actual operation information when actually operating and the unknown heat quantity and unknown oxygen quantity. Perform learning over click, by modifying the coupling coefficient between units constituting the neural network is obtained by solving the second problem.

The second aspect of the present invention is also used in the static blowing control method for converter steelmaking, in learning and outputting a neural network having an output layer constituted by the same number of units as the number of classes. Priority selection index number 1 for the output unit of the class corresponding to the unknown heat quantity or the unknown performance oxygen quantity
, The priority selection index is 0 for units of other ranks.
Are input as teacher data, and each output unit outputs the number of priority selection indexes from 0 to 1.

The second aspect of the present invention is, in the static blowing control method for converter steelmaking, in the learning of a neural network having an output layer composed of the same number of units as the number of classes, as a pretreatment. , If the frequency of data for which the amount of heat of unknown performance or the amount of oxygen of unknown performance corresponds to, for example, the frequency of data for which the number of priority selection indicators is 1 is low, the operating performance data is duplicated by copying to increase the data.
In order to make the frequency distribution of the learning data for each class almost uniform, the frequency distribution equalizing process of the learning data is performed in advance.

In the second aspect of the present invention, in the static blowing control method for converter steelmaking, when the output layer is composed of a single unit, the unknown heat quantity or the unknown oxygen quantity is output as a continuous quantity, At the time of learning, the amount of heat whose performance is unknown or the amount of oxygen whose performance is unknown is input to the above unit as a continuous amount.

In the second aspect of the present invention, in the static blowing control method for converter steelmaking, when the output layer is composed of a single unit, the unknown heat quantity or the unknown oxygen quantity is output as a continuous quantity, At the time of learning, the unknown heat quantity and the unknown oxygen quantity are divided into a plurality of classes each set with a predetermined difference width as a pretreatment, and the actual frequency of the class to which each of the unknown quantity of heat and the unknown quantity of oxygen corresponds If the number is small, the operation record data corresponding to that class is duplicated by copying and increased, and the frequency distribution of the learning data is made uniform in advance so that the frequency distribution of the learning data for each class becomes almost uniform. It is the one.

[0033]

In the first aspect of the present invention, unknown terms (unknown heat quantity and unknown oxygen quantity) that cannot be directly measured in the heat balance equation and the mass balance equation in the static blowing calculation are used by using the neural network. Since I tried to estimate
It is possible to estimate the unknown term with high accuracy without being affected by empirical rules and subjectivity. Therefore, by applying the unknown heat quantity and the unknown oxygen quantity to the heat balance formula and the mass balance formula, respectively, it becomes possible to determine the appropriate blown oxygen amount and coolant input amount, and the molten steel component concentration and It is possible to accurately adjust the molten steel temperature to the target value.

Further, for example, since the learning of the neural network is carried out periodically and using the latest performance information, it is possible to reliably maintain the estimation accuracy of the neural network, and it is always possible to accurately estimate the neural network. Since the heat quantity and the unknown oxygen quantity can be estimated, the control accuracy can be easily maintained.

Further, in the first aspect of the invention, in the learning, as a pre-processing, the unknown heat quantity and the unknown oxygen quantity are divided into a plurality of classes each set with a predetermined difference width, and the unknown heat quantity and the unknown result are calculated. When the actual frequency of the class corresponding to each of the oxygen amount is low, the operation result data corresponding to the class is duplicated and increased by copying, so that the frequency distribution of the learning data for each class becomes almost uniform, When the frequency distribution equalization process of the learning data is performed in advance, it is extremely effective in improving the estimation accuracy with respect to the operation pattern belonging to the class of which the record frequency is low.

In the second aspect of the invention, as described above, a neural network is constructed in which the operation information in converter steelmaking is input and the unknown heat quantity and the unknown oxygen quantity are output, and the unknown heat quantity and the unknown heat quantity are unknown using the neural network. When estimating the oxygen amount, at least one of the unknown heat amount and the unknown oxygen amount is divided into a plurality of classes set with the same difference width, and when the output layer is composed of the same number of units as the number of classes. , The self-learning of the neural network is performed based on the latest operation results so that a priority selection index number from 0 to 1, which is a criterion for selecting a corresponding class from each unit, is output. Since the network function has been modified, it is possible to follow changes in operating conditions and prediction targets, etc. Of the target unknown calorific value and uncertain target oxygen amount, and the measurement error and control error of the actual oxygen gas injection control equipment and coolant metering / injection equipment, etc. ) It becomes possible to realize static blowing calculation by a computer, which can improve the hit rate.

When the output layer of either one of the two neural networks is composed of a single unit, the neural network is self-learned based on the latest operation record, and the neural network By modifying the function, it becomes possible to follow changes in operating conditions and prediction targets in the same way, and it is possible to approach the operator's judgment method as described above, and at the same time, the unknown heat amount or unknown oxygen amount can be continuously applied. It is possible to estimate the quantity always with high accuracy.

In the second aspect of the invention, the number of units in the output layer, that is, the number of classes is determined by dividing the unknown heat quantity or the unknown oxygen quantity into equal parts. It is determined by comprehensively judging from the respective tolerances of the molten steel carbon concentration and the molten steel temperature, the control accuracy and the measurement accuracy of each of the oxygen injection control equipment and the coolant metering / injecting equipment, and the empirical knowledge of the operator.

Further, as described above, when the output layer is divided into a plurality of classes and the units corresponding to the number of the classes constitute the output layer, these output layer units have, for example, a priority selection index of 0 to 1. Output the number, and select the median value from the range of classes showing the maximum value (for example, the value closest to 1) in the output number of preferential selection indexes according to a certain rule, for example, the unknown heat quantity and / Or try to estimate the unknown oxygen content.

Further, in the second invention, in learning and outputting to the neural network having the output layer units corresponding to the number of classes, the priority selection index is given to the units of the class to which the uncertain heat amount and / or uncertain oxygen amount corresponds. When inputting the number 1 and the priority selection index number 0 to the units of other classes that do not correspond and each output unit outputs the priority selection index number 0 to 1, a plurality of output layer units It is possible to quantitatively output the corresponding class, ie, the unknown heat quantity or the unknown oxygen quantity, as a guidance value indicating the degree of possibility.

