JPH0559415A - Apparatus for estimating furnace heat variation in blast furnace - Google Patents

Abstract

PURPOSE:To provide an estimating apparatus for furnace heat variation in a blast furnace, which can be precisely executed as meeting to the actual condition in the blast furnace. CONSTITUTION:In the above estimating apparatus for furnace heat variation in the blast furnace 1, what is called, Kohonen type neural network 2 is applied. Then, in a neuron specified at the time of learning the neural network 2, distributing data P for wall dropping position for learning specifying this neuron, variation in solution loss accompanied with this, qualitative data of lowering of the furnace heat at this time and threshold value of the above variation obtd. based on this qualitative data, are allocated. Then, the neuron is specified based on a weight already learnt with the distributing data P for wall dropping position newly inputted at the time of estimating the lowering of the furnace heat, and when the variation in solution loss at this time exceeds the threshold value in the above neuron, alarm related to the lowering of furnace heat, is outputted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a blast furnace furnace heat change estimating device for estimating a change in furnace heat of a blast furnace, and in particular, to perform learning without needing teacher data for estimating the change in furnace heat. The present invention relates to a thermal change estimation device for blast furnaces using a neural network that uses a self-organizing feature mapping algorithm.

[0002]

2. Description of the Related Art When a so-called (wall drop) occurs, in which a deposit on a furnace wall of a blast furnace separates and falls for some reason, the deposit on the furnace wall induces a decrease in furnace heat, for example, a hot metal temperature, which causes a problem in operation. May have adverse effects. In such a case, conventionally, a skilled operator predicts a decrease in furnace heat based on information from various sensors arranged in the blast furnace, and takes an appropriate operation action to decrease the furnace heat according to the prediction result. I had to do it in advance. Therefore, the inventors of the present invention developed a blast furnace operating state estimating device that automatically estimates the decrease in the furnace heat without using a skilled operator, and filed an application as Japanese Patent Application No. 3-179275. This blast furnace operation state estimation device is provided with a so-called Kohonen type neural network to which the self-organizing feature mapping algorithm is applied. Further, the blast furnace operating state estimation device monitors each furnace wall temperature data from a plurality of temperature sensors provided on the furnace wall of the blast furnace, and based on a change in the furnace wall temperature data within a predetermined period in the past, the temperature concerned is monitored. Wall drops at the sensor positions are determined, and wall drop position distribution data indicating the presence or absence of wall drops at all temperature sensor positions is created. Therefore, at the time of learning of the neural network, a neuron to which the prepared wall drop position distribution data for learning is input to the neural network is specified, and the learning of the weight mainly associated with this neuron and neighboring neurons is automatically performed. And the learned weights are finally stored in the memory. At the same time, a furnace heat decrease index, which corresponds to the above-mentioned wall fall position distribution data for learning and empirically indicates a future decrease in furnace heat, is assigned to the specified neuron. The furnace heat drop index assigned to each neuron is averaged and the average value is stored in the memory. Therefore, at the time of estimating the reactor heat drop, new wall drop position distribution data is input to the learned neural network, and the neuron having the weight most approximate to the new wall drop position distribution data is specified. Then, the average value of the furnace heat decrease index assigned to the specified neuron is estimated as the degree of the furnace heat decrease of the blast furnace.

[0003]

However, the change in the furnace heat of the blast furnace as described above depends on only the wall fall position distribution data showing only the presence or absence of the wall fall obtained based on the furnace wall temperature distribution data. There are times when I cannot explain. This is because various operating factors are involved in the change in the furnace heat in a complicated manner. Other operating factors other than the wall drop position distribution data are, for example, changes in the furnace wall temperature at the wall drop position and changes in the so-called solution loss reaction in the furnace. The solution loss reaction is an endothermic reaction in which coke is wasted by carbon dioxide gas once generated in the furnace, and is represented by the following reaction formula. CO2 + C → 2CO + ΔH However, ΔH is negative reaction heat. However, in the invention of the previous application by the present inventors, the above-mentioned other operating factors such as the change of the furnace wall temperature or the change of the solution loss reaction at the falling position of the furnace wall as described above are not taken into consideration. Instead, there was a request to estimate the change in furnace heat more closely to the actual condition of the blast furnace. Therefore, the object of the present invention is to estimate not only the wall fall position distribution data but also the change in the furnace wall temperature at the wall fall position and the change in the solution loss reaction described above, which is more suitable for the actual condition of the blast furnace. An object of the present invention is to provide an accurate blast furnace heat change estimation device.

[0004]

In order to achieve the above-mentioned object, the main means adopted by the first invention is, in essence, the fact that the inside of the furnace is based on the temperature distribution data of the furnace wall within a predetermined period of the blast furnace. In a blast furnace thermal change estimation device for estimating temperature changes, at least the amount of carbon charged into the blast furnace, the amount of carbon used steadily in the blast furnace, and the amount of pig iron production within the predetermined period. From the solution loss change calculation means for calculating the change data of the solution loss in the furnace for each of the furnace wall temperature distribution data and the input data by giving a calculation weight to each of the input data forming the furnace wall temperature distribution data. A plurality of information processing units that are conceptually arranged in a grid pattern to output a function related to the sum of products of the calculation weights, and preset learning wall temperature distribution data for learning are stored in the plurality of information processing units. By changing the calculation weights of the information processing unit and the information processing unit at the address near the specified information processing unit by inputting into the unit and calculating, an environment with different calculation weights is formed between the information processing units. A first neural network having a data classification means for classifying the furnace wall temperature distribution data for each information processing unit corresponding thereto, and the above classification obtained corresponding to different light and light environments between the information processing units. The subsequent calculation weights are stored, and the change in the reactor temperature obtained from the change data of the solution loss and the change data of the reactor temperature at that time for each of the classified furnace wall temperature distribution data for learning is estimated. For storing the threshold value of the solution loss change amount for each of the specified information processing units, and the new furnace wall temperature distribution data in the first neural network. When a force is applied, the change in the furnace temperature of the blast furnace is made based on the threshold value of the solution loss change amount of the information processing unit specified on the basis of the calculation weight stored in the first memory means or the information processing unit close to the information processing unit. It is configured as a blast furnace thermal change estimation device according to the point including a first calculation means for estimating.

