0% found this document useful (0 votes)
190 views5 pages

Nozzle Clogging Prediction in Steel Casting PDF

Uploaded by

Prakash Sarangi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
190 views5 pages

Nozzle Clogging Prediction in Steel Casting PDF

Uploaded by

Prakash Sarangi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Copyright © 2002 IFAC

15th Triennial World Congress, Barcelona, Spain

NOZZLE CLOGGING PREDICTION IN CONTINUOUS CASTING OF STEEL

Jouni Ikäheimonen1, Kauko Leiviskä1, Jari Ruuska1, Jarkko Matkala2

1
University of Oulu, Control Engineering Laboratory, Oulu, Finland
2
Rautaruukki Steel, Raahe, Finland

Abstract: Submerged entry nozzle connects the tundish and the mold in the continuous
casting of steel. Continuous casting is usually done in series including 3–6 successive
heats. Casting of a heat takes about 35–50 minutes. The nozzle is changed after each
series and the new series is started with a new nozzle. About 360–720 tons of steel goes
through a nozzle during its lifetime.

The casting speed and the stopper rod position give the indication of nozzle clogging,
but they cannot, however, answer the question, how long time the casting can continue
and when the nozzle should be changed. In this paper, feedforward neural networks
with backpropagation training were used in modelling the nozzle clogging behaviour at
Rautaruukki Steel mill, at Raahe site. Copyright © 2002 IFAC

Keywords: Neural networks, prediction, fault diagnosis, quality control, steel industry

1. INTRODUCTION 2. NOZZLE CLOGGING PROBLEM

This paper shows the results of a research project that 2.1 What is nozzle clogging?
aimed to find reasons for nozzle clogging in the sub-
merged entry nozzle in continuous casting of steel, Submerged entry nozzle connects the tundish and the
and also to develop tools for predicting the occur- mold in the continuous casting of steel (Fig. 2). It is a
rence of this phenomenon. It is known that better tube with a diameter of about 15 cm and it divides
management of casting brings along considerable the molten steel into two directions on the both sides
potential for production increase and quality im- of the nozzle. The direct flow downwards should be
provement. avoided, because in this case aluminium compounds
in slag would stay in molten steel. The nozzle trans-
The casting plant at Rautaruukki Raahe Steel consists fers molten steel from the tundish to the mold and
of five single strand continuous casters. Casting separates steel from the atmosphere.
process starts by opening the steel ladle (see Fig. 1).
Then the steel flows in a continuous stream through Usually one cast series includes 3–6 successive heats
the ladle into the tundish. From the tundish the steel meaning the same number of ladles. Casting of a heat
flows through a submerged entry nozzle into a mold takes about 35–50 minutes. Each series is designed to
where the solidification of steel starts. Partly solidi- consist of the same steel quality. The nozzle is
fied cast strand is pulled out from the mold by rollers. changed after each series and the new series is started
The inner part of the cast strand is solidified in the with a new nozzle. About 360–720 tons of steel goes
secondary cooling zone by air-water spray cooling. through a nozzle during its lifetime. The risk of
At the end of the casting machine, a torch cutting nozzle clogging increases all the time so that when
machine cuts the cast strand to slabs. about 360 tons of steel has been cast (after the third
ladle) the nozzle condition must be monitored care- 2.2 Reasons and remedies
fully. The operators estimate, based on vision and the
casting plan, if the clogging should be removed or There are several reasons that contribute to nozzle
the nozzle changed in the middle of the cast series. clogging:
Nozzle clogging is not an instantaneous phenome-
non, but rather it develops with time. - metallurgical factors: steel cleanness and
aluminium content,
The clogging can be removed by pumping the - hydrodynamic factors: steel flow rate and
stopper rod and in this way continue the casting nozzle geometry,
series without changing the nozzle. Different nozzle - thermodynamic factors: cold steel and heat
materials and designs and dimensions of the nozzle transfer inside the nozzle,
are available. - nozzle material, and
- unpredictable disturbances and operational
faults.

Ladle Clogging mechanism depends on the cleanness of


steel (Miyazawa, 2001). With clean steels, the
clogging consists of solidified steel, but in other
cases powdered aluminium oxide dominates.
Tundish
Several methods have been used in avoiding nozzle
clogging (Rackers and Thomas, 1995, Pilet and
Bhattacharaya, 1984, Okamoto et al., 1982, Takasugi
Nozzle
et al., 1990):

- calcium silicate injection,


Mold - improving steel cleanness,
- argon injection into the nozzle, and
- nozzle material, construction and geometry.

The effects of nozzle clogging to mold level control


Strand has been studied by Dussud et al. (1998) and Graig et
al. (2001).

2.3 Nozzle clogging detection

There are some variables that can tell about increased


risk for nozzle clogging. Fig. 3 shows trends of cast-
Fig. 1. The continuous casting process.
ing speed and stopper rod position in two cases. The
uppermost figure shows the case without any nozzle
clogging and the lower one shows the opposite case.
Stopping rod The difference is clear and the experiences have
shown that these two variables can give the first indi-
Tundish cation of nozzle clogging.