In the second aspect of the invention, in the above learning, as a pre-processing, if the frequency of performance of the class corresponding to the heat quantity of unknown performance or the oxygen content of unknown performance is low, the data of the operation performance is copied by copying. When the frequency distribution of learning data is made uniform in advance so that the frequency distribution of learning data for each class becomes almost uniform by making it and increasing it, in order to improve the estimation accuracy for operation patterns with a low frequency of performance. It is extremely effective.

Further, in the second invention, when the output layer is composed of a single unit, the unknown heat quantity or the unknown oxygen quantity is output as a continuous quantity, and the learning result unknown heat quantity or the unknown result oxygen quantity is continuously outputted during learning. When inputting to the above unit as a quantity, appropriate learning can be performed on the neural network, and each estimated value can be output with high accuracy.

Further, in the second invention, when the output layer is composed of a single unit, even if the unknown heat quantity or the unknown oxygen quantity is output as a continuous quantity, the unknown heat quantity and If the unknown oxygen amount is divided into multiple classes that are each set with a specified difference width, and if the record frequency of the class to which the heat quantity of unknown record and the oxygen amount of unknown record corresponds is low, the operation that corresponds to that class If the frequency distribution of the learning data is preliminarily equalized so that the frequency distribution of the learning data for each class is almost uniform, the actual data is estimated for the operation pattern with a low frequency of performance. It is extremely effective in improving accuracy.

In the present invention, the theoretical heat input is sensible heat of hot metal or the heat of oxidation of carbon, silicon, manganese, phosphorus, etc. (combustion heat), and the theoretical heat output is quicklime, manganese ore, iron ore. , And the sensible heat of molten steel or slag at the time of blowing, as the theoretically generated oxygen amount, the amount of gaseous oxygen used,
The amount of oxygen in iron ore, etc. is calculated as carbon,
The consumption amount of silicon and the like due to combustion can be mentioned respectively.

[0045]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a diagram showing the outline of the configuration of a first neural network for estimating the unknown heat quantity applicable to the first embodiment of the first invention, and FIG. 2 similarly estimating the unknown oxygen quantity. It is a diagram showing an outline of the main part composition of the 2nd neural network which does. These neural networks are composed of three layers of an input layer, an intermediate layer and an output layer.

FIG. 3 is an explanatory view showing the outline of the apparatus configuration applied to this embodiment. Reference numeral 10 in FIG. 3 is a converter, and process information of the converter 10 is input to the process computer 12. The process computer 12 has first and second neural networks, which will be described in detail below, built therein.
By inputting the process information input online in 2 into both neural networks, the target unknown heat quantity and unknown oxygen quantity are respectively output.

The process computer 12 is connected to a learning server 14, and the learning server 14 collects process information from the process computer 12 and uses the information to perform self-learning of both neural networks. The correction information is calculated by performing the correction, and the weight coefficient of each neural network in the process computer 12 is corrected based on the correction information.

Since both neural networks have substantially the same configuration, the first neural network will be mainly described. In the first neural network, the input layer has n input units Ii (i = 1
.About.n), the intermediate layer has m intermediate units Hj (j = 1 to 1).
m), the output layers are each composed of one output unit O _A.

In this neural network, each of the n input units Ii has m intermediate units Hj.
Different weighting factors Wij (i = 1 to n, j
= 1 to m). That is, the input unit I1
, The intermediate units H1 to Hm are respectively combined with weighting factors W11 to W1m, and although not shown, the input units I2 to In-1 are similarly combined, and the last input unit In is weighted. They are connected by the coefficients Wn1 to Wnm.

The m intermediate units Hj forming the intermediate layer are the output units O forming the output layers.
_A and the weight coefficient W _jA (j = 1 to m) are combined.

In the above neural network,
Signals corresponding to n input units Ii
t _i (i = 1 to n) is input. This signal t _i is input as process information that is input online from the converter 10 to the process computer 12.
In this way is a signal input to each input unit Ii, for example, from the unit I1, the weighting factor is multiplied by the input signal t ₁ to the intermediate unit H1 ~Hm W11 · t ₁ ~
W1m · t ₁ is input respectively, and similarly, the intermediate units H1 to H are also input from the input units I2 to In.
A signal obtained by multiplying each of the input signals t ₂ to t _n by a weighting coefficient is input to Hm. Therefore, the signal u _j represented by the following formula (1) for obtaining the sum of i = 1 to n is input to each of the intermediate units Hj.

U _j = ΣW ij · t _i (1)

As described above, the signal u is sent to each intermediate unit Hj.
_jIs input, the intermediate unit Hj becomes an output unit.
O_ASignal to u_jA(J = 1 to m) is output. this
Signal u_jAIs the input signal as shown in the following equation (2).
u_jThe function f (u _j) To the weighting factor W_jAMultiply by
Given as a value. Also, the above function f (u_j) As
Is, for example, using the sigmoid function of the following equation (3)
You can

U _jA = W _jA · f (u _j ) (2) f (u _j ) = 1 / (1 + exp ( _{−u j} )) (3)

As described above, when u _{1A to} u _mA are input to the output unit O from each of the intermediate units H 1 to Hm, the output unit O _A gives the final output as an unknown calorific value:
And outputs the (Y _A). The output (Y _A) is given by the following equation (4). In this equation, the sigmoid function can be used as the function f.

[0057] _{(Y A) = f (u} 1A + u 2A + u 3A + ··· + u mA) ... (4)

The second neural network shown in FIG. 2 does not show the input layer, but the intermediate layers H1 to Hm.
An output unit O _B, the respectively coupled by the weighting factor W _jB (j = 1~m), as the final output, except that unknown amount of oxygen (Y _B) is to be outputted, substantially the It is the same as the first neural network. However, the numbers of units n and m can be set to numbers different from those of the first neural network.

[0059] Therefore, in the second neural network, each signal inputted to the output unit O _B is the (2)
The final output (Y _B ) can be represented by the following equation (5) corresponding to the equation and the following equation (6) corresponding to the above equation (4).

U _jB = W _jB · f (u _j ) (5) (Y _B ) = f (u _1B + u _2B + u _3B + ... + u _mB ) ... (6)

In this embodiment, in the first neural network, a signal as process information for each of the input units I1 to In, t ₁ = carbon concentration in molten iron, t ₂ = silicon concentration in molten iron, t ₃ = quicklime dosages, t ₄ = manganese ore charging amount, by inputting respectively the like, unknown amount of heat as an estimate from the output unit O _a (Y _a) is output.

Further, in the second neural network, a signal as process information, t ₁ ′ = amount of gaseous oxygen, t ₂ ′, for each of the input units I1 to In.
= Iron ore charging amount, t ₃ '= the molten iron in the carbon concentration, t _4' = molten iron in silicon concentration, by this inputting respectively the like, from the output unit O _B of the second neural network, unknown amount of oxygen as an estimate (Y _B ) is output.

Further, as described above, for both the neural networks, the actual values when actually operating, that is, the actually measured t ₁ , t ₂ , t ₃ , t _4, etc., and when the blow is actually stopped Learning is performed to correct the weighting coefficient of the coupling between the units using the uncertain calorific value and uncertain oxygen amount obtained by measurement. When learning, the initial value of the weighting coefficient between units can be set using a random number table.

As this learning method, a back propagation (error back propagation) method can be used. This back-propagation method configures the neural network so as to reduce the square value of the error δ between the unknown heat quantity and the unknown oxygen quantity estimated by the neural network and the unknown heat quantity and the unknown oxygen quantity based on the actual measurement. It changes the weighting factor between units, and is achieved by repeating forward calculation and backward error backpropagation calculation.

As described above, by repeating the learning for the first and second neural networks and periodically executing the learning, for example, even when the operating conditions change, correct estimation is always performed by the both neural networks. It can be carried out.

According to the present embodiment, as described above, the operation is repeated, and the first operation is performed based on the latest operation result, for example.
Since the second neural network and the second neural network are self-learned, the unknown heat quantity in the heat balance equation and the unknown oxygen quantity in the mass balance equation can always be estimated with high accuracy.

Therefore, by performing the static blowing control using the unknown heat quantity and the unknown oxygen quantity estimated by the above-described method, the molten steel component concentration and the molten steel temperature at the time of blowing stop can be adjusted to the target values with high accuracy. It becomes possible.

In the learning of the neural network, it is effective to perform the following frequency distribution equalization processing of learning data as a preprocessing for improving the estimation accuracy.

First, the unknown heat quantity and the unknown oxygen quantity are divided into a plurality of classes each set with a predetermined difference width, and the record frequency distribution is obtained for each of the classes to which the record unknown heat quantity and the record unknown oxygen quantity correspond.

Next, the frequency of the class with the highest actual frequency, F _A (MAX) for the unknown calorific value, F _B (MAX) for the unknown oxygen amount, and the frequency with the smaller actual frequency than that, When the unknown calorific value is F _A (X) and the unknown oxygen amount is F _B (X), the data of the operation results of each of the classes with a small actual frequency is F _A (MAX) / F _A (X), F for unknown oxygen content
_{B (MAX) / F B (} X) such that the frequency doubled, increasing by making a duplicate by copying.

At this time, the above F _A (MAX) / F _A (X), F _{B
(MAX) / F B (X} ) is if not an integer, truncation, rounding, by any of a predetermined constant integer rules of rounding up shall be integer.

By processing in this way, the number of learning data belonging to each class can be made substantially uniform. Therefore, it is possible to effectively improve the estimation accuracy with respect to the operation pattern belonging to the class with a low record frequency.

Next, a specific example to which this embodiment is applied will be described.

As the converter 10, a 230-ton upper-bottom blowing converter is used, and 230 tons of hot metal is put into the converter 10 to blow low carbon steel (carbon concentration 0.06% or less). The following is a description of the results of inputting the operation information of the above into the neural network and estimating the unknown heat quantity and the unknown oxygen quantity, respectively. The unknown amount of heat and the unknown amount of oxygen are generated by oxidation of iron and the like.

The operation information actually used as the input is the weight of the main contents of the furnace interior, such as hot metal, quick lime and iron ore, and their component values (contents of carbon, manganese, phosphorus, etc.). ), The amount of gas blown into the furnace (oxygen, propane, nitrogen, argon, etc.), the hot metal charging temperature, the molten steel temperature at the time of blowing, the slag basicity at the time of blowing, etc.

The estimation accuracy of the unknown heat quantity and the unknown oxygen quantity when this embodiment is applied are shown in FIGS. 4 and 5, and are shown in histograms in FIGS. 6 and 7.

The black circles in FIGS. 4 and 5 and the blank bar graphs in FIGS. 6 and 7 indicate the case where the frequency distribution equalization processing of the learning data is not performed in advance during learning (the method of the present invention (1-1)). The results are shown in FIG. 4 and FIG. 5 and the shaded bar graphs in FIG. 6 and FIG. 7 indicate the case where the frequency distribution equalization processing of the learning data is performed in advance (the method of the present invention (1- 2)).

In FIG. 4, the abscissa indicates the estimated unknown calorific value (10 ⁶ Kcal) according to this embodiment, and the ordinate indicates the uncertain calorific value (10).
⁶ Kcal). Further, in FIG. 5, the horizontal axis represents the estimated unknown oxygen amount (N m ³ ) according to the present embodiment, and the vertical axis represents the uncertain oxygen amount (N m ³ ).

In FIG. 6, the horizontal axis represents the difference between the amount of heat of unknown performance and the amount of heat of unknown heat estimation, and the vertical axis represents the frequency. Further, in FIG. 7, the horizontal axis represents the difference between the unknown amount of oxygen and the unknown amount of estimated oxygen, and the vertical axis represents the frequency.

In order to clarify the effect of this example, the results of estimating the unknown heat quantity and the unknown oxygen quantity by the conventional method for the same converter operation are also shown in FIG. 4 and FIG. The histograms are shown in FIGS. 8 and 9 corresponding to FIGS. 6 and 7. The unknown calorific value by this conventional method is estimated by the multiple regression equation shown in the following equation (7),
The unknown oxygen amount is estimated by using the multiple regression model formula of the following formula (8).

Unknown heat quantity = a ₀ + a ₁ t ₁ + a ₂ t ₂ + a ₃ t ₃ + ... a _n t _n (7)

Here, t ₁ to t _n are the same operation information as the input of the first neural network, and a ₀ to a _n are regression coefficients.

Unknown oxygen amount = b ₀ + b ₁ t ₁ ′ + b ₂ t ₂ ′ + b ₃ t ₃ ′ + ... b _n t _n ′ (8)

Here, t ₁ ′ to t _n ′ are the same operation information as the input of the second neural network, and b ₀ to b
_n is a regression coefficient. However, t _n and t _n ′ may be different.

As is clear from FIGS. 4 and 5 or from comparison between FIGS. 6 and 7 and FIGS. 8 and 9, the estimation accuracy of the present embodiment is much higher than that of the conventional method. .
Furthermore, the method of the present invention (1-
The method of 2) is the method of the present invention (1-1) in which the same treatment was not performed.
It can be seen that the estimation error is slightly smaller than that of.
In addition, the amount of heat with unknown results in FIG. 4 and the amount of oxygen with unknown results in FIG.
Blowing performance data is applied to the first and second neural networks and the multiple regression model equations shown in the above equations (7) and (8), and static blowing calculation is back-calculated and calculated. .

Next, before the start of blowing, the unknown heat quantity and the unknown oxygen quantity are estimated using the neural network of the present embodiment, and static blowing calculation is performed using the estimated values to calculate the required blowing oxygen quantity, FIG. 10 shows the result of investigating the simultaneous target appropriateness ratios of the molten steel carbon concentration and the molten steel temperature at the time of the blowing stop, while the required input coolant amount is obtained, and the calculation result is used for blowing.
Shown in. For comparison, the simultaneous target predictive value when performing blowing by the same method except that the unknown heat quantity and the unknown oxygen quantity are estimated using the multiple regression equations (7) and (8) above. Was also investigated, and the results are also shown in FIG. Here, the simultaneous target suitability is that the actual blown carbon concentration is within the range of the target blown carbon concentration ± 0.01% in the same heat, and the actual blown molten steel temperature is the target blown molten steel temperature ± 6. It is shown as a percentage of the total number of heats entering the range of ° C.

The static blowing calculation is conceptually performed as follows. That is, a heat balance formula (9) and an oxygen balance formula (10) in converter blowing are established.

Qin (heat input amount) -Qout (heat output amount) = ΔQ (unknown heat amount) (9) Vin (oxygen amount in input substance) -Vout (oxygen amount in released substance) = ΔV (unknown oxygen amount) (( 10)

Since these two equations (9) and (10) are binary simultaneous equations of the amount of blowing oxygen and the amount of injected coolant, solving these equations gives guidance values for static blowing calculation ( The required blowing oxygen amount and required input coolant amount) are obtained.

As is clear from FIG. 10, the conventional multiple regression method had a simultaneous predictive value of about 60%, but according to the methods (1-1) and (1-2) of the present invention, It can be seen that it has improved to 90%.

As described above in detail, according to the present embodiment, by using the neural network, the factors (unknown heat quantity, unknown oxygen quantity) which are known to be parameters from the conventional knowledge are effectively modeled. As a result, it has become possible to improve the accuracy of static blowing control. Further, since it is possible to always learn the above-mentioned neural network using the latest data, it is possible to maintain the estimation accuracy by the neural network. Therefore, always perform static blowing control with high accuracy. Is possible.

Next, a second embodiment according to the second invention will be described. The second embodiment uses the first and second neural networks used in the first embodiment as well as the third and fourth neural networks described in detail below.

FIG. 11 is a diagram showing the outline of the configuration of a third neural network for estimating the unknown heat quantity as in the first neural network, and FIG. 12 is the same as the second neural network in the unknown oxygen quantity. It is a diagram which shows the outline of the principal part structure of the 4th neural network which estimates.

As in the first embodiment, the converter control apparatus shown in FIG. 3 is applied to this embodiment, and process information of the converter 10 is input to the process computer 12.

The process computer 12 includes the first to fourth
3rd, 4th, 2nd of all neural networks
And the third or the first and fourth neural networks are combined to be used. And
In this process computer 12, by inputting the process information input online for each neural network of each of the above combinations, the target unknown heat quantity and unknown oxygen quantity are continuous quantities from the first and second neural networks, respectively ( As an analog value), the third and fourth
From the neural network, it is output as the number of priority selection indexes (1 to 0).

Since the first and second neural networks applied to this embodiment have already been described in detail in the first embodiment, the third and fourth neural networks will be described in detail below.

Each of the third and fourth neural networks is composed of three layers of an input layer, an intermediate layer and an output layer, and each layer is composed of a plurality of units.

The number of output layers of the third neural network is determined by equally dividing the amount of unknown heat by a predetermined width, and the number of output layers of the fourth neural network similarly determines the amount of unknown oxygen by a predetermined width. It is decided by dividing.

In the third and fourth neural networks, the predetermined widths (class difference widths) that define the number of output layers are all allowable errors of carbon concentration and molten steel temperature at the time of blowing, and oxygen gas. It can be decided by comprehensively judging from the control error and the measurement error of each of the injection control facility and the coolant metering / injection facility, and the empirical knowledge of the operator.
For example, in a 230 ton bottom blowing converter,
For example, the width may be 1 × 10 ⁶ Kcal, and the unknown oxygen amount may be, for example, 50 N m ³ . The output units of the third and fourth neural networks output the priority selection index numbers between 0 and 1.

The third and fourth neural networks will be described in more detail. Since the configurations of both are substantially the same, the third neural network will be mainly described.

In the third neural network, as shown in FIG. 11, the input layer has n input units Ii.
(I = 1 to n), the intermediate layer has m intermediate units Hj
(J = 1 to m), the output layer has k output units O _x (x
= 1 to k).

In this third neural network, like the first and second neural networks, each of the n input units Ii has m intermediate units Hj.
Different weighting factors Wij (i = 1 to n, j
= 1 to m), but the above m constituting the intermediate layer
Each of the intermediate units Hj constitutes k in the output layer.
Output units O _x (x = 1 to k) and weighting factor W
_They are connected by _jx (j = 1 to m, x = 1 to k).

Therefore, in the third neural network, the signal processing up to the intermediate layer when the corresponding signals t _i (i = 1 to n) are input to the n input units Ii is This is executed in the same way as in the case of the first or second neural network, and the signal u _j represented by the above equation (6) is input to each of the intermediate units H _j .

As described above, the signal u is sent to each intermediate unit Hj.
_{When j} is input, the intermediate unit Hj outputs signals u _jx (j = 1 to 1) for k output units O _x (x = 1 to k).
m, x = 1 to k) are output. This signal u
_jA is the above-mentioned input signal u _{j as} shown in the following equation (11).
_Is given as a value _obtained by multiplying the function f (u _j ) incorporating Also, as the above function f (u _j ),
For example, it can be given by the sigmoid function of the following expression (12).

U _jx = W _jx · f (u _j ) (11) f (u _j ) = 1 / (1 + exp ( _{-u j} )) (12)

As described above, k output units O _x (x
= 1 to k) from each intermediate unit H1 to Hm
_{When 1x} to u _mx are input, the output unit O _x is
The priority selection index number αx of 0 to 1 indicating that the calorific value in the range specified by x is suitable as the guidance value of the unknown calorific value used in the static blowing calculation is output. That is, the closer the priority selection index number αx is to 1, the more suitable the class is as a guidance value, and conversely, the closer it is to 0, the less suitable the class is. Finally, a class in which α x is closest to 1 is selected, and a representative value (for example, the median value) of the class is output to obtain a guidance value for the unknown calorific value.

In the fourth neural network shown in FIG. 12, the input layers are not shown, but the intermediate layers H1 to H1.
m l (lower case L) output units O _y (y = 1
To l) is the weighting factor _{W jy (j = 1~m, y} = 1~l) at respectively coupled, said output unit O _y is the oxygen content in the range of class y defines static blowing calculated The possibility of conforming to the guidance value of the unknown oxygen amount used for
It is substantially the same as the third neural network except that it outputs the priority selection index number βx of 0 to 1.

Therefore, in the fourth neural network, each signal input to the l output units O _y (y = 1 to l) can be expressed by the following equation (13) corresponding to the above equation (11). it can.

U _jy = W _jy · f (u _j ) (13)

In the present embodiment, as in the case of the first neural network described in the first embodiment, for the third neural network, a signal as process information for each of the input units I1 to In, t ₁ = carbon concentration in molten iron, t ₂ = silicon concentration in molten iron, t
Enter ₃ = quicklime input, t ₄ = manganese ore input, etc., respectively. As a result, from the k output units O _{x of} the third neural network, the priority selection index numbers α1 to αk corresponding to the unknown heat quantity class are output as values in the range of 1 to 0 and are used as guidance values.

[0111] Further, a fourth neural network, for each of the input unit I1 -In, as in the case of the second neural network, a signal as a process information, t ₁ '= gaseous oxygen, t _2' = Input iron ore input, t ₃ ′ = carbon concentration in molten iron, t ₄ ′ = silicon concentration in molten iron, etc., respectively. As a result, from the l output units O _{y of} the fourth neural network, the priority selection index numbers β 1 to β l corresponding to the class of the unknown oxygen amount are also output as the values 1 to 0, which are used as guidance values.

Further, as described above, with respect to the third neural network, the actual value when actually operating,
That is, by using the process information such as the carbon concentration in molten iron t ₁ , the concentration of silicon in molten iron t ₂ , the amount of quicklime input t ₃ , the amount of manganese ore input t ₄ , etc., and the calorific value of which the results are unknown, and to the fourth neural network. Similarly, the amount of gaseous oxygen
Using the process information such as t ₁ ′, iron ore input t ₂ ′, carbon concentration in molten iron t ₃ ′, silicon concentration in molten iron t ₄ ′, and the amount of oxygen whose performance is unknown, the weight of the bond between each unit Learn to modify the coefficient.

At that time, the amount of heat whose performance is unknown and the amount of oxygen whose performance is unknown
Output unit O corresponding to each _x(X = 1 to k),
O_yOne output unit in each (y = 1 to l)
1 for all other output units not applicable
Is input as teacher data.

Further, as pre-processing, the operation record data of the less frequent record class corresponding to the record heat quantity and record unknown oxygen amount are duplicated by copying to increase the frequency distribution of the learning data for each class. The learning data frequency distribution equalization process is performed so that As a specific method of this homogenization treatment, the method described in the first embodiment can be used. For the learning, the same backpropagation (error backpropagation) method as described in the first embodiment can be used, and in this case, the initial value of the weighting coefficient between units is a random number table. It can be set using.

In this embodiment, the estimation of the unknown heat quantity and the unknown oxygen quantity and the learning of the neural network based on the past operation record data are performed by any of the following combinations.

(A) Third and fourth neural networks (FIGS. 11 and 12) (B) Third and second neural networks (FIGS. 11 and 2) (C) First and fourth neural networks (Figs. 1 and 12)

In the combination of (A) above, the unknown heat quantity is k
K output units O _x (x
= 1 to k) as a priority selection index number of 0 or 1, and the estimated unknown calorific value is output from the output unit O _x as a priority selection index number of 0 to 1. Also, for the fourth neural network, at the time of learning, the unknown oxygen amount is divided into l output units O _y.
(Y = 1 to 1) is input as a priority selection index number of 0 or 1, and the estimated unknown oxygen amount is output as a priority selection index number of 0 to 1 from the output unit O _y .