Further, in order to achieve the above object, the main means adopted by the second invention is the gist thereof, in which the furnace wall is dropped due to the falling of the furnace wall based on the furnace wall temperature distribution data within a predetermined period of the blast furnace. In a blast furnace thermal change estimation device for estimating a change in internal temperature, a furnace wall drop position distribution data creating means for creating a drop position distribution data of a furnace wall based on the furnace wall temperature distribution data within the predetermined period, and Wall drop position temperature change calculation means for calculating change data of the furnace wall temperature at the drop position of the furnace wall for each created drop position distribution data, and calculation weight for each input data constituting the drop position distribution data Then, a plurality of information processing units conceptually arranged in a grid pattern to output a function relating to the sum of products of the input data and the calculation weights, and a plurality of preset falling position distribution data for learning are provided. The learning is performed by changing the calculation weights of the information processing unit specified by inputting to the information processing unit and performing calculation and the information processing units at the addresses near the information processing unit to form environments with different calculation weights between the information processing units. Neural network having a data classifying unit that classifies the dropout position distribution data for each information processing unit corresponding thereto,
The calculation weights after the classification obtained corresponding to different light and light environments between the information processing units are stored, and the change data of the furnace wall temperature for each of the classified falling position distribution data for learning is stored. Second memory means for storing the threshold value of the furnace wall temperature change amount for estimating the change in the furnace temperature, which is obtained from the change data in the furnace temperature at that time, for each specified information processing unit, When the falling position distribution data is input to the second neural network, the furnace wall temperature change amount of the information processing unit specified based on the calculation weight stored in the second memory means or an information processing unit close to the information processing unit. And a second calculation means for estimating a change in the temperature inside the blast furnace on the basis of the threshold value of 1.

Further, in order to achieve the above object, a third
The main means adopted by the invention of claim 1 is, in the blast furnace heat change estimation device for estimating the change in the furnace temperature based on the furnace wall temperature distribution data within the predetermined period of the blast furnace, within the predetermined period. For creating the falling wall position distribution data of the furnace wall based on the furnace wall temperature distribution data, and the furnace wall temperature at the falling position of the furnace wall for each of the created falling position distribution data. Wall fall position temperature change calculation means for calculating change data, and at least
Based on the amount of carbon put into the furnace of the blast furnace, the amount of carbon used steadily in the furnace, and the amount of pig iron produced within the predetermined period, the change data of the solution loss in the furnace is used to obtain the temperature distribution data of the furnace wall. A solution loss change calculating means for each calculation, and a calculation weight is given to each input data forming the dropout position distribution data, and conceptually arranged in a grid pattern so as to output a function related to the sum of products of the input data and the calculation weight. Calculation of the specified information processing unit and the information processing unit specified by inputting and calculating the preset falling position distribution data for learning into the plurality of information processing units A data amount for changing the weights to form environments with different calculation weights between the information processing units and classifying the learning dropout position distribution data for each corresponding information processing unit. A third neural network having a means, and storing the post-classification operation weights obtained in correspondence with different heavy and light environments between the information processing units, and storing the classified falling position distribution for learning. Reactor wall temperature change amount threshold value and solution loss for estimating the change in the furnace temperature obtained from the change data of the furnace wall temperature for each data and the change data of the solution loss and the change data of the furnace temperature at that time Third memory means for storing the threshold value of the variation amount for each of the identified information processing units,
When the new falling position distribution data is input to the third neural network, the furnace wall temperature of the information processing unit specified on the basis of the calculation weight stored in the third memory means or an information processing unit close to the information processing unit. It is configured as a blast furnace heat change estimation device according to a point including a third calculation means for estimating a change in the in-furnace temperature of the blast furnace based on each threshold value of the change amount and the solution loss change amount.

[0007]

According to the first aspect of the present invention, first, the solution loss change calculation means at least within a predetermined period, the amount of carbon charged into the furnace of the blast furnace, the amount of carbon used steadily in the furnace, and the pig iron. From the production amount, the change data of the learning solution loss in the furnace is calculated for each furnace wall temperature distribution data. Subsequently, at the time of learning the first neural network, the data classification means inputs the preset learning wall temperature distribution data for learning into a plurality of information processing units conceptually arranged in a grid, and specifies the data. An information processing unit having a small difference between the calculation weights associated with each of the specified information processing units and the furnace wall temperature distribution data is specified, and the calculation weights of the specified information processing unit and the information processing units at the addresses near the specified information processing unit are specified. Is automatically changed to form environments with different calculation weights between the information processing units, and the furnace wall temperature distribution data for learning is classified for each information processing unit corresponding thereto. At this time, the calculated weights obtained after the classification corresponding to different heavy and light environments between the information processing units are stored in the first memory means, and the solution for each learning furnace wall temperature distribution data is stored. The solution loss change threshold for estimating the change in the furnace temperature obtained from the change data of the loss and the change data of the furnace temperature at that time is also stored in the first memory means for each of the specified information processing units. To be done. When new furnace wall temperature distribution data is input to the learned first neural network when estimating the change in the furnace temperature, the first memory is used by the new furnace wall temperature distribution data. An information processing unit or an information processing unit close thereto is specified based on the calculation weight stored in the means. Therefore, the first computing means corresponds to the specified information processing section or an information processing section close to the specified information processing section, and based on the threshold value of the solution loss change amount stored in the first memory means, the furnace of the blast furnace. Estimate changes in internal temperature. Also, the second
According to the invention, instead of the furnace wall temperature distribution data, the solution loss change amount and the threshold value thereof in the first invention, the wall fall position distribution data and the wall fall position obtained based on the furnace wall temperature distribution data are obtained. The amount of change in the furnace wall temperature and the threshold value thereof are used. Therefore, similarly to the first aspect of the invention, it is possible to estimate the change in the furnace temperature that further reflects the actual condition of the blast furnace. Furthermore, according to the third aspect of the invention, the change in the furnace temperature is estimated in consideration of both the change in the solution loss and the change in the furnace wall temperature at the wall drop position. Therefore, it is possible to estimate changes in the furnace temperature that further reflect the actual condition of the blast furnace.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments embodying the present invention will be described below with reference to the accompanying drawings for the understanding of the present invention. FIG. 1 is a block diagram showing a blast furnace heat change estimating apparatus according to an embodiment of the present invention, FIG. 2 is a conceptual diagram showing a neural network included in the blast furnace heat change estimating apparatus, and FIG. FIG. 4 is a conceptual diagram showing neurons (information processing units) that constitute the neural network, FIG. 4 is a conceptual diagram showing temporal changes in the arrangement state of neurons of the neural network and the range of “bubbles”, and FIG. 5 is firing of the neural network. 6 is a graph showing the relationship between the distance from the specific neuron and the activity of the neuron output, FIG. 6 is a distribution diagram showing wall drop position data in the height direction and the circumferential direction of the blast furnace wall, and FIG. (a) to (d) are distribution charts showing wall drop position distribution data shifted in the circumferential direction at the same height of the blast furnace, respectively, and FIG. 8 is a wall expressed based on the learned load vector. FIG. 9 is a map showing the position distribution data for each arrangement of neurons, and FIG. 9 is an explanatory diagram showing the qualitative data of the reactor heat drop assigned to the 0th row and 0th column neurons and the accompanying variation in solution loss. 10 is a table diagram showing the estimation result when the furnace heat drop is estimated in consideration of the change amount of solution loss, and FIG. 11 shows the change amount of solution loss and the change amount of furnace wall temperature at the wall fall position. It is a table figure which shows the estimation result when estimating a furnace heat fall. The following embodiments are merely examples embodying the present invention and are not of the nature to limit the technical scope of the present invention.

Blast furnace furnace heat change estimation apparatus 1 according to the present embodiment
As shown in FIG. 1, the wall fall position distribution data created based on the furnace wall temperature data (furnace wall temperature distribution data) within a predetermined period from a plurality of temperature sensors 11 provided in the blast furnace 10 is configured. A so-called Kohonen-type neural network 2 that includes a plurality of neurons 5 (information processing units (FIG. 3)) that give arithmetic weights to each of the input data and output the data and that performs pattern recognition using a self-organizing feature mapping algorithm. (First, second, and third neural networks), the weight vector (computation weight) after the classification obtained corresponding to different heavy and light environments between the neurons 5 at the time of learning, and the wall for learning. A memory 3 (first, second and third memory means) for storing the threshold value of the solution loss change amount for each position distribution data and the threshold value of the furnace wall temperature change amount at the falling position of the furnace wall. And the change data of the solution loss and the change data of the furnace wall temperature at the wall fall position for each of the plurality of wall fall position distribution data classified for each neuron 5, and these change data are assigned to the neuron 5. At the same time, when new wall fall position distribution data is input to the neural network 2 at the time of estimation, based on the threshold values of the solution loss change amount and / or the furnace wall temperature change amount of the neuron 5 that conceptually fires, Control unit 4 for estimating the furnace heat drop of the blast furnace 10
(First, second and third computing means) and temperature sensor 11
And an input / output port 12 for outputting an alarm output related to the estimated furnace heat drop to the blast furnace 10 while inputting the furnace wall temperature data from

The arithmetic and control unit 4 receives furnace wall temperature data at predetermined sampling times from a plurality of temperature sensors 11 provided on the furnace wall of the blast furnace 10 within a predetermined period, for example, one hour before a certain time. The average value of is calculated for each temperature sensor 11. The average value of these furnace wall temperature data is 1
It is obtained at each time and stored in a memory not shown. And
The arithmetic control unit 4 determines the wall drop of the furnace wall in the vicinity of the installation position of the temperature sensor 11 based on the average value of the furnace wall temperature data using the wall drop criterion described below. Generally, the furnace wall of the blast furnace 10 is cooled to a temperature of, for example, 150 ° C. or lower by a cooling means arranged on the furnace wall. Therefore, when the furnace wall is in a steady state, the furnace wall temperature is 150 ° C.
Although the temperature is continuously maintained below, when the deposits on the furnace wall separate and fall off from the furnace wall, the hot air in the furnace directly hits the furnace wall at that position, and the temperature of the furnace wall at that position rises. Therefore, as the wall drop criterion, for example, the average value of the furnace wall temperature data obtained every hour is maintained at 150 ° C. or lower for at least 24 hours before 1 hour before the time and for 1 hour before that time. When the average value at a certain time becomes 100 ° C. or more higher than the average value of the furnace wall temperature data obtained previously, it is determined that a wall drop has occurred near the temperature sensor 11 that has detected the furnace wall temperature data. Method is used. Then, the arithmetic control unit 4 sets a value of 1 to the position corresponding to the furnace wall position in the data map of FIG. 6 where it is determined that a wall drop has occurred, and sets a value of 0 to a position where no wall drop has occurred. It is supposed to do. That is, as shown in FIG. 6, the wall drop position pattern U (data area of 1) surrounded by the broken line is formed, and the wall drop position distribution data P is set by the value of 0/1 set in all the temperature sensors 11. Is configured. FIG. 6 shows a state in which the furnace wall of the blast furnace 10 is expanded in the circumferential direction, and temperature sensors 11 are provided corresponding to the furnace walls at the positions marked with 0/1 in the figure. That is, the furnace wall of the blast furnace 10 has 8
A total of 80 temperature sensors 11 at the circumferential direction position r (r = 1 to 8) of the azimuth and the height position s (s = 1 to 10) of 10 steps.
Have been deployed.

Further, the arithmetic and control unit 4 has seven walls having the same shape in the direction of the height position s of the furnace wall, which are shifted to the circumferential position r by one position based on the wall fall position distribution data P. The drop position distribution data p is created. A mode in which the wall fall position distribution data p is created from such wall fall position distribution data P is shown in FIGS. In the figure, the asterisk marked in the wall drop position pattern U and the wall drop position pattern u surrounded by broken lines indicates that a value of 1 is set, and a value of 0 is set in the portions other than each pattern. Indicates that it is set. Note that, in FIG. 7, only three wall fall position distribution data p ((b) to (d)) which are deviated from the wall fall position distribution data P ((a)) by two directions in the circumferential direction are illustrated. Naturally, similar wall fall position distribution data relating to the remaining orientations are also created. Then, the arithmetic control unit 4 appropriately extracts, from the temperature sensors 11 (the wall drop position pattern U) that have determined that a wall drop has occurred, a temperature rise that is particularly large from the one hour before, and extracts from these. The average value Tn of the furnace wall temperatures is calculated, and the average value Tn of the furnace wall temperatures and all (80
Point) is the difference between the average value Ta of all the furnace wall temperatures from the temperature sensor 11 and the variation Td of the furnace wall temperature at the wall drop position.
(Change data). As a matter of course, the larger the amount of change Td of the furnace wall temperature, the greater the degree of wall drop, which is likely to cause a decrease in furnace heat. That is, the means for executing the procedure of calculating the change data of the furnace wall temperature at the falling position of the furnace wall for each falling position distribution data by the calculation control unit 4 is the wall falling position temperature change referred to in the second and third inventions. It is a calculation means. On the other hand, the solution loss cannot be directly measured because it is an endothermic reaction occurring in the furnace. Therefore, the arithmetic control unit 4 is used effectively without waste in the furnace within a period of 1 hour before a certain time to, the amount of carbon Ce introduced into the furnace,
That is, the amount of carbon Cs used in a steady state and the amount of pig iron production M
From the (pig iron production rate), the solution loss Sn (to) at that time to is calculated using the following formula. Sn (to) = K · (Ce−Cs) / M (1) where K: positive coefficient Note that the amount of carbon Ce charged into the furnace, the amount of carbon Cs used in the steady state, the pig iron The production amount M is obtained as follows. M: total amount of iron put into the furnace, blast temperature of blast blown into the furnace, blast humidity, CO, CO ₂ , N in the furnace top gas
₂ Calculated from gas concentration of H ₂ . Ce = Cg = Ctg-Ca (2) Here, Cg: carbon amount produced by gasification of injected coke Ctg: carbon amount in furnace top gas ... blast temperature, blast humidity, CO in furnace top gas , CO ₂ , calculated from gas concentrations of N ₂ . Ca: Carbon amount produced by decomposition of limestone Cs = Ch + Co + Cr (3) Here, Ch: Carbon amount required for decomposition of water vapor contained in blast ... Calculated from blast temperature and blast humidity. Co: amount of carbon consumed by oxygen contained in the blast ...
Calculated from oxygen concentration in the furnace, blast temperature, and blast humidity. Cr: amount of carbon required for reduction of each oxide of silicon, manganese, phosphorus ... Calculated from content rates of silicon, manganese, phosphorus contained in hot metal and pig iron production amount. Then, the arithmetic control unit 4 determines the solution loss Sn (to) at the time point to and the past 24 from the time point to.
The difference from the average value Sa of the solution loss obtained during the time is obtained as the change amount Sd (change data) of the solution loss as shown in the following equation.

[Equation 1] In a blast furnace, the amount of change S in the above solution loss is generally
It is considered that the larger the d is, the more the above endothermic reaction occurs, and the furnace heat is likely to decrease. That is, the means for executing the procedure for calculating the solution loss change data for each furnace wall temperature distribution data by the operation control unit 4 is the solution loss change calculating means in the first and third aspects of the invention.

Next, an outline of the Kohonen type neural network 2 described above will be described. The neuron 5 proposed by Professor Zevo Kohonen, as shown in the model in FIG. 3, has a weight vector W _{i, j} (i is a neuron number, j is a neuron number) for each input constituting the wall fall position distribution data. Input signal number). The activity O _i of the output of the neuron 5 is obtained from the sum of inner products of the weight vector W _{i, j} and the input signal X _j as the input vector, as shown in the following equation (5). O _i = 1 / (1 + e ^−Ii ) ... (5) However, I _i = ΣX _j W _{i, j} Furthermore, FIG. 2 shows a conceptual basic structure of the neural network 2 in which the above-mentioned neurons 5 are connected. The neural network 2 includes 80 input signals X ₁ , ..., X ₈₀ (temperature sensor 11,
Each of the 0/1 data corresponding to the furnace wall temperature data from 11, ... Is individually connected to the 25 neurons 5, and the output signals O ₁ , ..., O _{25 of} each neuron 5 are individually connected.
To the input side from the feedback signals F ₁ , ..., F ₂₅
A feedback coupling is formed to feed back. As shown in FIG. 5, the feedback coupling has a function of dragging and increasing the activity of a neuron 5 adjacent to the specific neuron 5 when the neuron 5 is fired and identified, and the activity O _i thereof increases. It has. In this way, when a certain neuron 5 is activated, the neurons 5 having relatively high activity that are dragged around the particular neuron 5 are collectively referred to as “bubble”.
(The neuron 5 corresponding to the position indicated by B in the figure is called).