Using the casting speed and the stopper rod position


Nozzle cannot, however, answer the question, how long time
the casting can continue and when the nozzle should
Clogging areas be changed. Figure 4 shows a block diagram for the
system that aims to estimate also the time available
for undisturbed casting.

Mold level
3. DATA ACQUISITION AND ANALYSIS

This paper reports the study that concerns with the


possibilities to predict nozzle clogging or, more
exactly, to estimate the amount of steel that can be
Fig. 2. Location of the submerged entry nozzle in the cast without changing the nozzle. This is based on
continuous casting process. historical data collected from Rautaruukki Steel
Mill’s converter plant and two casters. The first
analysis revealed that nozzle clogging exists with Does the rod position show
aluminium-killed steel grades that have gone via the the risk of clogging? Small risk
stirring station to the casters. Modelling was done No
with feedforward networks trained by backpropaga- Yes
tion.
Casting continues
High risk
3.1 Data

Data was collected from casters 5 and 6, from 5800


heats from each. All the heats were not used in mod-
elling. Clogging occurs with aluminium-killed heats Use neural nets to calculate
how many tons can be cast
and other heats were left outside. Four steel grade
groups were concerned and they will be called grades
1, 2, 3 and 4 in the following. The heats that had
gone via the ladle furnace were omitted. Silicon-
killed heats were also left without further considera-
Is the prediction reasonable? Casting continues
tion as also heats with serious disturbances. Yes
No
3.2 Correlation analysis
Nozzle change
The analysis considered only these heats where
nozzle clogging existed. First attempts were made Fig. 4. Nozzle clogging detection based on the stop-
keeping all the heats for casters 5 and 6 together. per rod position and neural network model.
Correlations were, however, weak, and therefore the
data was first divided according to the caster and
further into smaller data groups based on four steel In correlation analysis, a total of 67 variables were
grades (1–4). considered. All main temperatures and compositions
were included starting from the converters until the
So, correlation analysis was made separately for two casters together with other process variables and
machines and four grades. It meant eight data groups some additional variables that were calculated from
and the results were used in defining the potential process variables. The most significant are listed in
variables for nozzle clogging models. Table 1 later in Chapter 4. It was also found that in
some cases data from two successive heats were
200 needed.
Ca sting speed
150

100 4. MODELLING APPROACH

50 Stopper rod position Nozzle clogging was modelled using neural net-
works. Modelling with neural networks consists of
0
two phases: training and testing. In training, network
4
27
59

166
220
274
328
379
432
474
507
113

parameters are updated aiming to minimise the dif-


Distance in the strand ference between estimated and actual response val-
ues. In this case, training utilised backpropagation
200 procedure. In testing, response is calculated using
Casting speed constant network parameters. Testing is done with a
150
data set that was not used in training. Training is
100 continued, until the error in the testing phase reaches
50 its minimum. Excessive training must be avoided.
Stopper rod position
0
1672
1742
1814
1870
1932
2005
2077
2134
2189
2219

4.1 Pre-Processing for Neural Networks

Distance in the strand Before training, all input and output variables were
scaled. Usually, the scaling is done in intervals -1 –
Fig. 3. Examples of casting speed and stopper rod +1 or 0 – +1. Most of the inputs were scaled using
position in undisturbed case (upper figure) and in the Matlab-function prestd() that scales the variables
a nozzle clogging case (lower figure). to the average value of zero and the standard devia-
tion 1. The serial number of the heat is scaled inside
the interval 0.16–1.16 by dividing it by 6. The serial
number starts from 1 and the longest series count to
seven heats. The mold width (koklev) was scaled knowledge. The original results are shown in project
inside the interval 0 – +1 using equation 1: reports (Ikäheimonen, Leiviskä, Ruuska, 2001).

Koklev=(koklev-1000)/800 (1)
Predicted
The output variable of the network is the amount of M easured
cast tons with a certain nozzle without pumping. Its
values were scaled inside the interval 0.16–1.16 by
dividing the actual measurement value by 720 that is
the total amount of six heats in tons.

