JP2982174B2 - Blast furnace operation support method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a blast furnace operation support method, and more particularly to learning control of a blast furnace operation state.

[Problems to be Solved by Conventional Technique and Invention] There are the following methods as conventional learning control methods.

(1) Learning in a knowledge system This learning method is symbol processing, and is not suitable for data processing. Although conversion from data to symbols is possible, it is still at the level of sophistication of knowledge, and it is difficult to acquire new knowledge.

(2) Learning in a fuzzy system This learning method is merely a tuning of a membership function, and it is difficult to acquire new knowledge (the same applies to techniques such as adaptive control).

(3) Learning in Neural Network This learning method is effective for pattern recognition of operation, but it is difficult to accumulate operation data according to the learning result because learning depends on the coupling coefficient between units.
In addition, it is difficult to recognize an operation state in which the learning speed is low and the occurrence frequency is low, and it was not possible to grasp past cases.

The present invention has been made in order to solve such problems, and it is extremely easy to acquire new knowledge. Further, automatic classification of operation data, accumulation of operation characteristic extraction data, and automatic retrieval of accumulated operation data It is another object of the present invention to provide a blast furnace operation support method which enables guidance of an operation amount based on past accumulated data.

Means for Solving the Problems A blast furnace operation support method according to the present invention scans a sensor attached to a blast furnace and collects main raw data including a temperature, a gas component, and a shaft temperature at each point of the furnace body. And also
Operational data including blast furnace air flow rate, blast temperature, sinter ratio of small grains, coke ratio and raw material charging conditions are collected, and based on the main raw data, the temporal and spatial conditions in the blast furnace are analyzed. Detects abnormal furnace conditions, detects abnormal furnace conditions, detects the presence or absence of an operation change based on the operation specifications, and detects the changed items.Then, the main raw data, operation data including the above-mentioned characteristic amounts and operation specifications The process of classifying into a class having a plurality of classes according to the presence or absence of abnormal furnace conditions, the presence or absence of an operation change, and the items to be changed, the latest operation data, and under the same classification as the latest operation data, A multidimensional scale with each class to which the plurality of operation data belongs is obtained based on the feature amount, and when the multidimensional scale of the latest operation data is smaller than the multidimensional scale between the classes, the latest Operation data Between the operation data belonging to the class having the smallest multidimensional scale and updating the operation data of the blast furnace of the class; and obtaining the multidimensional scale of the operation data of the current operation and each class to the current operation. Identifying a close class, and extracting and outputting operation data in the class as a case in the same manner.

The multidimensional scale (MDS) refers to “the overlap between the spatial coordinates of an operation data group representing a certain operation and the spatial coordinates representing another operation,” as described in an example described later. Mitsuki distance. More specifically, for the current operation data group, it is determined to which state (class) they belong. For example, the state of the blast furnace is normal, shelf hanging occurs, If it is desired to determine which of the states belongs to excessive heat or the like, the similarity or closeness to the data group (class) constituting these states is obtained as a scale.

Also, replace the latest operation data with the corresponding oldest data in the same class.

Also, when the multi-dimensional scale based on the latest operation data is larger than the multi-dimensional scale between classes, a new class is created based on the latest operation data, and the one with the smallest multi-dimensional scale between classes is selected. Both classes are combined into one class, and the number of classes is fixed.

[Operation] In the present invention, the main raw data and the operation data are collected, the characteristic amount in the blast furnace is extracted based on the main raw data, the abnormal furnace condition is detected, and the operation change is performed based on the operation data. Is detected, and the operation data including the main raw data, the characteristic quantity, and the operation data is classified according to the presence or absence of the abnormal furnace condition, the presence or absence of the operation change, and the change item. Then, the latest operation data, under the same classification as the latest operation data, a multidimensional scale of each class to which a plurality of operation data belongs is obtained based on the feature amount, and the latest When the multidimensional scale of the operation data is smaller than the multidimensional scale between the classes, the operation data is assigned to the class having the smallest multidimensional scale with the latest operation data.

The latest operation data is replaced with the oldest data of the corresponding class.

When the multidimensional scale of the latest operation data is larger than that between the classes, a new class is created, while the two classes at the position where the multidimensional scale is the smallest are combined into one class.

In this way, the operation data of each class is constantly updated with new operation data, a similar case is extracted from a class close to the operation data of the current operation, the operation amount is determined based on the case, and displayed on a display device, for example. Guidance to workers.

[Embodiment] FIG. 2 is a block diagram of a control device for implementing a blast furnace operation support method according to an embodiment of the present invention and its related equipment. In the figure, (1) is the blast furnace to be controlled, (1)
0) is a blast furnace control device. This blast furnace control device (10)
Means a data scanner (11) (11a, 11b,... 11n), a sensor data collecting means (12), a CPU (13) as an arithmetic unit, C
A system ROM (14) that stores a program that controls the operation of the PU (13), a RAM (15) that stores various data,
(16) and CRT (17). Reference numeral (20) denotes a drive control device, which adjusts a calculation result of the blast furnace control device (10) or hot air supplied to the blast furnace (1) by manual operation of an operator.

FIG. 1 is a flowchart showing a blast furnace operation support method in the blast furnace control device of FIG. Hereinafter, a description will be given of an example of the prediction of the abnormal furnace condition of the blast furnace and the guidance of the operation amount corresponding thereto.

A. Reading of operation data Each sensor is scanned by the data scanner (11) and various measurement data is taken in by the sensor data collection means (12). In this type of blast furnace, about 300 pieces of sensor information are taken in at a fixed cycle and stored in the RAM (15).

B. Operation feature extraction Next, the operation state features are extracted from various measurement data stored in the RAM (15). When extracting the characteristics of the operation state, the characteristics of the sensor information, the detection of the abnormal furnace condition, and the detection of the operation change are performed.

(B1) The feature extraction of sensor information extracts features in terms of time (change, variation) and space (each direction of furnace height, circumference and radius) based on the experience of the operator.

(B2) Abnormal furnace conditions (blow-through, wall drop, slip) are detected by the furnace pressure for blow-through, shaft temperature or CS body temperature for wall drop, and furnace mouth sounding sensor information for slip. I do.

(B3) The detection of the operation change is to detect the change of the operation amount performed by the operator, and the operation amount is defined as the blast flow rate, the blast temperature, the blast humidity, the ratio of the small sinter ore, the coke ratio, Raw material charging conditions (number of rotations, rotation angle) and the like.

C. Recognition (classification) of operations Here, classification of abnormal operations, discrimination, and recognition based on operation distance space are performed.

The distance from the past operation example is calculated by the following equation. The target of the calculation is the change of the operation amount (item and quantity), abnormal furnace condition (item and degree), operation condition (load drop condition, circumferential balance condition, gas condition, furnace heat level, residual iron condition, etc.) , Operation specifications (input and output to and from the blast furnace), and the like.

 Find min (dj).

In calculating the distance (dj), the following equation may be used instead of the equation (1). The calculation based on the following equation focuses on only the changed items, and can treat the items that do not originally affect the operation change or the like as zero, so that there is an advantage that disturbance from the unaffected items can be removed.

 Find max (dj).

In this feature extraction value, “−1” represents belonging to a contradictory set. For example, if a good operation is “1”, a bad operation is “−1”. “0” indicates unknown or no correlation.

D. Reliability evaluation of operation recognition Check how much the guidance content of this system up to the last time matches the current operation status. The operation is the same as the above-mentioned C term. The degree of coincidence includes the operation amount guidance up to the previous time and the current degree of coincidence (α ₁ ), the degree of coincidence between the abnormal furnace state prediction up to the previous time and the present time (α ₂ ), and the predicted operation state accompanying a change in the operation amount. There is a current degree of coincidence (α ₃ ), and the degree of coincidence is determined each time and is filed in the RAM (15) as a past operation database. Then, the recognition (affiliation class) up to the previous time is corrected so that the results with respect to the estimation match, and the data is reconstructed based on the recognition reliability evaluation. For example, when the degree of coincidence is small, the weight coefficient of the equation (1) or (2) is changed, or the present case is placed in a different class.