The neural network 2 is
Each load vector W of Ron 5 _{i, j}Is input
Network such as changing for each wall fall position distribution data
It is necessary to study ku. Therefore, the load
Cutle W_{i, j}Is learned by the following equation (6). dW_{i, j}/ Dt = aO_iX_j-G (O_i) W_{i, j}(6) However, a = learning coefficient and a> 0. here,
The first term on the right side is the product of the activity of the input and the output of neuron 5.
The load vector W proportionally_{i, j}Heb that changes
It is a term that follows the law of ebb). The second term on the right side is the front
Note Load vector that will only increase if only the first term
W_{i, j}Is a term for suppressing. In addition, the neuro
As shown in FIG. 4, the channel 5 is roughly divided into a two-dimensional lattice with m rows and n columns.
It is arranged just in case. The neural network shown in FIG.
In work 2, neurons are arranged in a two-dimensional lattice of 10 rows and 10 columns.
5 is provided, which facilitates the explanation of the learning mode.
This is because the neural network applied to this embodiment is
The neurons 5 of the network 2 are arranged in a two-dimensional lattice of 5 rows and 5 columns.
It is equipped. Then, at the time of actual learning,
And output activity O_iWhether or not to activate
Chi, O_iIs set to {0, 1} and the suppression term is simplified
I.e. g (O_i) Is {0, a}, and the load vector is
Toll W_{i, j}Learning has been simplified. That is,
Using equations (7) and (8), O_i= 1, g (O_i) = A (inside the range of bubbles), dW_{i, j}/ Dt = a (X_j-W_{i, j})… (7) O_i= 0, g (O_i) = 0 (outside the range of bubbles), dW_{i, j}/Dt=0...(8)_{i, j}Learn. On here
The learning coefficient a and the size of the range of the bubble described above are included in the above learning.
In order to bundle, it is reduced as the number of learnings. That is,
The learning coefficient a and the size of the bubble range are a function of the time t.
Has been set. In Fig. 4, the range B of bubbles at the beginning of learning₁
Is the range B of bubbles in the middle of learning₂From the end of learning bubble range B₃
To a specific neuron 5 that fired at this time_SNext to
Secondly, we will conceptually show the situation of narrowing. Thereby, before
Load vector W_{i, j}Is similar to that of neighboring neuron 5
Similar to the above load vector W_{i, j}The value of is a neuron
Different light and heavy environments are formed so as to change smoothly between 5
To be done.

Therefore, an example in which the neural network 2 having the above-described outline is applied to the blast furnace 10 will be described below. However, in this embodiment, the change amount Sd of the solution loss is an operation factor other than the furnace wall temperature distribution data.
The case of using will be described. When the furnace wall temperature data from each temperature sensor 11 of the blast furnace 10 is input to the arithmetic control unit 4 at each sampling time, the arithmetic control unit 4 calculates the average value of the furnace wall temperature data in the previous hour every hour. Then, the wall fall position distribution data P having a value of 0/1 is created by determining the wall fall for each temperature sensor 11 based on the wall fall standard based on the average value of the obtained furnace wall temperature data. Further, the arithmetic control unit 4 creates seven wall fall position distribution data p based on the wall fall position distribution data P by shifting the wall fall position distribution data P by one azimuth in the circumferential direction r. When the learning step of the neural network 2 is executed, the wall fall position distribution data P arbitrarily created may be input to the arithmetic control unit 4 to create the shifted wall fall position distribution data p. On the other hand, prior to the learning of the neural network 2 _, the initial value of the weight vector W _{i, j} of each neuron 5 is set. At this time, a random number normalized to 0 to 1 is used as an initial value of the weight vector W _{i, j} , and these values are appropriately dispersed. Then, the learning step is executed. In the learning step, the wall fall position distribution data P for learning and the wall fall position distribution data p obtained by shifting the data in the circumferential direction are sequentially input to the neural network 2 and are input to the respective wall fall position distribution data P, p. Each neuron 5 having the weight vector W _{i, j} stored in the memory 3 having the closest value (distance) is fired (specified). At this time, the arithmetic control unit 4 extracts the neuron 5 _s having the smallest difference between the weight vector W _{i, j} associated with each of the fired neurons 5 and each wall drop position distribution data P, p. In the above case, since the wall drop position distribution data P, p is normalized to a value of 0 or 1, the wall drop position distribution data P, p and the weight vector W _i, for each neuron 5 are The distance D _i between the vectors of _j is calculated using the following equation (9).

[Equation 2]Subsequently, the extracted neuron 5_s, And its circumference
Weight vector W of side neuron 5_{i, j}Is next (10)
Modified by the formula. W_{i, j}(T + 1) = W_{i, j}(T) + a (t) (X_j(T) -W_{i, j}(T)) (10) where a (t) is empirically determined as described above.
The value ranges from 0 to 1. Furthermore, a (t) is time
It is set to decrease with the elapse of t. As well
In addition, the range of bubbles increases with the learning time as described above.
B₁, B₂, B ₃It is reduced like
Huron 5_SThe wall fall position distribution data P and p are
May be characterized as physiologically self-organizing.
Be similar.

As described above, the weight vectors W _{i, j} obtained corresponding to different heavy and light environments between the neurons 5 after the learning wall fall position distribution data P, p are classified are stored in the memory. 3 is stored. Then, the series of learning steps as described above is executed for each group of learning wall drop position distribution data P and p prepared in advance in plural groups. Also at that time, the load vector W _{i, j} stored in the memory 3 is changed and rewritten and updated in the memory 3 each time. Furthermore,
The wall fall position distribution data P of all the groups prepared in advance,
If a series of learning steps by p is one learning of the neural network 2, learning using the same wall fall position distribution data P and p for all the groups described above is repeated a predetermined number of times, for example, 500 times. The learning of the neural network 2 is completed. The learning as described above is executed for all the above-mentioned wall drop position distribution data P for learning and the wall drop position distribution data p accompanying them without the need for teacher data, and one of them is represented. The wall drop position distribution data P, p are classified into any of the neurons 5 _s that have fired. That is, the means for executing the method of the neural network 2 for classifying the learning wall fall position distribution data P or p by the feedback connection into the corresponding neurons 5 _s is the data classifying means. In this case, the falling wall distribution data P or p for learning is not classified into the neuron 5 that has not fired, but the neuron 5 is affected as bubbles of the neighboring fired neurons 5. Therefore, the closer the neuron 5 is to the fired neuron 5, the more similar the load vector W _{i, j} is set to that of the fired neuron 5.

Subsequently, each of the plurality of groups of the same learning wall fall position distribution data P and the wall fall position distribution data p is sequentially input to the learned neural network 2 for each group, and the load of the memory 3 is inputted. The change amount Sd of the solution loss corresponding to the wall drop position distribution data P for each group obtained by extracting each neuron 5 from the neurons 5 sequentially extracted based on the vector W _{i, j} is the operation control unit 4 Assigned by. Further, the arithmetic and control unit 4 estimates the decrease in the furnace heat from all solution loss changes Sd assigned to each neuron 5 and the qualitative data (change data in the furnace temperature) regarding the decrease in the furnace heat at that time. Change Sd of solution loss
The threshold value Sti (i = 1 to 25) is calculated for each neuron 5, and the threshold value Sti is stored in the memory 3. The solution loss change amount Sd corresponding to the wall fall position distribution data p deviated in the circumferential direction is the same as the solution loss change amount Sd set in the learning wall fall position distribution data P. Is. It should be noted that FIG. 8 shows an example of a map A in which the wall fall position distribution data represented based on the learned calculation weights is shown corresponding to each neuron 5. In the map A shown in FIG. 8, m rows and n columns (5 rows 5
The position of each rectangular frame shown in the column corresponds to the arrangement of the neurons 5 of the neural network 2 (FIG. 4). still,
In the figure, a plurality of dots shown in each frame are the temperature sensor positions at the circumferential position r and the height position s, and indicate that the black-painted dots are the wall drop positions. The larger the value, the greater the degree of wall drop at that position.

Here, for example, focusing on the neuron 5 in the 0th row and 0th column, as shown in FIG. 9, as qualitative data of the furnace heat drop in the near future, the "reactor heat drop large (solution loss change amount Sd Sd = 3.25) "and" Reducing furnace heat (above change Sd = 3.17, 3.42) "2 cases,
"No furnace heat drop (the above-mentioned change amount Sd = -0.96, ~ 2.
29) ”is assigned to a total of 9 cases. Here, since the alarm output is made for “a large decrease in the furnace heat” and “during the decrease in the furnace heat” in which the decrease in the furnace heat is actually detected, the threshold value of the change amount Sd of the solution loss is set for the neuron 5 concerned. 2.30 is set as Sti and stored in the memory 3.

When the estimation of the furnace heat drop is executed by using the neural network 2 which has been learned as described above, the furnace wall temperature data from each temperature sensor 11 of the blast furnace 10 is calculated and calculated at each sampling time. Input to 4. Similar to the learning step, the arithmetic and control unit 4 obtains the average value of the furnace wall temperature data within the previous hour every hour, and based on the average value of the obtained furnace wall temperature data, the wall drop criterion is calculated. Then, the wall drop for each temperature sensor 11 is determined, and the wall drop position distribution data P having a value of 0/1 is created. Further, the arithmetic control unit 4 creates seven wall fall position distribution data p based on the wall fall position distribution data P by shifting the wall fall position distribution data P by one azimuth in the circumferential direction r. That is, the calculation control unit 4 executes a procedure for creating falling wall position distribution data (wall falling position distribution data) of the furnace wall based on the furnace wall temperature distribution data within a predetermined period during learning and estimation execution. The means is a furnace wall dropout position distribution data creating means. At the same time, the arithmetic control unit 4
Is introduced into the furnace based on the temperature and humidity of the air blown into the furnace, the concentration of each gas in the furnace top gas, the amount of coke introduced, the amount of iron, the content of elements such as silicon in the hot metal, etc. The amount of carbon Ci, the amount of carbon Cs used in a steady state, and the amount of pig iron production M are calculated, and the solution loss Sn (to) at that point in time in the furnace is calculated from the formula (1), and then (4) From the formula, the change amount Sd of solution loss
Is calculated. Subsequently, the wall drop position distribution data P and the wall drop position distribution data p are sequentially input to the learned neural network 2 and stored in the memory 3 having the value (distance) closest to each wall drop position distribution data P, p. Each neuron 5 having the weight vector W _{i, j} stored in is identified. At this time, the arithmetic control unit 4 causes the neuron 5 having the smallest difference between the respective weight vectors W _{i, j} and the wall drop position distribution data P, p among the specified neurons 5.
To extract. If the extracted neuron 5 is, for example, the above-mentioned neuron 5 in the 0th row and 0th column, the operation control unit 4 is assigned to the neuron 5 and stored in the memory 3 and the amount of change in the solution loss. The future decrease in furnace heat of the blast furnace 10 is estimated based on the threshold value Sti of Sd. In this case, if the calculated solution loss change amount Sd is greater than or equal to the threshold value Sti (= 2.30), the calculation control unit 4 issues an alarm output command to the blast furnace 10 regarding a decrease in furnace heat.

In this embodiment, all neurons 5 are used for learning.
Although an example in which is fired is shown, if there is a neuron 5 that did not fire during learning, the neuron 5 does not fire during estimation. Using the blast furnace heat change estimation apparatus 1 of the present embodiment, 114 cases were actually verified for fluctuations of the furnace heat of the blast furnace 10, and FIG.
As shown in, the correct alarm is 94 (= 39 + 55)
There were 6 missed alarms, 6 missed alarms and 14 missed alarms, which was a good predictive predictive ratio of 82.5% (= (39 + 55) / 114).