In the combination of (B) above, the third neural network for estimating the unknown heat quantity is (A) above.
In the second neural network, the estimated unknown oxygen amount is output from one output unit O _B , for example, by the equation (6), as in the case of the first embodiment. It is output as a sigmoid function as shown, and the unknown oxygen amount is input as an analog value as it is.

For the output at the time of estimation and the input at the time of learning, the conversion index (1 to 0) of the actual value used at the initial learning of the neural network is used to convert it into an analog value of the unit actually used. Then, the estimated value is output, and at the time of learning, the actual value is input as the converted value of the analog value in the unit actually used, using the conversion index (1 to 0) used for the initial learning.

In the combination of (C), the fourth neural network for estimating the unknown oxygen amount is output or input in the same manner as in the combination of (A), but in the first neural network, the As in the case of the first embodiment, the estimated unknown calorific value is one output unit O _A.
Is output as, for example, a sigmoid function as shown in the equation (4), and the uncertain amount of heat is directly input as an analog value.

In outputting and inputting the unknown calorific value in this case, the conversion index (1 to 1) of the actual value used in the initial learning is used.
The conversion with the analog value of the unit actually used using 0) is the same as in the case of the unknown oxygen amount of the combination of (B).

As described above, by repeating the learning with respect to the neural network and performing the learning,
Even when the operating conditions change, the neural network can always make a correct estimation.

According to the present embodiment, as described above, the operation is repeated and, for example, based on the latest operation result, for example, for each of the combinations of the neural networks of (A), (B) and (C) described above. Since self-learning is performed, the unknown heat quantity and the unknown oxygen quantity at the time of static blowing calculation can always be estimated with high accuracy.

In the construction and learning of the third and fourth neural networks, as described above, the output layer has carbon concentration and molten steel temperature tolerance at the time of blowing stop, oxygen gas blowing control equipment and coolant metering. -By comprehensively judging from the control error and metering error of the input equipment, the empirical knowledge of the operator, etc., the unknown heat quantity and / or the unknown oxygen quantity are divided into a plurality of classes each having the same width, and the number of these classes Since the output layer is configured with the same number of output units and one priority selection index number consisting of, for example, a value of 0 to 1 is output to each output unit, the model formula fixed as in the conventional case is used. As compared with the method of outputting the estimated value as a continuous analog value, the estimation accuracy of the unknown heat amount and the unknown oxygen amount can be improved.

Further, according to the present embodiment, as compared with the method of the first embodiment, which corrects the coupling coefficient between the units forming the neural network by self-learning and outputs the estimated value as a continuous analog value. , It becomes possible to realize a blowing calculation method that is much closer to a judgment method as if an operator selects an option while maintaining the target end point carbon concentration and / or molten steel temperature target ratio by a computer. .

Next, a specific example to which this embodiment is applied will be described.

As the converter 10, a 230 ton top-bottom blowing converter was used, and 230 tons of hot metal was put into the converter 10 to blow low carbon steel (carbon concentration 0.06% or less). The operational information of was input to the neural network to estimate the unknown heat quantity and the unknown oxygen quantity.

Here, the combinations (A) to (C) were used as the neural network combinations. As the output layer of the third neural network, the class width of the unknown calorific value is 1 × 10 ⁶ Kcal for the reason described above.
Then, as shown in parentheses in FIG. 11, 1 × 10 ⁶ ~
The output layer was composed of units corresponding to 8 classes up to 9 × 10 ⁶ Kcal. For the same reason, the output layer of the fourth neural network has a class width of unknown oxygen amount of 5
0 N m ^3, and as shown in parentheses in FIG.
The output layer was composed of units corresponding to 16 classes from 00 to +400 Nm ³ .

The blowing data of the low carbon steel was prepared for 500 charges, and the first to fourth neural networks were constructed by initial learning. At that time, 1 was input to the output unit of the class corresponding to the actual data of the unknown heat quantity or the unknown oxygen amount of the third and fourth neural networks, and 0 was input to the output unit of the other classes not corresponding. .
Regarding the first to fourth neural networks, if the number of data of the corresponding class is small, the number of data of the class is increased by making a copy so that the number of data of all classes becomes almost uniform. The learning data was homogenized by the same method as described in the example.

After the above pretreatment, the number of units in the intermediate layer was adjusted, the results were verified using the operation data, and the neural network having the highest estimation accuracy was selected to optimize the configuration. .

After the structure of the neural network was determined, the latest operation data of 500 charges was input and re-learning was performed, and the neural network was used in the actual operation thereafter.

The estimation accuracy when estimated using the neural network configured as described above is shown in FIGS. 13 and 14, respectively.
It was shown to.

In FIG. 13, the abscissa indicates the difference between the heat quantity of which the result is unknown and the heat quantity of which the estimation is unknown estimated by the present embodiment (× 10 ⁶ Kcal).
), And the vertical axis is frequency. In addition, in FIG. 14, the horizontal axis represents the difference (N m ³ ) between the unknown amount of oxygen and the unknown amount of oxygen estimated in this example, and the vertical axis also represents the frequency. Further, in FIGS. 13 and 14, the histograms in the shaded areas surrounded by broken lines show the estimation accuracy when the third and fourth neural networks are used, and the histograms in the blank areas surrounded by solid lines are the first histograms, respectively. And the second
When the learning data is homogenized as the preprocessing by using the neural network of (the method of the present invention (1-
The estimation accuracy of 2)) is shown.

As is clear from FIG. 13, there is almost no difference in the estimation accuracy between the case where the first neural network is used and the case where the third neural network is used for estimating the unknown heat quantity, Further, it can be seen from FIG. 14 that there is almost no difference in estimation accuracy between the case where the second neural network is used and the case where the fourth neural network is used for estimating the unknown oxygen amount.