E. Accumulation of operation data (acquisition of knowledge) Operation data is acquired by the following methods (1) and (2). Since the storage capacity of the computer is limited, the following method is used to expand knowledge as the user experiences operation.

(1) Calculation distance dj <distance between all classes; In this case, operation data having a small difference from the conventional experience is replaced with similar old data.

FIG. 3A is an explanatory diagram showing a method of accumulating operation data when the calculated distance (dj) is smaller than the distance between all classes. In the figure, A, B, and C are classes, respectively. Each class is composed of barycentric coordinates G (marked with ●), the number of belonging knowledge, the distance between knowledges (marked with x) in the class, the distance to other classes, and the like.

When the latest knowledge (operation data) N is input, the distances d1, d2, and d3 between the knowledge N and each class are obtained. When the distance is shorter than the distance between the centers of gravity of each class,
The knowledge N and the class belong to the class having the shortest distance. In the illustrated embodiment, knowledge N belongs to class A. Then, for example, the class A
When the number of knowledges in the class exceeds a predetermined number, the oldest knowledge in the class A is deleted. In this way, the knowledge in the class is updated with the latest knowledge.

(2) Calculated distance dj ≧ distance of all classes; In this case, a new class is set as the operation data that has a large difference from the previous experience, and the old class with a small difference in experience (distance) is used. Aggregate (degenerate). This indicates that the distance lg between classes increases, which corresponds to an increase in knowledge.

FIG. 3B is an explanatory diagram showing a method of accumulating operation data when the calculated distance (dj) is larger than the distance lg between all classes. When the latest knowledge (operation data) N is input,
As in the case described above, the distances d1, d2, and d3 between the knowledge N and each of the classes A, B, and C are obtained. If the distances are all greater than the distance between the classes, the new class is determined. Create D. Then, a new class is created by aggregating classes whose distances are short. In this embodiment, classes B and C are aggregated to create a new class B '. Also in this new class B ', when the number of knowledges exceeds a predetermined number, old knowledges are deleted in order.

FIG. 4 is an explanatory diagram of an image of a database of past operation data stored in the RAM (15). The presence / absence of a furnace condition abnormality, the presence / absence of an operation change corresponding to the presence / absence of a furnace condition, and change items in the presence of an operation change are respectively filed and discriminated by flags. And
Each change item is classified into classes 1 to m, and each class is classified into operation data 1 to operation data n.

Each operation data is classified into operation specifications, feature extraction data, main raw data, operation amount change history, abnormal furnace condition occurrence history, operation date, and the like. The operating specifications are: blending (SR ratio, coke ratio, small SR ratio), air flow (flow rate, temperature, humidity),
And charging (swirl mode) and tapping (component, temperature) data. Feature extraction data includes circumferential balance state, charge state, gas flow state, pressure distribution state, furnace heat level, gas utilization rate, shaft efficiency, sol loss C amount, tuyere tip temperature, residual iron amount state, blow-through, slip , And data such as wall falling.

The main raw data is composed of data such as the temperature, gas components, and the Schft pressure at each point of the furnace body. Operation change history
It is composed of a pair of data groups consisting of the time when the operation was changed and the items changed. The abnormal furnace condition occurrence history is composed of a pair of data groups including the time when the abnormal furnace condition occurred and the contents thereof.

Each of the above operation data is filed in chronological order,
When calculating the distance according to the above equation (1) or (2), the average value is calculated. Further, as described later, when the data is displayed as similar data in the past, the data is output as it is.

F. Prediction of operating state In order to predict abnormal reactor conditions and provide guidance on the amount of operation, the operating state is predicted in advance.

(1) A state change accompanying a change in the operation amount (particularly when the guidance is not followed) (a state change is obtained by a linear operation on the change amount).

(2) State change due to abnormal furnace condition (calculation is the same as (1) above).

