JPH0637650B2 - Blast furnace blast flow control method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an operating method for controlling a ventilation state of a blast furnace by a flow rate of air blown therein.

[Conventional technology]

For blast furnace operation, the blast flow rate should be determined first from a certain production plan, and operating at a specified air flow rate means that the productivity of the furnace is always constant and the furnace condition is kept stable. I have

On the other hand, in order to send a certain amount of air to the blast furnace, the ventilation condition inside the blast furnace must be stable and good at all times. There is. Therefore, when the air permeability in the furnace deteriorates, the air flow rate is temporarily reduced from the specified air flow rate (hereinafter referred to as the specified air flow rate) to ease the air flow conditions and increase again when the air flow rate is restored. It is necessary to perform the operation of returning by wind.

Such operations are conventionally performed by a blast furnace operator based on information from various sensors installed in the blast furnace to make a comprehensive judgment such as "poor ventilation" or "deterioration in the near future", and The method of determining and controlling the timing and amount of wind reduction has been adopted. The same applies to the return operation when the wind is reduced.

However, standardization of operations is difficult because there are individual differences based on the ability and experience of operators in determining the timing and amount of wind reduction and increase, and since evaluation is not quantitative, appropriate actions are not always taken. became. Moreover, in order to do so, a large amount of sensor information must be constantly monitored, and there is a problem that the operator's operational load is high.

In order to solve such a problem, Japanese Patent Laid-Open No. 52-10551
Proposed a formula for predicting the pressure loss level in the blast furnace, and compared it with a predicted value of the pressure loss and a dangerous value, and showed a blast flow control method that can maintain a stable pressure loss level. However, that method largely depends on the accuracy of the pressure loss prediction formula, and it is still difficult to accurately determine the blast flow rate change amount. Also, although we are only thinking about preventing blow-through, operations that slightly reduce the air flow rate must be comprehensively judged by looking at other operating factors such as slip, Anzatsu drop, furnace heat level, etc. .

Traditionally, an expert system (Daikai Sho 62-2) that creates physical models and embodies the know-how of operators
70712) has been used to predict ventilation abnormalities such as blow-throughs and slips, and how to process measured data into information that makes it easy to determine the ventilation status has been studied. However, none of them is capable of automatically adjusting the blast flow rate set value by linking a large number of observation events with blast flow reduction and increase actions in an accurate causal relationship. .

[Problems to be Solved by the Invention]

The present invention has been made to solve such problems of the prior art, and not only slip and blow-through but also Anzatsu drop, in-furnace residue, changes in furnace heat level, etc. regarding all operations of the blast furnace. It is characterized by creating a network structure that clarifies the correspondence between each piece of information and the action of reducing airflow and increasing airflow based on information, and aims at the following (1) to (3). It is a thing.

(1) To improve the accuracy of the amount and timing of actions to reduce or increase wind.

(2) Taking in all kinds of information to automatically control the flow rate of air flow and consequently reduce the load on the operator.

(3) Furthermore, the movement of each measurement data and the causal relationship between the action, ie, reduced air volume, increased air volume, and maintaining the current state are clear, and in the area where any operator performs the same action, “If ... The action is instructed by a so-called expert system that has the rule group as a knowledge base. This simplifies the system,
In areas where the causal relationship between the movement of each measurement data and the action is unclear, clarify the correspondence between each measurement information and the action using a neural network to improve the accuracy of action judgment.

[Means for Solving the Problems]

The present invention temporarily reduces the air flow rate from a specified value to control ventilation of the blast furnace, avoids abnormal conditions such as slip, hanging, and blow-through, and then increases the air flow to restore the specified air volume. In the air flow rate control method, a neural network of two or more layers is constructed by using various data as an input layer, and an output layer that outputs the output of each unit of the air reduction amount or the air increase amount divided into a plurality of stages in advance. The optimal mutual coupling coefficient is determined by learning in advance, and the sensor data for determining the ventilation state etc. installed in the blast furnace is converted into operation data calculated based on the sensor data at a predetermined timing. Use the neural network by inputting and processing each of them into data that shows recent changes and matches the operator's judgment criteria. Te is a blast furnace blast rate control method and controls the actual sensor data and the operation data to determine the wind reduction amount or increase air volume in real time based on its value automatically.

In the second invention, in such a blast furnace blast flow rate control method, an expert system is configured in parallel with the neural network, and a process (determination means) for determining the movement of each measurement data is provided to store the data. When it is determined that the causal relationship between the increase and decrease in air volume is clear, the expert system determines the cause, and when it is not clear, the neural network determines the air volume reduction or air volume increase in real time. Characterize.

The point of the present invention is to learn in advance the deterioration of air permeability, whose causal relationship was unclear, in advance, using a neural network as a teacher, with a trained operator as a teacher to learn the relationship between the wind reduction amount, the wind reduction timing or the wind increase amount, and the wind increase timing. By doing so, the amount of action and the timing are made more appropriate and accurate.

In addition, for the part where the change in the measurement data is large and the action against it is clear and invariant, a rule of the type “if ...” is created and the amount of action is determined.

Neural networks are used to make action decisions (increase wind, reduce wind, maintain current status, etc.) when many changes in measurement data are small and the changes (directions) are irregular and complicatedly related. As a means, the causal relationship with the phenomenon is clarified, and the accuracy of the action, that is, the timing of the action and the accuracy of the action amount are improved.

Here, the neural network will be briefly described with reference to pages 115 to 124 of Nikkei Electronics (August 10, 1987).

A neural network is a network that imitates the human brain. Multiple units corresponding to brain neurons are connected intricately. The pattern recognition function and the knowledge processing function can be embedded by properly determining the operation of each unit and the connection form between the units.

The structure of the unit is a model of a neuron, which is shown schematically in FIG. The structure is simple and consists of a part that receives input from other units, a part that converts the input according to a certain rule, and a part that outputs the result.

A variable weight W is attached to each of the connecting portions with other units.
_{Add ij} . This is to show the strength of the bond. Changing this value changes the network structure. Learning the network means changing this. W takes positive, zero, and negative values. Zero means no result.

When a unit receives input from multiple units, the total sum is used as the input value. The sum of input net is Is.

Each unit applies the total sum net of inputs to the function f and converts it. This input / output characteristic may differ for each unit. However, generally, the function shown in FIG. 6 (b) or (c) is often used. FIG. 6C is called a sigmoid function. With a differentiable, pseudo-linear function Can be expressed as The area is 0 to 1, and approaches 1 as the input value increases and approaches 0 as the input value decreases. When the input is 0, it becomes 0.5. Add the threshold θ, In some cases.

The network has a hierarchical structure as shown in FIG. An intermediate layer is provided between the input layer and the output layer. This intermediate layer is called a hidden layer. There are neural nets with other structures, but here, we limit ourselves to hierarchical networks that can use backpropagation algorithms.

It is connected in the direction of the network input layer, middle layer, and output layer. There is no bond within each layer. There is also no coupling from the output layer to the input layer. It is a feedforward or positive network.

Although it has a structure similar to that of the Perceptron, the capacity is dramatically improved by providing an intermediate layer and making the input / output characteristics of each unit nonlinear.

The network is trained as follows. Input data is given to each unit of the input layer. This signal is converted in each unit, propagates to the intermediate layer, and finally emerges from the output layer.
The output value is compared with the desired output value, and the coupling strength is changed so as to reduce the difference.

Learning is difficult with the middle class. This is because it is not known which W (coupling strength) causes the error. Backpropagation is a learning algorithm for such multi-layer networks. It was proposed by Ramel Hart et al. Of the University of California in 1986.

Actual output value (O _pj ) when given pattern p
And the desired output value (t _pi difference It is defined as This represents an error in output unit j. t
_pj is teacher data given by a human.

To train, you can change the strength of all bonds to reduce this error. Here, the change amount of W _ij when the pattern p is given is Decide.

By modifying this formula (5), ΔW can be formulated. The result is ΔpW _ij = ηδ _pj O _pi (6). Here, O _pi is the input value from the unit i to the unit j, and δ _pj is the unit j and the output unit is the intermediate unit. In the case of the output unit, δ _pj = (t _pj −σ _pj ) fi ′ (net _pj ) ...
(7) For the intermediate unit, Is. Expression (8) is a recursive function of δ.

The calculation of ΔW starts with units in the output layer and moves to units in the intermediate layer. In the intermediate unit, ΔW cannot be calculated unless the preceding ΔW is determined. Therefore, the calculation can be performed only by going back to the last input layer. In this way, learning proceeds in the opposite direction to the processing of the input data, that is, backward (backward).

Therefore, learning by backpropagation proceeds as follows. Input learning data and output the results (forward). Change the strength of the bond (backward) to reduce the resulting error. Input the learning data again.
This is repeated until it converges.

ΔW _ij (n + 1) = ηδ _pj O _pj + αΔW is a general formulation of ΔW.
_ij (n). n represents the number of times of learning. The first term on the right side is ΔW just obtained, and the second term is added to reduce the error vibration and accelerate the convergence.

[Action]

The function of the present invention is to determine and control the air flow reduction amount or the air flow amount in real time from various blast furnace operation data by the above-mentioned means. For example, the amount of wind reduction may not be a continuous value, for example, ONm ³ / min
(No ^{change), - 100Nm 3 / min,} -200Nm 3 / mi
n, −300 Nm ³ / min, etc. may be set in a predetermined number of steps, and simply reduced airflow −100 Nm ³ / min and 0 Nm ³ / min
It is also possible to select two outputs (no wind reduction).

〔Example〕

The present invention will be described below with reference to the drawings. FIG. 1 is a system configuration example for realizing the present invention.

A large amount of air is blown from the blower 3 to the blast furnace 1 via the hot air stove 2, and the flow rate is measured by the blower flow meter 4,
The blade angle of the blower 3 is adjusted by the control device 5 and is controlled to be constant. Blast furnace process computer 6
Has a function of centrally managing the detection end information 8 at each location in the furnace. Therefore, various measurement data of the blast furnace and operation data calculated based on them, which are necessary in the present invention, are all present in the process computer 6 including past data for a certain period. A dedicated computer 7 for carrying out the present invention has an input means 9 for each data 12, a processing means 10 for processing them, and an action amount determining means 1 including a neural network, and the determined air reduction amount or air increase amount 13 is a dedicated computer. 7 is sent to the blast furnace process computer 6, and the blast furnace process computer 6 gives the set value of the blast flow rate to the control device 5.

The above is a specific configuration example for carrying out the method of the present invention. The computers 6 and 7 may of course be the same machine. The program of the present invention for controlling the flow rate of air is executed by the computer 7 in a 5-minute cycle. It must have a real-time property corresponding to it.

Next, detailed examples will be described.

First, the data input means 9 is used. Here, the following measurement data and blast furnace operation data 12 are input from the process computer.

(1) Blower pressure: Hot air pressure of the hot-blast stove output (2) One-way index: An index showing fast and slow charging intervals, defined by the following equation.

(3) Furnace top gas temperature: Gas temperature of uptake part (4) Furnace top gas utilization rate (ηCO): The higher ηCO means that the ratio of CO gas is smaller and the ratio used for reduction is higher.

(5) Shaft pressure: The pressure of the shaft of the blast furnace divides it into two stages, the upper stage and the lower stage, and there is a detection end.

(6) Skin flow temperature: Measured by a thermometer projecting from the furnace wall into the furnace by several tens of millimeters, which is data that clearly shows the gas passing state near the wall.

(7) Stave heat load: Data indicating the amount of heat removed from the stave, which is expressed by the following equation.

STL = Fx (To-Ti) × k F: Stave cooling water amount To: Outlet water amount Ti: Inlet water temperature k: Constant (8) Stave temperature: Thermometer embedded in stave at each location of furnace body, vertical multistage and It is attached at a plurality of positions in the circumferential direction, and it is possible to detect that an inert zone (Anzatz) is attached to the furnace wall or that it is dropped.

(9) N ₂ concentration in furnace top gas: A change in the solution loss reaction due to an irregular fall of the charge in the furnace is detected.

(10) Calculated amount of slag: The amount of slag obtained by calculation from the amount and composition of the charge and the speed of charging (11) Actual amount of slag: The actual amount of slag, which is an operator with no measuring means The amount is visually measured and the computer 6
Is manually entered in.

(12) Charging circuit: the number of times of charging within a fixed time (13) Tap hot metal temperature: It is a guide for the state of heat balance in the furnace.

(14) Time from tapping to slagging (15) Slip size (16) Density index: Absolute value within Σ in equation (2) (17) Number of empty TOPED under furnace (18) Hot Metal Inventory: Hot Metal Inventory in Steel Works Each of these data is the information that the operator always needs to observe the state of the blast furnace such as the ventilation state.
However, the operator observes not only the current values of these data but also the change situation from the past to the present, and makes a good / bad decision for each observation event. The data processing means 10 processes the data so as to match the operator's observation event.

The example from here will be explained by focusing on the target when the wind volume is reduced from the specified air volume.

(1 ') The wind pressure condition is classified into three stages of bad, moderately bad, and good according to the degree to which the blowing pressure of (1) rises.

○ The current blast pressure is the average blast pressure over the last 3 hours plus 3
When it is larger than σ ... Bad ○ The current blast pressure is the average blast pressure for the last 3 hours plus 2
Greater than σ and less than 3 blast pressure averages over the last 3 hours …… Slightly worse ○ Current blast pressure averages over the last 3 hours blast pressure plus 2
When it is smaller than σ ... Good (2 '): Information regarding unloading is also classified into the following three categories: bad, slightly bad, and good.

○ When the size of slip (15) is 3 m or more (slip occurred), or when the unilateral index (2) is larger than plus 15 (hanging on the shelf), or the current sparseness index (16) is the last 10 When the average charge is greater than 3σ ... Bad ○ The current sparseness index is the average of the last 10 charges plus 2σ
When it is larger and less than plus 3σ of the average of 10 charges recently ... Slightly bad ○ The current sparseness index is minus 2 of the average of recent 10 charges
When it is larger than σ and smaller than the average plus 2σ of 10 charges recently ... Good (3 '): Similarly, regarding the information of the top gas temperature (3) regarding the gas flow, the current top temperature is the last 3 hours. When it is larger than the average plus 3σ ... Bad ○ When the current furnace top temperature is smaller than the plus 3σ of the last 3 hours average and is larger than plus 2σ of the last 3 hours average… Somewhat bad ○ The current top temperature is the latest 3 When it is smaller than the time average plus 2σ and larger than the last 3 hours average plus 2σ ... Good (4 '): The same applies to (4) ηco information regarding the gas flow.

(5 '): Similarly, the same applies to (5) shaft pressure information regarding gas flow. However, the shaft pressure is determined by the pressure difference between the upper and lower stages.

(6 '): Similarly regarding gas flow (9) Top gas N ₂
The same applies to the concentration of.

(7 '): (6) Skin flow temperature (SKT) related to furnace temperature, (7) Stave heat load (STL), (8)
Regarding the stave temperature (STT), the difference between the average value of 1 to 2 hours ago and the average value of the latest 30 is ΔSKT, ΔST.
L, ΔSTT are calculated respectively, and ΔSTL> 1 million Kcal or ΔSTL> 100 ° C. or ΔSKT> 50 ° C. or STL> 1 million Kcal or STT> 400 ° C. or SKT> 800 ° C .... Bad ○ ΔSKL, ΔSTT, ΔSKT 1 to 2 for each
Within 2σ with respect to σ before time, and STL <1000
10,000 Kcal and STT <400 ° C and SKT <800 ° C …… Good ○ Other than the above …… Slightly bad (8 ′): The following processing is performed based on the information on the remaining slag inside the furnace. Classify into three stages.

○ The time (14) from tapping the tap currently being tapped to tapping is 90 minutes or more, or the residue level (= (1)
Actual slag amount- (10) Calculated slag amount) is 200 tons or more …… Bad ○ 70 minutes to 90 minutes or the same as above, when the level is 150 tons or more …… Slightly bad ○ Other than the above (9 '): For information regarding the shortage of furnace heat, processing is performed based on (13) tap representative hot metal (HMT) temperature as follows.

○ When HMT <1470 has lasted two taps in the past …… Bad ○ 1470 <HMT <1480 is one of the last two taps
When tapped out ... A little bad ○ When both recent 2 taps are 1480 <HMT ... Good (10 '): Regarding (12) information on the number of times of charging, the following is the actual number of times of charging of the last 8H and theory. Processing is performed as follows based on the number of times of charging.

○ Current actual charging times-theoretical charging times are more than average + 3σ of difference in the last 8 hours …… Bad ○ Current actual charging times-average differences of theoretical charging times are more than ± 2σ in the last 8 hours When ……… Good ○ Other than above… Slightly bad (11 ′) For information on hot metal balance, perform the following processing based on the number of empty furnace topped cars in (17) above and (18) hot metal inventory To do.

○ When the number of empty trucks is <2 or when the hot metal inventory is <3000 tons …… Bad ○ When the number of empty trucks is> 4 or when the hot metal inventory is <2000 tons …… Good ○ Other than the above …… Slightly bad Although an example of the processing has been described, it can be summarized as shown in Table 1 that pattern information can be processed such that a blank in the table is filled in one place in each row.

The processing method shown here is an example, and any information used by the operator to observe and determine the air permeability may be used. Raw data before classifying into 3 stages is also acceptable.

It is necessary to classify into three or more stages for processing, and set the boundary value with the largest change amount of the boundary value to a value at which any operator takes an action.

Next, another embodiment of the present invention will be described with reference to FIGS. FIG. 11 has an expert system 22 in parallel with the neural network 11 and a judging means 21 for judging which of them is to be used.

In the above example, the part classified as "bad" is a situation in which any operator determines that the operation is bad, and is an area in which the action to be taken is clear. If any data enters the area judged to be "bad", the amount of action is set according to the expert system 22 that has the prepared "if ... if" type rules as the knowledge base. .

For example, if (IF): If the blast pressure is bad, in this case (THE
N): Reduce the air flow by 100 Nm ³ / mm If (IF): If the unloading is bad, in this case (THEN):
Blowing amount 100 Nm ³ / mm Wakashi lower (IF): the air blowing pressure is poor, and bad load down, if, in this case (THEN): air volume of 200 Nm ³ / mm Wakashi lower (IF): residue level Is bad, in this case (THE
N): Reduce the air flow rate by 100 Nm ³ / mm If (IF): Shaft pressure is bad and N ₂ is bad or η _co is bad, in this case (THEN): Air flow rate is reduced by 100 Nm ³ / mm And it's very easy to decide the rules,
The boundary value of the data judged as "bad" is almost unchanged, the maintenance is easy, and the rule outlook is easy.

On the other hand, it is very difficult if no instrumentation data is judged to be "bad" and only "slightly bad" or "good". It became very complicated, and maintenance and revision of rules became almost impossible, and the accuracy of action judgment as an output was poor.

In this embodiment, the neural network is used to determine the action in such a case. If there is even one item judged as "bad", the expert system can take action, so in the network of FIG. 12, only fine adjustment wind reduction is necessary.
In the output layer, there are two levels: 0 (maintain the current state) and -100 Nm ³ / min.

Two types of input layers of "slightly bad" and "good" among the three-step judgment of the processed data of (1 ') to (11') are provided. Here again, the input layer has a value of either 0 or 1. Each is a through unit.

The output layer, the furnace situation is very bad state, already because it dealt with "Wakashi ... if ... the" type of rule, 0 the extent of the wind reduction (as specified air volume maintenance), - 100Nm ³ / minute Use only 2 units divided into 2 stages. The output of the output layer unit is 0 or 1.

Next, the layer structure of the neural network will be described.

Although the neural network is linear in the two-layer structure, it becomes non-linear in the three-layer structure or more, and the information processing capability is significantly improved. Therefore, in general, a network having three or more layers with an intermediate layer is used. In this embodiment, one hidden layer is provided to have a three-layer structure. The number of hidden layer units was 5.

Since the output function of the output layer unit is such that the output of the output layer is 1 or 0, and only one of the output layers is set to 1, the threshold value as shown in FIG. It was a function.

Considering each unit as the i-th unit of the k-th layer as shown in FIG. 4, first, the output of each unit of the (k-1) -th layer, To each weight (mutual coupling coefficient) Sum of things multiplied And take the threshold function The output is the product of.

The output of the k-th layer i-th unit is given by the following equation.

Note that θ is a threshold value. A unit that always takes 1 is provided on the input side, and its weighting coefficient is set to −θ so that it is reflected in the equation (11), and the threshold value can also be corrected by learning.

When the data processed as described above is given to the input layer, each layer unit sequentially calculates the output from the lower layer to the upper layer in accordance with the equation (13).

The output result which looked at the output of the output layer in time series
Shown in the figure. The wind reduction amount previously assigned to the unit in which 1 stands is output to the process computer 6. In this way, automatic control when the wind is reduced becomes possible.

In such a neural network, it is necessary to determine the mutual coupling coefficient in advance. The learning method is the backpropagation method of Ramelhart et al.

Learning is performed by a skilled operator as a teacher. That is, a certain pattern (reference input) having an input layer and a teacher output are given, and the forward propagation and the backward propagation are repeatedly calculated a plurality of times to converge.

There is also a method in which the weighting coefficient of the initial value is given once by the reference input and the teacher output of some patterns, and fine correction is performed online.

Although the embodiment in the case of reducing the wind has been described above, the action of reducing the wind once and then recovering it by the increase in wind can be similarly controlled by using the neural network. However, it is different from the processing part of the data, and it is necessary to be able to output the determination result that the processing is improved or deteriorated.

FIG. 8 shows an example of operation results when the air flow rate control method of the present invention is applied. Action instructions are given accurately. In addition, Fig. 8, Fig. 9 and Fig. 10 show the trends of the fluctuation amount of Si in the pig iron and the fluctuation amount of the charging material charging interval, respectively, but both are decreasing, and stable operation of the blast furnace is shown. It became possible.

〔The invention's effect〕

According to the present invention, the action amount and timing of the air flow reduction and the air flow increase in the air flow rate control are more appropriate, and the air permeability control accuracy is improved. In addition, the load on the operator has been dramatically reduced by automating what was conventionally operated by the operator.

[Brief description of drawings]

FIG. 1 is a block diagram of an apparatus for implementing the method of the present invention, FIG. 2 is a flow chart of a neural network of an embodiment of the present invention,
FIG. 3 is a graph of the unit internal function of the neural network in the embodiment, FIG. 4 is a flow chart showing the input / output relation of the neural network in the embodiment, and FIG.
FIG. 6 is a chart showing an output example of the neural network in the embodiment, FIG. 6 (a) is a schematic system diagram of the neutral network, (b) and (c) are graphs showing a functional form, and FIG. 7 is a neural network. FIG. 8 is a flow chart showing the operation of the network, FIG. 8 is a graph showing an example of operation results, FIG. 9 is a graph showing the trend of the fluctuation amount of Si in the pig, and FIG. 10 is a graph showing the fluctuation amount of the charging material charging interval. Fig. 11 is a graph showing a trend, Fig. 11 is a block diagram of another embodiment of the present invention, and Fig. 12 is a flow chart showing the input / output relationship of the neural network. 1 ... Blast furnace, 2 ... Hot air furnace, 3 Blower blower, 4 Blower flow meter, 5 Blast flow controller, 6 Blast furnace process computer, 7 Special computer, 8 Blast furnace detection end, 9 ... Data input means 10 ... Data processing means 11 ... Neural network 22 ... Expert system

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Toshiro Sawada 1 Kawasaki-cho, Chiba-shi, Chiba Inside Kawasaki Steel Co., Ltd. (72) Inventor Keiji Kobayashi 1 Kawasaki-cho, Chiba-shi Chiba Steel Co., Ltd. Chiba Steel In-house

Claims (2)

[Claims] 1. A blower flow rate is temporarily reduced from a specified value in order to control ventilation of a blast furnace to avoid abnormal states such as slipping, hanging, and blow-through, and from there, the wind is increased again to a specified air volume. In the return air flow rate control method, a neural network of two or more layers is constructed in which various data are used as input layers, and the output layer is the output of each unit that is divided into a plurality of levels of wind reduction or increase in volume. The network is pre-learned to determine the optimum mutual coupling coefficient, and the sensor data for determining the ventilation state installed in the blast furnace and the operation data calculated based on the sensor data are input at a predetermined timing. And process it into processing data that matches the operator's judgment criteria so that you can understand what kind of change each of these shows based on the data. The management, on the basis of the processing data, the use of neural networks to determine the wind reduction amount or increase air volume in real time, automatically blast furnace blower flow control method and controls with that value. 2. The air flow rate for controlling ventilation of the blast furnace is temporarily reduced from a specified value to avoid abnormal conditions such as slipping, hanging, and blow-through, and from there, the air is increased again to a specified air volume. In the return air flow rate control method, a neural network of two or more layers is constructed in which various data are used as input layers, and the output layer is the output of each unit that is divided into a plurality of levels of wind reduction or increase in volume. The network is pre-learned to determine the optimum mutual coupling coefficient.On the other hand, an expert system is constructed in parallel with the neural network, and the sensor data for determining the ventilation state installed in the blast furnace and the The operation data calculated by the above is input to the data input means at a predetermined timing, and based on the data, the recent changes If the judgment means judges that the relationship between the movement of each measurement data and the increase / decrease in air volume is clear from the processed data, the expert system is processed. Input the processed data into, and if it is not clear, enter the processed data into the neural network,
A blast furnace blast flow rate control method characterized in that an expert system or a neural network determines the amount of reduced airflow or the amount of increased airflow in real time based on the processing data, and automatically controls with that value.