4.2 Network Structure

Modelling used feedforward networks with only one


hidden layer. These networks model steady-state
dependencies between input and output variables.
Two activation functions were applied, namely a
hyperbolic tangent (y = tanh(x)) and a linear function Fig. 5. An example of test results.
(y=x). Modelling used Matlab’s NN-toolbox. The
first layer included non-linear tanh-functions and the
second layer a linear y= x function. This made non- 5.2 Cross-Testing
linear modelling possible.
Cross-testing considered applying the model devel-
oped for caster 5 to the data from caster 6. Only
4.3 On the Training Data grade 1 showed reasonable results. This confirms the
earlier results claiming that different variables effect
Three principles were used while selecting the input the nozzle clogging in different casters.
variables for the neural networks:
Table 1. The most significant variables. N denotes
- The variables had a correlation with cast how many times the variable was included in 26
tons over 0.20 models considered.
- They were not the set points for process Variable N
controllers Serial number of heat in the series 15
- They were not correlating with each other. Casting speed*mold width 10
The limit was chosen as 0.50. Mold width 9
Tundish temperature of the first heat 8
The amount of data limits the size of the network. Steel temperature after stirring station 6
The number of training parameters should not exceed Nitrogen content at stirring station 6
the number of training points. In practice, network Average tundish temperature 5
modelling is difficult if the number of data point is Aluminium oxide content at stirring 4
less than 60, because training requires 30 points at Temperature deviation in tundish 4
minimum and almost the same number is needed for
testing. With five inputs and 30 training points, a
conventional network can include only 5 neurons. 5.3 Testing with Successful Cases
These limitations were especially met in the caster 6.
Modelling used only data from cases where clogging
occurred. Models for caster 6 were tested using data
5. MAIN RESULTS from corresponding successful heats. The result was
as expected. It is impossible to tell how many tons
5.1 Significant Variables could have been cast in successful cases, if casting
had continued with the same nozzle. Therefore, the
Several models were developed for both casters. The predicted cast tons given by the models remain in
best models can predict the cast tons with ±60 tons these cases lower than actual. The average error is
accuracy in over 80 % of cases. Models for caster 6 about one heat, 120 tons. The result is important from
show a better accuracy. Fig. 5 shows a typical exam- the model application point of view: models never
ple. Table 1 shows a summary of 12 models for predict too high cast tons. The opposite result could
caster 5 and 14 models for caster 6. It shows the vari- lead to erroneous operator actions: cast could con-
ables that were used in the models. There are no sur- tinue in spite of increased clogging risk. Fig. 6 shows
prises in this table; the listed variables can be an example run.
assumed to effect on nozzle clogging also by a priori
1 000
Results seem promising; in several cases the cast tons
are estimated with the accuracy of ±60 tons in more
800 than 80% of cases. Not so encouraging was the find-
Actual cast tons ing that different variables become dominating in
600 different machines. The models were trained using
clogging cases, only. When testing the models with
400 successful castings, it was found that the models
never gave too high predictions.
Predicted cast tons
200

REFERENCES
0
Actual-predicted
-20 0 Craig, I.K., F.R. Camisani-Calzolari, P.C. Pistorius
0 50 1 00 150 2 00 2 50 300 (2001). A contemplative stance on the automa-
tion of continuous casting in steel processing.
Fig. 6. An example from testing the NN-model with Control Engineering Practice 9, pp. 1013-1020.
data from successful cases. Dussud, M., S. Galichet, L.P. Foulloy (2000). Appli-
cation of fuzzy logic control for continuous
casting mold level control. IEEE Trans. on Con-
6. CONCLUSION trol Systems Technology, 6. no. 2, pp. 246-256.
Ikäheimonen, J., K. Leiviskä, J. Ruuska (2001).
Nozzle clogging problem causes production losses Report B32, University of Oulu, Control Engi-
and quality impairment in continuous casting, espe- neering Laboratory. (In Finnish)
cially with aluminium-killed steels. Variations in Miyazawa, K. (2001). Continuous casting in Japan.
stopper rod position and casting speed provide the Science and Technology of Advanced Materials,
operator with the first information on increased risk 2, pp. 59-63.
of nozzle clogging. This cannot, however, answer the Pielet, H.M., D. Bhattacharya (1984). Thermody-
question, how long time the casting can continue and namics of nozzle blockage in continuous casting
when the nozzle should be changed. In this paper, a of calcium-containing steels. Metallurgical
system is proposed that aims to estimate also the time Transactions B, 15, pp. 547-562.
available for undisturbed casting. Rackers, K.G. and B.G. Thomas (1995). Clogging of
continuous casting nozzles. Steelmaking Confer-
The estimate of the amount of steel that can be cast ence Proceedinfs, pp. 723-734.
without changing the nozzle is predicted using neural Okamoto, K., T. Nakamura, M. Kondo (1982).
network models. These models are based on data Development of alumina-graphite immersion
collected from Rautaruukki Steel Mill’s converter nozzle for continuous casting. Iron and Steel
plant and two casters; numbers 5 and 6. A total num- Engineering, no. 9, pp. 47-52.
ber of 5800 heats and 67 variables were analysed. Takasugi, H., T. Masaoga, A. Shirayama, T. Mori, H.
However, the number of variables effecting nozzle Murakami (1990). Prevention of alumina build-
clogging is quite small. Feedforward networks with up in submerged entry nozzles for continuous
backpropagation were applied. Separate model for casting of low carbon Al-killed steels. 6th Inter-
both casters were needed and, in addition to this, data national Iron and Steel Conference, Nagoya, pp.
had to be divided in four quality groups. This shows 462-469.
that different variables dominate in different cases.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy