Application of Expert System for Blast Furnace Operation Control

Application of Expert System for Blast Furnace Operation Control

Copyright © IFAC 10th Triennial World Con!!;ress . Munich. FRG. 198 7 APPLICATION OF EXPERT SYSTEM FOR BLAST FURNACE OPERATION CONTROL M. Shibata, K...

1MB Sizes 12 Downloads 211 Views

Copyright © IFAC 10th Triennial World Con!!;ress . Munich. FRG. 198 7

APPLICATION OF EXPERT SYSTEM FOR BLAST FURNACE OPERATION CONTROL M. Shibata, K. Hashimoto and K. Wakimoto Fukuyama Works , Nippon Kokan K. K. . 1 Kokan -cho , Fukuvama City, Japan

Abstract. Ve have applied the technology of expert systel which is based on Knowledge Engineering to the furnace condition diagnosis systel, which was one of the lost ilPortant systels of the operation control systel on the blast furnace. This expert systel is constructed by two systels, the first is the 'Abnorlal Furnace Condition Prediction SYStel', the second is 'Furnace Heat Monitoring and Control Systel'. In these systel we used the operator's knowledge and experience, and a large quantity of sensor data that are gathered by an on-line process cOIPuter. The ilPortant character of this systel is that it is a real-tile expert systel. Ve developed and applied this systel to the relined No.S Blast Furnace in Fukuyala Vorks and have achived good results. Kevwords. Expert systel; artificial intelligence; cOIPuter control; steel industry; sensors

knowledge

base;

INTRODUCTION On the iron and steel laking process ,the new electronics and cOIPuter technology have been introduced early with the advancelent of these techniques. and the autolatic control and the systelatization laking use of cOlputers have been progressed on every field of the works . And nowadays. needs for high technology have been required lore and lore strongly because of the decrease of the steel deland and the request for the high quality of the steel products. In this situation, the knowledge engineering technology that is an application fields of Artificial Intelligence has been focused on because it could be used as the new systel technique for the saving lan power or as the new control technique for the ilProvelent of the equiPlent and operations. This tile we would like to introduce the outline of the on-line real-tile expert systel which was developed by NKK as the Blast Furnace (B.F.) operation supporting systel. This systel has been applied to the No.S B.F. in the Fukuyala Vorks which was refired in February 1986. and has achieved good results. BACKGROUND OF THE DEVELOPMENT EXPERT SYSTEM

OF

THE

B. F. operation is the process which lakes pig iron by reducing iron ore with charging coke frol the top and blowing blast fro. the bottol. It is the first process of the works, and it is lost ilPortant because it supplies high quality pig iron to the next stage in the steel laking process.

For this reason lany hundreds of sensors are installed to observe the furnace condition and, fro I the data, we have constructed the large operation lanagelent systel using lathelatical lodels and statistical processing techniques for supporting the high level B.F . operation. This lanagelent systel is cOIPosed of the following three functions as illustrated in Fig. 1 : 1) data analysis function which gathers and analyzes the sensor data 2) analyzing function of the furnace static condition for getting the lost suitable lethod when the operation restriction changes greatly 3) diagnostic function of the furnace condition But the actual reaction in the furnace is very cOIPlicated, so SOle operation depend on the experienced operator's knowledge especially during an abnorlal condition of the furnace which disturbs the furnace operations seriously. For this background. we planned to introduce the knowledge engineering technology which is a relarkable new cOIPuter technique because it lakes it possible to use hUlan knowledge and experience effectively. Ve have constructed the 'Furnace Condition Diagnosis Expert Systel' which supports the B.F. operation by on-line real-tile . This systel has lainly the following three ails: 1) to realize the highly accurate systel by appling the Artificial Intelligence(AI) technology 2) to standardize and inherit the B.F . operation technique 3) to deal flexibly with the frequent delands for the systel lodifications

28i

288

M. Shibata, K. Hashimoto and K. Wakimoto

(Phenomena (Noise)

(Control foctor>

H

rri

OI,lrlbulion of

In-furnace lop

molerlal

Row

In blost furnace>

materials property

~

I Burden condition

11

(Operation conlrol system)

(Sen.sor dolo)

Burden descendln0rvelocity

r-

• Throat 90S temperature

4'

IChonqtl of I permeability

Scaffold

1Chonoe of

the furnace

90S

Furnace heol monitor

• Shoft

-Ino and conlrol

temperature

,I flow

• Top 00' pressuret• Shaft pressure • Bios t pressure

l

Chonoe

'Blo,1 volum. ·Blosl molslurt

'Blo,1 l-.mperolure ·Oxyo.n volume ·Fuel volume

--l TopplnQ

m.lhod

.}-

i'

nl

of reduction

I

• Top 90S temperature • Top oos contents

ChonQe of

Coke proper lies

• SloO and hot metal residual condition

system

11

system

. Inner profile of

'"-

dlo~ : losls

tExpert system) Abnormal condltlon prediction system

-

• Skin flow temperature

Furnace condition

~

flow reslslance

I

of heQI condllton

~ f-

I

• Tuyere temperature

I

• Conlents of 'IOO~

I

Chonoe

I-

ond hot metol

In-furnace condition onolysls (Mathematical model) -Burden distribution slmulollon model

I Tolo! JudQInQI

I

loperallon

Instructions

oRis' model (Operation dlooroml 'Shope ond poslllon of cohesive zone model

j

I I

Control

Data anal sls system ·Collecllon of operalion data '0010 analysis

• Hol metol temperature

I" i g.!. Concept of operation control system

OUTLINE OF THE EXPERT SYSTEM Funct i on aruL Structure

~

Preprocessino fOI Inference (ProcedlXol colculollon melhod)

t..hJL System

This system is composed of two systems. The first is the 'Abnormal Furnace Condition Prediction System (AFS)' which predicts the occurrence of burden slip and channeling in the furnace. The second is the 'Furnace Heat Monitoring and Control System (HCS)' which judges the in-furnace heat level and instructs operators of the proper method of the operation. This is an observational and controltype expert system, and has some characteristics for example 1) it handles a lot of time series data 2) a real-time process is needed in it. To realize this, we separate the functions of the system into two parts. (see Fig. 2) One is the preprocessing part for inference that is executed on the process computer using the conventional method. Another is the inference part which is performed on the AI processor using knowledge engineering technology. The for.er part has the functions of gathering sensor data and registering them in the time series data base, calculating the process of making the fact data for inference, and displaying the results of the inference. The latter part has the functions of inferring the furnace condition by the fact data received fro. the former part and the rules in the Know I edge Base (KB) . And we use a software tool for constructing an expert system in the inference part. This tool is based on LISP language and supports various knowledge representation methods such as 'production rule', 'frame-based model and 'black-board model

L:::::="+';======-____J

(Process computer)

Inference for 10furnace condition

(AI -meThod) A.I processor

;~:~~e r:~v~~~p~~trt

Knowled(Je

syslem

(Experl SHELL)

Fig.2. SYstem configuration of Furnace Condition Diagnosis Expert System

Knowledge t..hJL ~

Representat i on aruL Structure

~

The domain-specific knowledge and heuristic rules of thumb acquired by the blast furnace engineers and operators are stored in the KB of the AI processor. Basically we use the 'production rule' for the knowledge representation in this syste., and we construct it as the analyzing type expert system which selects one of the most suitable hypotheses that are arranged beforehand by analyzing the given data. We

Blast Furnace Ope ration Control

think it is a proper aethod . Our aajor probleas are to judge the furnace condition by the sensor inforaation and after that to predict the occurrence of the abnoraal furnace condition and to decide the suitable operation action. Besides it we aake the best of the 'black-board aodel' for aeaorizing the tiae series data as results and line of reasoning , and the inforaation transfer between the knowledge units(KU) described after. And the ' fraaebased aodel' is used in HCS for representing the static knowledge concerning the teaperature and pressure of the various parts in the furnace . The KB is divided into the several KUs which are coaposed of soae rule sets, by the attribute of the functions and sensors. (see Fig.3, an exaaple of AFS) And on the whole it takes the fora of a hierarchical construction of the KU . The aias are as follows: 1) Because the expert's knowledge is constructed hierachically,it aakes it easy for us to represent it in the saae style . 2) It aakes clear the rule to collect and add by dividing them to the KU, and we can easily check the validity aaong the rule sets. 3) If we don't divide the rules to the KU ,the inference tiae increases aore and aore according to the nuaber of them. So it enables us to iaprove the efficiency of the reasoning. The followings are exaaples of the actual rules used in AFS: IF "burden descending velocity is slower than XX" THEH "channeling is likely to occur (CF.value X.XX)". IF " the integral calculus of burden descending velocity is smaller than YY" THEH "channeling is likely to occur (CF . value Y. YY)" . We created about 100 rules in the KB of AFS, and about 300 rules in the KB of HCS .