In the above embodiment, the variation S of the solution loss is an operation factor different from the furnace wall temperature distribution data.
Although d is used, instead of the variation Sd of the solution loss, the variation Td of the furnace wall temperature at the wall drop position described above may be used as the another operation factor.
Also in this case, as in the case of using the change amount Sd of the solution loss, the change amount Td of the furnace wall temperature previously obtained corresponding to the wall fall position distribution data P for learning and the decrease in the furnace heat at that time are calculated. Qualitative data is prepared, and the wall fall position distribution data P for learning is input to the neural network 2 for learning. Then, based on the qualitative data of the furnace heat drop accompanying the falling wall position distribution data P for learning that specifies a certain neuron 5, the change amount Td of the furnace wall temperature for estimating the future furnace heat drop is calculated. Threshold value Tti (i = 1
25 to 25) are allocated to the neuron 5 and the memory 3
Stored in. Also at the time of estimating the furnace heat drop, when the new wall fall position distribution data P is input to the learned neural network 2 and a certain neuron 5 is specified, the wall fall position distribution data P is obtained together with the wall fall position distribution data P. The change amount Td of the furnace wall temperature at the position is compared with the threshold value Tti of the neuron 5, and when the change amount Td exceeds the threshold value Tti, an alarm regarding the decrease in the furnace heat is output. Further, as another operation factor as described above, it is also possible to use the change amount Sd of the solution loss and the change amount Td of the furnace wall temperature together, and in this case, the neuron 5 fired during learning of the neural network 2 , The change amounts Sd, Td associated with the wall fall position distribution data P that fired the neuron 5 and the qualitative data of the reactor heat decrease at that time are assigned, and based on the qualitative data, the change amounts Sd, Td. Thresholds Sti and Tti of
The storage in the memory 3 is the same as that in each of the previous embodiments. However, at the time of estimating the furnace heat drop, the threshold values Sti and Tti of the neuron 5 that are fired by the wall fall position distribution data P input to the neural network 2 and the Sd obtained along with the wall fall position distribution data P, Td is compared with each other, and calculated as the influence values X and Y of the furnace heat decrease due to the respective equations (11) and (12). X = Sd-Sti ... (11) X: Influence value of furnace heat decrease due to change of solution loss Y = Td-Tti ... (12) Y: Influence value of furnace heat decrease due to change of furnace wall temperature at the falling wall position , The total influence value F considering each influence value X and Y together is obtained from the following equation. F = αX + βY (13) α: Weighting coefficient for changes in solution loss β: Weighting coefficient for changes in furnace wall temperature at the wall drop position The threshold value Fti related to the total influence value F is set in advance for each neuron 5. , Here, a value of 0 is uniformly set for all the neurons 5. The weighting factors α and β are α = 1.0 and β = 0.01 here. Therefore, if the total influence value F obtained from the equation (13) at that time exceeds 0, an alarm related to the decrease in furnace heat is output. This embodiment also has the same 1 as the first embodiment.
The wall fall position distribution data P based on 14 furnace wall temperature distribution data was applied, but as shown in FIG. 11, 100 (= 43 + 57) correct alarm outputs, 4 false alarms, and 10 missed alarms. 87.7 (= (43+
It was possible to obtain a better predictive predictive value such as 57) / 114).

As described above, the blast furnace heat change estimation device 1 according to each embodiment is obtained from the furnace wall temperature distribution data from the temperature sensor 11 of the blast furnace 10 or the furnace wall temperature distribution data as an operation factor. Since not only the wall fall position distribution data P but also the change amount of the solution loss in the furnace and the change amount of the furnace wall temperature at the wall fall position were applied to the Kohonen type neural network 2, the decrease in the furnace heat of the blast furnace 10 It can be performed more accurately according to the actual condition of the blast furnace. The blast furnace heat change estimation device 1 has the advantages that the conventional device has, for example, (1) The calculation weights necessary for estimating the decrease in the furnace heat due to falling walls can be automatically changed. (2) It is possible to accurately estimate the decrease in furnace heat by sufficiently reflecting the pattern of the wall drop position. (3) Even if the furnace conditions change due to maintenance, etc.
The setting of the calculation weight can be easily changed by re-learning. It goes without saying that it has In addition, in each of the above-described embodiments, the wall drop position distribution data p, which is based on one wall drop position distribution data P and is deviated by one direction to the circumferential position r, is also used. Only the distribution data P may be applied to the neural network 2.
Even in this case, a corresponding effect can be obtained. Further, in the above embodiment, an alarm is issued when the calculated solution loss change amount Sd or the like exceeds the threshold value of the fired neuron 5, but in addition to this, for example, the change amount Sd and the like are added to these. It is also possible to control the blast furnace 10 to compensate for the furnace heat drop in accordance with the difference from the corresponding threshold value.

[0022]

As described above, the first invention considers not only the furnace wall temperature distribution data but also the amount of change in solution loss in the furnace, and the second invention is created based on the furnace wall temperature distribution data. Considering the amount of change in the furnace wall temperature at the wall fall position as well as the collected wall fall position distribution data, the changes in the furnace temperature are estimated by the neural network using the so-called self-organizing feature mapping algorithm. It is possible to accurately estimate the change in the furnace temperature according to the actual condition of the blast furnace. In the third aspect of the invention, the neural network is used to estimate the change in the furnace temperature in consideration of any of the wall fall position distribution data, the solution loss change amount, and the wall fall position change amount. The accuracy of estimation of changes in furnace temperature is further improved.

[Brief description of drawings]

FIG. 1 is a block configuration diagram showing a blast furnace heat change estimation device according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram showing a neural network included in the blast furnace thermal change estimation device.

FIG. 3 is a conceptual diagram showing a neuron (information processing unit) that constitutes the neural network.

FIG. 4 is a conceptual diagram showing a temporal change in arrangement state of neurons and a range of “bubble” in the neural network.

FIG. 5 is a graph showing the relationship between the distance from a specific neuron that fired and the activity of neuron output of the neural network.

FIG. 6 is a distribution chart showing wall drop position data in the height direction and the circumferential direction of the furnace wall of the blast furnace.

7 (a) to 7 (d) are distribution charts respectively showing wall drop position distribution data shifted in the circumferential direction at the same height of the blast furnace.