The estimation accuracy of the unknown heat quantity and the unknown oxygen quantity is as follows. The unknown quantity of heat or the unknown quantity of oxygen obtained by inputting the above process information into the neural network in each heat of the past operation and estimating and calculating. The evaluation was made by the difference (error) between the amount and the estimated unknown calorific value or the estimated unknown oxygen amount according to the present example. When the third and fourth neural networks were used, the median value of the class whose priority selection index number was closest to 1 was used as the output value.

FIG. 13 and FIG. 1 showing the results of this example.
As is clear from comparison between FIG. 4 and FIG. 8 and FIG. 9 showing the results of the conventional method, it can be seen that the estimation accuracy is extremely high according to the present embodiment.

In order to further clarify the effect of the present invention, the unknown heat quantity and the unknown oxygen quantity are estimated by the conventional method using the model equations (7) and (8) for the same converter operation. In the present invention (first invention method), only in the first and second neural networks outputting the continuous quantity (analog value) shown in the first embodiment (first method). 3) combinations of the neural networks of (A) to (C) (second invention method (A) to
Table 1 shows simultaneous target predictive ratios of blown carbon concentration and blown molten steel temperature when static blowing calculation is performed by estimating unknown heat quantity and unknown oxygen quantity. Indicated.

Here, as in the case of the first embodiment, the simultaneous target hit ratio is the same charge (same heat), and the actual blown carbon concentration is within the target value ± 0.01%. In addition, it is a percentage of the total number of charges when the actual blown molten steel temperature falls within the range of the target value ± 6 ° C.

[0139]

[Table 1]

As is apparent from Table 1 above, the simultaneous target predictive ratios of the first invention method and the second invention methods (A) to (C) are almost at the same level, which is dramatically improved as compared with the conventional method. You can see that

According to this embodiment described in detail above, in the static blowing calculation in converter steelmaking, the unknown heat amount and the unknown oxygen amount are estimated by the neural network, and the units forming the neural network are self-learned. Since the weighting coefficient of the coupling is modified, the estimation accuracy can be maintained high even when the operating conditions change with time.

Further, in the construction and learning of the third and fourth neural networks, as described above, the output layer includes the carbon concentration at the time of blow stop and the tolerance of the molten steel temperature, the oxygen gas injection control equipment and the coolant. Comprehensively judge from control error and measurement error of metering / input equipment, empirical knowledge of operator, etc., and divide unknown heat quantity and / or unknown oxygen quantity into multiple classes each having the same width, and The output layer consists of the same number of output units, and each output unit has 0
Since one priority selection index number consisting of a value of ~ 1 is output, the carbon concentration at the end point and the carbon concentration at the end point are compared to the conventional method of outputting estimated values as analog values by a fixed model formula. The target precision of the molten steel temperature can be improved.

In this embodiment, the estimated value is output as a continuous analog value while correcting the coupling coefficient between the units forming the neural network by self-learning.
Compared with the method of the first embodiment, a blowing calculation method that is much closer to a judgment method as if an operator selects an option while maintaining the target end point carbon concentration and / or molten steel temperature of the equivalent end point. Can be realized by a computer.

As a result, the simultaneous target hit ratio for simultaneously hitting the molten steel carbon concentration and molten steel temperature at the target value at the time of blowing stop is
It was possible to improve from about 60% of the method using the conventional model formula to about 90%. In addition, as a result, the increase in the dissolved oxygen content in iron and steel in the slag due to the excessive reduction of the blowing carbon concentration can be reduced, the yield of steel output can be improved, and the basic unit of deoxidizing aluminum can be reduced. It was

The present invention has been specifically described above, but the present invention is not limited to the above-mentioned embodiments, and various modifications can be made without departing from the scope of the invention.

For example, in the first embodiment, the case where the first neural network that outputs the unknown heat quantity and the second neural network that outputs the unknown oxygen quantity is separated is shown. The output layer may be composed of two units, that is, the unit O _A that outputs the unknown heat amount and the unit O _B that outputs the unknown oxygen amount, which are integrated.

In the second embodiment, the unknown heat quantity is output.
Output the unknown oxygen amount with the third neural network
A configuration in which two of the fourth neural networks are separated.
In this case, the neural network is integrated.
And multiple units O that output unknown heat quantity_XAnd unknown acid
Multiple units O that output elementary quantities_YThe output layer is composed of two groups.
May be done. Also, the first neural network and the fourth
A neural network, or a second neural network
Network and third neural network respectively
One output unit O for each unit_AAnd multiple
Output unit O _YOr with multiple output units O
_XAnd one output unit O_BYou can configure the output layer with and
Yes.

Further, in the above-described embodiment, the first and third neural networks (FIGS. 1 and 1) that output the unknown heat quantity are used.
1) and the second and fourth neural networks (FIGS. 2 and 12) that output the unknown oxygen amount have the same input layer and intermediate layer configurations, but needless to say, they may be different from each other. Yes.

Further, in the above-mentioned embodiment, the case of using the neural network having the three-layer structure has been described, but the present invention is not limited to this, and the number of layers may be four or more.
Further, the number of priority selection indexes is not limited to 0 to 1, and an arbitrary numerical range can be set.

Further, although the case of the upper-bottom blowing converter has been described in the above embodiment, the present invention can be applied to other blowing furnaces such as the bottom-blowing converter or the upper-blowing converter. Needless to say.

[0151]

As described above, according to the first invention,
Since the unknown heat quantity and unknown oxygen quantity contained in the model formula used for static blowing control can be estimated with high accuracy at all times, the static blowing to keep the molten steel component concentration and molten steel temperature at the target value at the target value It is possible to always perform high precision control.

According to the second invention, in addition to the effects of the first invention, in the static blowing calculation of the converter, the class of the unknown heat quantity and / or the unknown oxygen quantity is selected by the neural network. As a result, it is possible to perform operations close to the operator's sense and judgment without lowering the target accuracy.