Pattern Recogn i z i ng Process IlL t..h..tL Sensor I2.a1JL The pattern recognition of the tiae series data that has previously been eapirically perforaed by huaan experts was realized on the coaputer. This process is, we think, originally the problea which can be treated by the AI aethod, but there is no proper tool for it because of the enoraasness of the data and the restriction of throughput. So we developed it using the conventional aethod as aatheaatical calculation. This is actually perforaed in the following two stages as the functions of the preprocessing part for inference . The first stage is the process of the saoothing the sensor data. The sensor arrangeaent on B. F. is illustrated in Fig.4. We use over 200 sensor data in this systea . The various sensor data, such as pressure and teaperature in the blast furnace, are affected by disturbances and pseudo periodic changes due to the influence of the charging of raw aaterials . To remove these changes, the saoothing treataent using the statistical aethod has been done (the linear regression process) . The process of the top gas teaperature is shown in Fig.5 as the exaaple of the process. In the regression process , we use H data which is aeasured every 1 ainute, T(t): t=1.2," " /H, and aake fitting to the following linear equation: f(t)=Co+C,t as

ainiaize the value /J

S=

2-

L.lTe (t) -f (t)]

~'1

so decide the coefficient CO,Cl, the present value froa f(H) . It aethod of the least squares.

and get is the

In the secondary stage, we extract the special alternation pattern of the sensor data which causes the change of the furnace condition using the results of the priaary stage. For exaaple, 1) coaparison of inclinations 2) coaparison of the level 3) coaparison of the variance 4) coaparison of the integration 5) forecasting (fitting) the typical change pattern 6) calculation for detecting the change value, are executed. The exaaple is shown in Fig. 6 .

--

-- ---

IF burden descendIng velocit y is low THEN IF va riation of burden descending ve locity is large

THEN ..

The results of inference are displayed on the CRT. AFS indicates the occurrence probability of burden slip and channeling, and HCS shows the operation instruction and CF.value of the heat level and transition . The foraer probability can be given froa CF.value as the following conversion: Probability=f(o<. · CF . value+~,

()

(C<, ~

,t

: fixed value) Fig . 3 . Composition of knowl e dge bas e of AFS

And the explanation of line of reasoning can be displayed at the saae tiae. The

290

M. Shibata, K. Hashimoto and K. Wakimoto

Top 90S composition

~ Top QCs pressure

\1lITD~ ~

Item

Top QCs temperature BUrden level

rf

CD

CD

Shotl temperature Skin flow temperature

@

Throat 90S temp-

®

orature

@ @) @

Pressure los5 Blost pressure

®

Shaft temperature

-~.,,-

Shoft pressure

®

Shoft pressure

Top

(j)

Pressure loss

velocity

pressure

90S

~

1(

Gos utilization ratio

Coollno stove temperature

~'

!!l!L. Exomple

dl

of

Tuyere body temperature

processing

~= dl

Tuyere blost volume

Kb

1

I!

:k:-r

I

AvoroQe' Ps

~"

sensors doto

~~

,Psll1 ~I

Exomple

of

Pressure loss

Shaft pressure

I~Tll

S haft temperature

===l

Burden descending

(f) Skil flow lemperalure

Fixed -type probe I skin flow temperature

I

CD

Burden descendlnQ velocity

Shaft temperature



Ii) Comparison of vorlunce

®

@ Gas ullllzallon rallo

.i,/

11) Comparison of level

S e nsors

Under· beJl probe, throoT 90S temperature ond QOs

I

!I Cdmporbon of nc"ollons

Kb = Pb 2 _P 0 2 Vn

I

when furnace

condition Is stable

Psltl-Ps

bosh

Vbosh:8osh gas velocity Pb:Blost pressure Po:Top

QOS

pressure

0

s s TllltklTljltk-II "I

Psftl-Ps>Ps· or

Ukb =

:::; IKbltl-KbJ2

I:-n

n 0 :::; Kbltl

PsOI-Ps..: -Ps-

~coltkl- ~coltk-41

'=-n Kb =--n-'-'-

4· "I

Blast pressure Blast moisture I 02

Fig.5. Example of second stage sensor data preprocessing

Blast temperature

Ruult of 8.F. Abnormal Condition

Fig.4. Blast furnace sensors for expert system

curr ent time 20:36

41

38

Probability -I.

63

58

Slip

(OCI 186

8F lop gas lemperalure ,---,----,----,---,----,,----,

Result of B.F. Heat Condition current time 20:20

171.75 h~---l-----.j/-\

ronk 3

0.95

ronk 3

0.70

Heal Transition

ronk 4

0.60

ronk 4

0.53

Action Iype

Moisture

157.5

80

90

G/NM3

01 -4.00

-4.00

The reason of diagnOSis for channeling

Regression dolo

129

lost time 20:00

Heat Level

10PERA T 1ONI

143.2 5 f----+---=~

lost tim e 20:34

ChOMelin9" Probability -I.

100

110

120

Reason

130

Imlnl

Probabilily

Prababilily of chonneling from burden descending

19%

velocity

Fig.5. Example of first stage sensor data preprocessing

Probabilily of channeling from pressure loss

10%

Probabilily of chonnelinq from temperature Probobilily af much amount of residual hot

19% 0%

metol and sloQ

example of display is shown in Fig.7. Moreover various transition graphs and the current results of inference are displayed by request. These examples are shown in Pig.8 and 9. REPRESENTATION OF UNCERTAINTY In the complicated process of blast furnace operation, the judgement of furnace condition wi II differ slightly even in observing the same data. Because operators have there own empirical knowledge. The judgement includes some uncertainty more

Channelinq Iby sensors)

Probabilily of channeling

Ihis time

Tolal

)(

41%

Tolol

)(

41%

Fig.7. Example of results display by Furnace Condition Diagnosis System or less. And the furnace condition is not necessary abnormal even though there are changes in the sensor information, and on the contrary the sensor data do not necessary change before the occurrence of the abnormal furnace condition. This system

Blast Furnace Operation COlltrol

I.0

O.B

~iurd.n slIP-cF

0 .6 ~

0 .4 0 .2

.

eJ'":"

r'



, ';J",!

,

l"\...'" ....

-

~

(Cho,...,eUnQ)

c-- -

:3

1-

,

- r - c-- r-- -

BlSden descending velocity

~~

k

- -

r--

-

1B~U~~ ~ ~ ~r h ~ ~~ ~ l""",

-IBO

-150

-120

r--

OA

r--

-

0.3

~ ~rlF~ ~ f ~

0.2

~

-90

-

c--

-60

-

-30

ImlnJ

7

-

AI

A2 'C 1515

6 L

TRANSITION HIGH E :3 2 5 4 Aa

~

LOW I C

Ba

A

Z 0.0

400

WR. Ma l .

I - - - I - -- - \ i Bz

81

C

C

C

C

80

:C" ,'

C

C

Do

o

1505 5

~N WR MOl WR.MO!

,

X Range

XU

X Sensor data

XL"-XU

Certainty factor

Fig.10. Enlarged .embership function of in-furnace heat transition

r - - - -' -t----j 8

XL

0

6 .00

W~MOl ii

0

.0

0 .1

Fig.B. Infered results of AFS (CRT display)

ACTION MATRIX

I 21 I I 11 I I I I

I nvtnInJ

.

-

5

f--

4

,i~fT1~~ L -,

-b.G -t--;:;

,J'...-.,

~-.:."\..J-

-;,-:n'±rb \-nct

,

Level of In-furnace'

heat transition

" \.J- -,:\:1- :

n ' ;

0 .0

+-

"

----'chl~i-T ~:.--~b~!.;':'~~-!r'=Nv

291

-5.00

-~

it is measured in an iron runner outside of the furnace. For this. the temperature includes noise from the radiation. so the relationship between the hot metal temperature and the furnace heat level is not lucid. Moreover. we thought the heat level itself also has a little vagueness _ Therefore we introduced the membership function that was enlarged to three dimensions as illustrated in Fig.10.