FIG. 8 is a map diagram showing wall fall position distribution data expressed based on a learned weight vector for each arrangement of neurons.

FIG. 9 is an explanatory diagram showing qualitative data of the furnace heat drop assigned to the 0th row and 0th column neuron and the accompanying change amount of solution loss.

FIG. 10 is a table showing an estimation result when a furnace heat drop is estimated in consideration of a change amount of solution loss.

FIG. 11 is a table diagram showing an estimation result when a furnace heat drop is estimated in consideration of a change amount of solution loss and a change amount of a furnace wall temperature at a wall fall position.

[Explanation of symbols]

1 ... Blast furnace furnace thermal change estimation device 2 ... Neural network 3 ... Memory 4 ... Operation control unit 5, 5 _S ... Neuron 10 ... Blast furnace 11 ... Temperature sensor

Claims (3)

[Claims] 1. A blast furnace thermal change estimation device for estimating a change in in-furnace temperature based on data of temperature distribution in a blast furnace wall within a predetermined period of time, wherein a blast furnace thermal change estimation device is provided in the blast furnace at least within the predetermined period. Solution loss change calculation means for calculating change data of solution loss in the furnace for each furnace wall temperature distribution data based on the input carbon amount, the carbon amount used in the furnace in a steady state and the pig iron production amount, and A plurality of information processing units conceptually arranged in a grid to give a calculation weight to each input data forming the furnace wall temperature distribution data and output a function related to the sum of products of the input data and the calculation weight, and preset By inputting the calculated learning furnace wall temperature distribution data to the plurality of information processing units and performing calculation, the calculation weights of the specified information processing unit and the information processing units at the addresses near the specified information processing unit are changed. A first neural network having data classification means for forming environments with different calculation weights between the information processing units and classifying the learning furnace wall temperature distribution data for each information processing unit corresponding thereto; Storing the operation weights after the classification obtained corresponding to different light and light environments between the information processing units, and the change data of the solution loss for each of the classified learning furnace wall temperature distribution data And first memory means for storing the threshold value of the solution loss change amount for estimating the change in the in-reactor temperature obtained from the change data in the in-reactor temperature at that time for each of the specified information processing units, and When the furnace wall temperature distribution data is input to the first neural network, the information processing unit specified based on the calculation weight stored in the first memory means or information close thereto Blast furnace thermal change estimation apparatus characterized by comprising comprises a first calculating means for estimating a change in the furnace temperature of the blast furnace, based on the solution loss change amount of the threshold of physical unit. 2. A blast furnace heat change estimation device for estimating a change in furnace temperature due to falling of a furnace wall on the basis of furnace wall temperature distribution data within a predetermined period of the blast furnace Furnace wall drop position distribution data creating means for creating the drop position distribution data of the furnace wall based on the data, and the change data of the furnace wall temperature at the drop position of the furnace wall is calculated for each of the created drop position distribution data. A wall drop position temperature change calculating means and a conceptual weight are arranged in a grid form so as to give a calculation weight to each input data constituting the drop position distribution data and output a function related to a sum of products of the input data and the calculation weight. A plurality of information processing units and the information processing units specified by inputting and calculating preset falling position distribution data for learning into the plurality of information processing units and the information processing units at the addresses near the information processing units Data classification means for changing the calculation weights to form environments with different calculation weights between the information processing units, and for classifying the falling position distribution data for learning for each information processing unit corresponding to the data The neural network of 2 and the operation weight after the classification obtained corresponding to different heavy and light environments between the information processing units are stored, and the reactor for each of the classified falling position distribution data for learning is stored. A second threshold for storing the furnace wall temperature change amount for estimating the change in the furnace temperature obtained from the change data of the wall temperature and the change data of the furnace temperature at that time is stored for each of the specified information processing units. A memory means and an information processing section specified based on the calculation weight stored in the second memory means when new falling position distribution data is input to the second neural network, or an information processing section close to the information processing section. Section above furnace wall temperature change amount of blast furnace thermal change estimation apparatus characterized by comprising comprises a second calculating means for estimating a change in the furnace temperature of the blast furnace based on a threshold. 3. A blast furnace furnace heat change estimation apparatus for estimating a change in furnace temperature based on furnace wall temperature distribution data in a predetermined period of a blast furnace, wherein the furnace is based on the furnace wall temperature distribution data in the predetermined period. Furnace wall drop position distribution data creating means for creating wall drop position distribution data, and wall drop temperature change for calculating change data of furnace wall temperature at the drop position of the furnace wall for each created drop position distribution data The calculation means and at least the change amount of solution loss in the furnace based on the amount of carbon charged into the furnace of the blast furnace, the amount of carbon used in the furnace in a steady state and the pig iron production amount within the predetermined period. A solution loss change calculating means for calculating each furnace wall temperature distribution data, and a calculation weight for each input data constituting the drop position distribution data, and a product of the input data and the calculation weight. Is specified by inputting a plurality of information processing units that are conceptually arranged in a grid pattern to output a function related to the sum of Information corresponding to the above-mentioned drop position distribution data for learning, by changing the calculation weights of the information processing unit and the information processing units at the addresses near the information processing unit to form environments with different calculation weights between the information processing units. A third neural network having a data classifying means for classifying each processing section, and storing the calculated weights after the classification obtained corresponding to different light and light environments between the information processing sections, The change in the furnace temperature obtained from the change data of the furnace wall temperature and the change data of the solution loss and the change data of the furnace temperature at that time for each of the learned falling position distribution data for learning To store the threshold value of the furnace wall temperature change amount and the solution loss change amount for each specified information processing unit, and new drop position distribution data to the third neural network. When input, the blast furnace is based on the respective thresholds of the furnace wall temperature change amount and the solution loss change amount of the information processing unit or the information processing unit close to the information processing unit specified based on the calculation weight stored in the third memory means. And a third calculation means for estimating the change in the furnace temperature of the blast furnace.