Further, according to the present invention, when classifying data into classes, it is possible to incorporate even less frequent operation patterns by uniforming learning data, and improve the end point precision. Becomes

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram showing a first neural network applied to an embodiment according to a first invention.

FIG. 2 is a schematic configuration diagram showing a part of a second neural network applied to the same embodiment.

FIG. 3 is an explanatory diagram showing an outline of a device configuration applied to the embodiment.

FIG. 4 is a graph showing the estimation accuracy of unknown heat quantity according to the example.

FIG. 5 is a graph showing the estimation accuracy of the unknown oxygen amount according to the example.

FIG. 6 is another graph showing the estimation accuracy of the unknown heat quantity according to the example.

FIG. 7 is another graph showing the estimation accuracy of the unknown oxygen amount according to the same example.

FIG. 8 is a graph showing the estimation accuracy of unknown heat quantity by the conventional method.

FIG. 9 is a graph showing the estimation accuracy of the unknown oxygen amount by the conventional method.

FIG. 10 is a graph showing the simultaneous target blow-in stop ratio in the embodiment of the first invention and the conventional method.

FIG. 11 is a schematic configuration diagram showing a third neural network applied to one embodiment according to the second invention.

FIG. 12 is a schematic configuration diagram showing a part of a fourth neural network applied to the embodiment.

FIG. 13 is a graph showing the estimation accuracy of unknown heat quantity according to the example.

FIG. 14 is a graph showing the estimation accuracy of the unknown oxygen amount according to the example.

[Explanation of symbols]

 10 ... Converter 12 ... Process computer 14 ... Learning server 16 ... Main lance 18 ... Sublance

Claims (7)

[Claims] Claim: What is claimed is: 1. A static converter for steelmaking, wherein the amount of injected oxygen and the amount of coolant input are determined based on the heat balance and mass balance in the converter steelmaking, and the molten steel temperature and the concentration of molten steel components at the time of blowing are appropriately adjusted to target values. In the dynamic blowing control method, the operation information in converter steelmaking is used as input, and the unknown heat quantity, which is the difference between the theoretical heat input quantity and the theoretical heat output quantity, and the unknown oxygen quantity, which is the difference between the theoretical generated oxygen quantity and the theoretical consumed oxygen quantity. When an unknown heat quantity and an unknown oxygen quantity are estimated using the neural network by using a neural network as an output, the actual operation information when actually operating and the unknown heat quantity and the unknown oxygen quantity are used. A method for static blowing control in converter steelmaking, characterized in that the neural network is learned and the coupling coefficient between units forming the neural network is modified. 2. The method according to claim 1, wherein pre-processing is performed when learning the neural network.
The unknown heat quantity and the unknown oxygen quantity are divided into a plurality of classes each set with a predetermined difference width, and when the record frequency of the class to which the unknown record heat quantity and the unknown record oxygen quantity correspond is low, that class The operating temperature data corresponding to the above is duplicated by copying to increase the frequency distribution of the learning data so that the frequency distribution of the learning data for each class is almost uniform. Static blowing control method. 3. A static converter for steelmaking in which the amount of blown oxygen and the amount of coolant injected are determined based on the heat balance and the mass balance in the converter steelmaking, and the molten steel temperature and the concentration of molten steel components at the time of blowing stop are appropriately adjusted to the target values. In the dynamic blowing control method, the operation information in converter steelmaking is used as input, and the unknown heat quantity, which is the difference between the theoretical heat input quantity and the theoretical heat output quantity, and the unknown oxygen quantity, which is the difference between the theoretical generated oxygen quantity and the theoretical consumed oxygen quantity. When constructing a neural network that outputs, and estimating the unknown heat quantity and the unknown oxygen quantity using the neural network, at least one of the unknown heat quantity and the unknown oxygen quantity is set to a plurality of classes set with the same difference width. Configure the output layer of the neural network that divides and estimates at least one of the unknown heat quantity and the unknown oxygen quantity with the same number of units as the above classes, and configure the output layer. From each unit, so that the priority selection index number, which is the criterion for selecting the corresponding class, is output, using the actual operation information when actually operating, and the unknown quantity of heat and the unknown quantity of oxygen, A method for controlling static blowing in converter steelmaking, which comprises learning the neural network and correcting a coupling coefficient between units constituting the neural network. 4. The unit according to claim 3, wherein when learning or outputting a neural network having an output layer composed of the same number of units as the number of classes, a unit of a class to which an unknown amount of heat or an unknown amount of oxygen corresponds. The priority selection index number 1 is input to the units of the other classes, and the priority selection index number 0 is input as teacher data, and each output unit outputs the priority selection index numbers 0 to 1. A method of static blowing control in converter steelmaking. 5. The learning according to claim 3, wherein, when learning a neural network having an output layer composed of the same number of units as the number of classes, as a preprocessing, a heat quantity of unknown performance or an oxygen quantity of unknown performance is applied to If the actual record frequency is low, increase the number of copy by copying the operational record data and make the learning data frequency distribution uniform so that the learning data frequency distribution for each class is almost uniform. A method of static blowing control in converter steelmaking. 6. In claim 3, when the output layer is composed of a single unit, the unknown heat amount or the unknown oxygen amount is output as a continuous amount, and the learning result unknown heat amount or the unknown result oxygen amount is continuously output during learning. A static blowing control method in converter steelmaking, characterized in that the quantity is input to the unit. 7. In claim 3, when the output layer is composed of a single unit, the unknown heat quantity or the unknown oxygen quantity is output as a continuous quantity, and at the time of learning, the unknown heat quantity and the unknown oxygen quantity are pretreated. Is divided into a plurality of classes that are each set with a predetermined difference, and when the record frequency of the class to which each of the unknown quantity of heat and the unknown quantity of oxygen corresponds is small, the operation result data corresponding to that class is calculated. A static blowing control method in converter steelmaking, characterized in that the frequency distribution of learning data is preliminarily made uniform so that the frequency distribution of learning data for each class becomes almost uniform by making duplicates by copying. .