:

Actually there are 7 ranks of furnace heat level and 5 ranks of furnace transition. and each rank can include B F an index of uncertainty. We can set El Eo r---t---- ' - - - CF.value for all ranks of current WR . MOl. HUM. / / furnace heat level and transition >\\ ' . / , T B MAX. I - 1470 using this function from the sensor :~ ;;. El 60.00 :DC;: Ez R G c 2 ~ . ,.. , information which is measured pe-10.00 U /. . .. o 0 .00 riodically . And we determine the .ost I-- 1450 El Ez FI Fz Go proper operation action of the furnace I heat condition fro. the heat level and transition for operation instructions. By introducing the me.bership functions we can treat uncertainty of the judging knowFig. 9. Infered results of HCS (CRT display) ledge and process unifor.ly.so we don't have to increase the number of rules. uses a Certainty Factor (CF.value) and thereby adjusting and refraining of the the .e.bership functions that handle rules efficiently. fuzziness set as an index to represent the the uncertainty. thereby i.proving accuracy of the system . REALIZATION OF REAL-TIME EXPERT SYSTEM Er- 1495

roo . . . CS(/ . 1485 LI-';00. J1.~ fo . :3 ";.~ "

V

E

4

.

.~ ~' .

/

E

/

,~

01

f--+--- -I - - - i 'HR. MO!

-10.00 50 .00

/

~

The CF . va I ue is used in AFS. and has so.e advantages such as the easiness of co.bination calculation and being able to keep the data regularity during calculation. but has so.e defects. such as the difficulty for setting values because of nonlinearity of the calculation. In this syste. we think the CF.value as one of the para.eters of each rule. and present the definite guideline for arranging the value to operators. and prepare the usable .anmachine interface. In HCS. inference is done using the expanded .e.bership function which is based on the idea of 'fuzzy set ' . For exa.ple. the hot .etal te.perature which is an i.portant function for judging the furnace heat level cannot be .easured in the furnace so

Genera II y. a usual expert s ystem is a dialog off-line batch type system. so it is not appropriate for real-time operations. There are so.e restrictions in the real-tile proc ess such as the data gathering from sensors at short cycle. the treating of time series data and the limitation of the time. so it is difficult to realize this by only the usual expert system method (knowledge engineering technology). Therefore it is necessary to divide the functions into the part of using the expert system .ethod and one of using the conventional .ethod . But there are so.e difficult proble.s such as the linkage between the two syste.s and how to divide the functions. Additionally there is a big proble. of the reasoning speed of the expert system itself.

292

M. Shibata . K . H ashimoto and K. Wa kimoto

We resolved these proble.s by the following .ethods. 1) The ti.e series data processing as s.oothing and pattern recognition that is not suitable for the knowledge engineering technology is perfor.ed in the process co.puter using FORTRAN language. 2) We .ake a scheduler on the process co.puter to control the reasoning process on the AI processor. Fro. this we can separate nu.erous data gathering and processing fro. the inference part. 3) We planned to increase the reasoning speed by introducing the AI special processor and dividing KB into s.aller KUs. Actual inference cycle is every 2 .inutes on AFS and every 20 .inutes on HCS.

fl ow

ou tline I

I Decision

I

of the torQe'

: 0

to m o k~ cleor the system functions and

I

scope

I

I Know ledQe

acquisition

I

~ -- ------- --- - --- - - ----- - : 0 10 inv estioote t he repor ts and the I techn ic al books about Ihe fie ld, and the

:

op erat or manu als

I

IO ta coU ect the knowl edQe

f ro m I h ~ upert

I ~ o------ ---- --pr ----- to orronQe the -thinkino ocess --of the

expert

ArronQemen t an d sys temat izot ion

I

of the knowledQe

• checkino the r elat ion and the contradict ion

I

o the patte r n recogn itio n of the sensor dat a • f ilt erino :

-level, diff er ent ial

: 0 to repre sent the v agueness in the knowledQe ~- ------ - -----------------

lDeS iQninQ the structure J : 0 oroupino and str ucturino the rules I

of the rules

SYSTEM DEVELOPMENT ACQUISITION

AND

KNOllLEDGE

An outline of the syste. develop.ent is illustrated in Fig.l1. In the develop.ent of the expert syste. it takes a lot of ti.e to deter.ine the target and coverage of the syste.. to acq'u ire do.a in-spec i f ic knowledge fro. experts. and organize it for use by a progra •. In particular there are following the.es concerning the knowledge acquisition: 1) how to .ake a highly accurate syste. by expressing the vagueness of e.pirical knowledge on the rules of KB 2) how to acquire the knowledge that is not conscious clearly for expert ( the knowledge is co •• on sense for hi. ) 3) how to divide the intensive knowledge by so.e conditions Moreover we think it is necessary for the develop.ent of an expert syste. to get the strong and overall assistance of experts (operator and staff of the factory ). lie have developed the syste. for about one year by 2 knowledge engineers(KE) and 3 experts. The real load of the develop.ent was about 12 aan-.onths except of progra •• ing load. (ordered to the .aker)

: 0

.z ~ IDiVid in
I

J

I

prot otype system

.I

I

.~

l

: 0 t o de vide th e systtm f un ctions t o the : pr eproc essing par t and th e in f er ence P Ori I- _ ______ __ _ _ __ ___ ___ __ __ _ ___ _ _

Con struction of the

IE'II oluotion

to consider the infer enc e speed

~- ---------- . ---------- ---- : 0 t o realize the online reoltime pr oc essing

and tunino

I

I {

: 0 t o des cri be the rules and the Blo Ck-Ba r d I m odeles I

I o ta formulate the reo l sys tem and th e t es t ] system

~- --- - -- -- ---- - --- - -- -- -- - -- : 0 to tune up the sy st em us lno the off- line I tes t sy st em I

I

- to che ck the Validit y of the sy st em

:

- funino the value of the Cer tainty F oclor

I~---

App linQ and fOisin O the le vel

I

--------- --- ----- ---- - -: 0 oddinO and co rre ct ino the r ule s

,I

Fig . 11. Pro c edur e of the dev e lopment

expert system

syste. and study a good deal such as the i.portance of the part of KE and the efficiency or the li.itation of the expert syste •. There are .any subjects to resolve but we plan to raise the level of the expert systel .ore and .ore and to expand the syste. to a higher operation control syste. for the B. F. REFERENCE

EVALUATION OF THE SYSTEM AFS was put into operation when the reblowing of No.5 Blast Furnace in Fukuyama lIorks was started on February 19th. 1986 . And the syste. has been able to predict the occurrence of burden slip. channeling and unstable condition of the furnace with good accuracy. The hitting ratio of the abnor.al condition has been over 80 percent . HCS was put into operation a little later than AFS. and it has been used effectively with the field test and refining the rules on the actual operation. CONCLUSION lie have applied the expert syste. technolOgy based on knowledge engineering to the observation and control of the B.F. operation which is a continuous process. with good results. And in each phase of syste. develop.ent we could obtain various knowledge for the construction of the expert

Sato.T .• (1985). Operation control syste. for blast furnace at OhgishilRa . II..K..K... Technical Report. 107.1-11. Fukuda . T .• (1984) . Develop.ent and application of blast furnace operation contro I syste. in Hi rohata Works . lLo..n.. a.ruL s..t..e..e.L Vo1.70. No.1. 51-57 Kajikawa. Y.. (1984). Develop.ent and appl ication of blast furnace ther.al control syste • . lLo..n.. a.ruL s..t..e..e.L Vol.70. s-800 Yaaaaoto.R .• (1983) . Develop.ent of blast furnace abnoraal condition prediction syste •. lLo..n.. a.ruL s..t..e..e.L Vo I . 69. s-782 Moore.R.L .• (1985) Adding real-tille expert syste. capabilities to large distributed contrl systems. Control Engineering / April. 118-121 Zadeh.L.A . • (1984) Making cOlputer think like people. llEE....SpectrUl. Vo1.25. No . 8 Okada.J . . (1983) DeveloPlent of nuclear reactor a c cident diagnostic syste. based on knowledge engineering technique . J.At.Energy Soc. Japan. Vol . 25. No . 6 Ishizuka.M .• (1983) Treahent of uncertain knowldege. Journal Q.,LSl..C.£.., VoI.22 . No.9