G. Reference of similar example to prediction In the state of prediction (1 hour after) from the past (for example, 8 hours before), a plurality of past operation cases are extracted.

Similar cases are extracted for the operation status only, and also for changes in the operation amount and abnormal furnace conditions, if any.

H. Abnormal reactor condition guidance Even if there is no abnormal reactor condition in the current operation, if there is an abnormal reactor condition in the same case, an alarm is output as abnormal reactor condition prediction.

This warning is the result of comprehensive evaluation based on the reliability of the case (plurality) and the D term. The comprehensive evaluation is performed by a weighted numerical sum for each affiliation degree (by fuzzy processing).

If the current operation involves abnormal furnace conditions or changes in the amount of operation, guidance will be provided in accordance with similar cases including these.

Misoperation, special abnormal furnace condition, etc. have few cases in the same manner.

I. MV guidance Basically the same as H. In a similar case in the past, a gain between the operation amount and the state amount is statistically obtained, and the operation amount is determined so as to be a target value (set value or current state maintenance). In addition, an output pattern of the operation amount is determined from a time delay of the state amount with respect to each operation amount set in advance and a control cycle of the present system (for example, a finite setter).

FIG. 5 is an explanatory diagram in the above-mentioned guidance, and shows a case where a similar past case is read out when the current data changes significantly. As a past case, a blow-by occurred when the data changed, and wind was reduced (300 Nm ³ ) as a measure against it, and the data was restored to a normal value by the processing. Therefore, by referring to this past case, it is possible to guess the next process to be performed by the operator.

Next, an outline of the actual operation of the learning control method in this embodiment will be described.

(1) Steady-time operation; (1-1) Detection of abnormal furnace condition; Slip, blow-through and wall drop are checked every minute. When these abnormalities occur, the message is displayed on CRT (17) and Give voice guidance. Then, by operating the screen key and the function key, the base data of the abnormality detection is displayed as shown in FIG.
That is, the present main change data is displayed, and a similar past case is displayed. When a slip or the like occurs, the display is displayed.

(1-2) Confirmation of operation change; Check the blast temperature, humidity, volume, furnace top pressure every minute,
The number of turns and the turning angle are checked for each dump, and the coke ratio, PC ratio, SR compounding ratio, and small particles for each charge (cycle)
Check the SR mixing ratio. Of course, the operation itself of the operation change is performed by the operator, and the change is automatically detected. At this time, when a change in operation is detected, the fact is displayed on the CRT (17).

(1-3) Extraction of operation characteristics: The above-mentioned spatial characteristics and temporal characteristics are extracted every 30 minutes. Then, in the operation feature extraction, the presence / absence of a furnace condition abnormality, the presence / absence of an operation change, and the feature of the operation state are extracted. As the operation state, operation specifications (; blast furnace input / output amount), charging state, gas flow state... (; Blast furnace state amount) are extracted. Then, a similar operation in the past is extracted by performing pattern matching, and a prediction of an abnormal furnace condition and a calculation of an operation amount are performed based on a past example. If it is necessary to change the manipulated variable, a message to that effect is displayed on the CRT (17),
Guided by voice. Thereafter, operating the screen key and the function key to display the present data and the past similar cases on the CRT is the same as the above-described case.

(2) At the time of request (2-1) Operation amount guide; By operating the screen key and the function key, the current data and the past control amounts (plural) and operation amounts (plural) for similar cases in the past can be obtained. Display as a function of time.

FIG. 6 is an explanatory diagram showing the display at this time. The display data can be automatically scaled, and the time coordinate can be manually shifted and automatically shifted. On the right side of the CRT screen, the characteristics of current data and past data information are displayed.

The characteristics of the current data include spatial features, temporal features,
The characteristics of the operation, past actions, etc. are displayed. As past data, past data information (time, control amount, operation amount, other specifications, ...), change of operation amount (item, operation amount), reason for selecting past data (important items, close classes ,...), Determination of operation amount (past data, identification, target operation), knowledge (knowledge at abnormal time, knowledge at stable time) and the like are displayed.

(2-2) Confirmation of learning data; Operating the screen key and the function key causes the learning data to be displayed on the CRT (17). FIG. 7 is an explanatory diagram showing the display at this time. The data to be displayed here is data (average value, such as class number, presence or absence of abnormal furnace condition, presence or absence of operation change, its items, operation specifications, etc.). ), Operation data (accumulated) number, operation data index, date and time, and the like. Then, a key operation is performed in accordance with whether the furnace condition is good or bad and whether raw data is present or not, thereby displaying a time chart of the current data and the past data. Further, when the time chart is displayed, a list of display items or the like can be displayed, or by changing the display items, a time chart having contents corresponding thereto can be displayed.

[Effects of the Invention] As described above, according to the present invention, new measurement data is automatically classified and accumulated in the corresponding class, replaced with the oldest data in the class, and the measurement data is separated from each class. Creates a new class, while two classes in close proximity are combined into one class. For this reason, class operating data adapts to plant changes and, over time, operating experience (knowledge)
Increases. In other words, the operating state can be automatically learned by the hand, and therefore, accurate guidance of the operation amount can be provided based on the similar result in the past, which is the learning result.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart showing a blast furnace operation support method according to an embodiment of the present invention, and FIG. 2 is a block diagram of a control device for implementing the blast furnace operation support method and related equipment. FIGS. 3 (a) and 3 (b) are explanatory diagrams for acquiring new data, FIG. 4 is an explanatory diagram showing the configuration of a past operation database, FIG. 5 is an explanatory diagram of operation guidance, FIG. FIG. 7 and FIG. 7 are explanatory diagrams of display examples of the CRT in confirming the operation amount guide and the learning data, respectively.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-59-127105 (JP, A) JP-A-2-76033 (JP, A) Toshiro Terano, two others, "Introduction to Applied Fuzzy Systems," First Edition , Ohmsha, May 25, 1989, pp. 56-69, section 3.2 "Blast Furnace Control" Takayuki Saito, "Multidimensional Scaling Method", Measurement and Control, Society of Instrument and Control Engineers, 1983 Month, Volume 22, Issue 1, P126-131 (58) Fields investigated (Int. Cl. ⁶ , DB name) G05B 13/00-13/04 G06F 15/18 560 G06F 9/44 580 JICST file ( JOIS)

Claims (3)

(57) [Claims] 1. Scanning a sensor attached to a blast furnace,
The main raw data including the temperature of each point of the furnace body, the gas components and the shaft temperature were collected, and the blast furnace blast flow rate, blast temperature,
The operation data including the sinter ratio of small grains, the coke ratio and the raw material charging conditions are collected, and based on the main raw data, feature amounts for the temporal and spatial conditions in the blast furnace are extracted. The abnormal furnace condition is detected, the presence or absence of an operation change and the change item are detected based on the operation data, and the operation data including the main raw data, the characteristic amount and the operation data are converted to the abnormal furnace condition. A process of classifying each class into classes having a plurality of classes according to the presence / absence, presence / absence of operation change, and items to be changed, and the latest operation data and a plurality of the operation data under the same classification as the latest operation data. A multidimensional scale with each class to which the class belongs to is determined based on the feature quantity, and when the multidimensional scale of the latest operation data is smaller than the multidimensional scale between the classes, the multidimensional scale between the latest operation data and Most Updating the blast furnace operation data of the class by assigning the operation data to a class having a small dimensional scale; and identifying a class close to the current operation by obtaining a multidimensional scale of the operation data of the current operation and each class. A process for extracting and outputting operation data in the class as a similar case. 2. The method according to claim 1, wherein the latest operation data is replaced with the corresponding oldest data in the same class.
Blast furnace operation support method as described. 3. When the multidimensional scale based on the latest operation data is larger than the multidimensional scale between classes, a new class is created based on the latest operation data, and the one with the smallest multidimensional scale between classes is selected. The blast furnace operation support method according to claim 1 or 2, wherein both classes are combined into one class, and the number of classes is fixed.