You are on page 1of 6

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/230582537

Online Prediction of Longitudinal Facial Cracks


in Thin Slab Casting Using Principal
Component Analysis
Article in Journal of Iron and Steel Research International October 2008

CITATIONS

READS

192

Some of the authors of this publication are also working on these related projects:

Surface quality of AHSS steel grades. Online precdiction and effective countermeasures.

View project

Available from: Dirk Lieftucht


Retrieved on: 31 October 2016

Online prediction of longitudinal facial cracks in thin slab casting


using principal component analysis
Dirk Lieftucht, Matthias Arzberger, Markus Reifferscheid, Jochen Schlter
(SMS Demag AG, Eduard-Schloemann-Strae 4, D-40237 Dsseldorf, Germany)
Abstract:An online algorithm to detect longitudinal facial cracks (LFC) has been developed and integrated into the mould
monitoring system (MMS) of a thin slab caster. The algorithm is based on the thermocouple readings in the mould and applies a
hybrid model approach using the principal component analysis (PCA) in terms of defect detection and an expert system including
fuzzy rules regarding defect identification and isolation. The model based technology is tolerant to distorted thermocouple readings
and is consequently leading to a low false alarm rate. Results from thin slab casters in different Compact Strip Production (CSP) lines
prove the beneficial usability of the described concept to evaluate process stability regarding longitudinal facial crack appearance
dependent on different caster settings and process conditions.
Key words:Continuous casting,thin slab casting,mould monitoring system,longitudinal facial crack detection,multivariate statistical
process control (MSPC),principal component analysis (PCA),expert systems,fault diagnosis,statistical analysis

1. Introduction
Quality demands are rising in Compact Strip
Production (CSP) due to extension of the product
mix from soft non-alloyed and high strength low
alloyed steel grades to higher sophisticated
multiphase steel grades with excellent surface
quality properties to full fill also the demand of the
automotive industry [1]. Increasing abilities to
monitor the steel casting and rolling process lead to
the demand of an ongoing reduction regarding defect
appearance on the final product.
Early works on longitudinal facial crack (LFC)
detection are known from conventional continuous
casting, where the defect can easily be monitored by
visual inspection of the cold slab [2]. The visual
inspections correlate with abnormal drops of
thermocouple recordings in the mould of the slab
caster. In general the LFC appearance is a harmful
product defect. In some rare cases when the strand
shell is weakened so much by the LFC that the
thermo-mechanical stress during casting is tearing
open it even can result into a breakout, i.e. a
disruption of the casting process.
To meet the required quality demands the process
data and the LFC defect have to be monitored and
denoted in a data base to enable a learning effect
regarding the best process settings to rigorously
reduce the amount of defect appearance. Challenges
are in the fully automatic system and foremost the
reduction of false alarm rates. More recent
developments
handle
fault
detection
of
thermocouple signal patterns without the real defect
identification [3]. However, it is common
understanding that in quality detection systems the
implementation of qualitative and quantitative
human understanding is required to achieve an
automated quality control [4].
The main aims of this work are therefore to develop
are:

1. fully automatic detection of the LFC,


2. robust signal handling, i.e. tackling the task
to deliver correct detection results even if
faulty thermocouple signals are present,
3. online process feedback regarding LFC
occurrence to deliver qualitative and
quantitative human process understanding.
The developed model operates with the
thermocouples of the mould monitoring system
(MMS) of the CSP caster and is running as a
additional MMS function at three different casters
on 5 strands.

2. LFC characteristics and basic algorithm layout


The LFC algorithm focuses on cracks which arise in
the mould. The proof of defect origin is done by
metallographic investigations indicating the
presence of mould powder components and
consecutive reaction products within the crack itself.
Broad Face Direction

Depression type

Vcast

Bleeder type depression

Heat crack

Fig. 1

Different types of LFC appearances

These LFC appear with a length between some


centimetres and several metres on the broad faces of
the slab. Three groups can be distinguished as
shown in Fig. (1).
The upper sketch in Fig. 1 shows the most common
LFC in thin slab casting which appears in
combination with a depression having shallow sides
and a total width of approximately 20 mm. The
crack itself runs in the bottom of the depression in
casting direction. If the crack tears open in the
mould the LFC bleeder type depression is present as
shown in the middle sketch. Finally the lower sketch
is highlighting the typical heat crack with a
characteristic sharp edge morphology. Crack contour
in casting direction is often changing the position in
width direction.
All described crack types are accompanied by a
local characteristic disturbance in the heat extraction
in the mould which is monitored by the MMS
thermocouples arranged in columns as shown in Fig.
2.

Fixed Side
Narrow
Face
Left

Loose Side

Narrow Face Right

Fig. 2

Casting Direction

50 mm to the next thermocouple column can be


detected.
The model structure describing most important
algorithm functions is shown in Fig. 3
Step 1: Pre-processing
Pre-processing

Signal disturbances

Normalise and
validate signals
Step
Step 2:
2: Detection
Detection
Search for defect
temperature patterns

Principal Component Analysis (PCA) of


thermocouple signals

Gradient analysis (Fuzzy)

Step
Step 3:
3: Isolation
Isolation
Kind, location and
time of defect
Step 4: Identification

LFC

Time-variant
behaviour and
counter action

Roll out

Temperature
drift

Other

Fig. 3
Hybrid, multi-step model structure for LFC
defect detection

Step1: At first a data pre-processing handles all


signal disturbances at special caster situation e.g.
start cast, regular casting operation and cast end.
Explainable signal patterns occurring during normal
casting operation are shown in Fig. 4. The
highlighted examples include a characteristic signal
pattern of an LFC and a bleeder type LFC and signal
disturbances like defect thermocouple, contact
problem, shielding problem in case of EMBr
operation. Additionally typical drift patterns are
visualised like slag roll out local accumulation of
mould powder on the mould surface and its sudden
removal -, cast speed changes and influence of
narrow side position. Final part of the data preprocessing is the transformation from a time into a
location based system
LFC

Bleeder type Roll out


LFC

Cast speed
changes
(drift)

EMBR
(shielding)

Thermocouple
contact

Narrow face
influence

MMS thermocouple arrangement

The crack detection is performed by a column wise


evaluation of thermocouple readings within the
casting format and contact length in casting
direction. For reliable LFC identification under
industrial conditions three operating thermocouples
are required in one column. Acknowledging a 10%
rate of distorted thermocouples minimum five
thermocouples in one column are recommended as
the minimum installation.
Most characteristic signal patterns arise if the defect
directly appears in front of a thermocouple column.
Due to the width of the heat extraction disturbance
resulting into a crack and the two dimensional heat
flow in the copper plate cracks with a distance of

Start of cast
Defect Thermocouples
Shielding
Narrow side
influences

Fig. 4

Damaged
thermocouple

Different types of thermocouple signal pattern

Step 2: The detection algorithm has to be able to


incorporate defect or distorted thermocouple signals

and to separate them from the regular signal pattern.


Multivariate statistical process control (MSPC)
techniques take relations between different variables
into account and have a low level of false alarms.
PCA is one applied MSPC technology often in
industrial applications [5]. The PCA analysis
requiring normalized data with cero mean and unit
variance is applied for fault detection.
The fault detection operates with the prediction error
of the PCA model which is based on defect free
training data set. The prediction error - difference
between predicted and measured value - is
highlighted for an example of a thermocouple
reading of one column in Fig. 5. The PCA model
can not mimic the LFC signature properly, i.e. the
LFC signature is characterized by a large deviation
between predicted and measured value at multiple
thermocouples. Anyway sometimes other signal
patterns roll outs, temperature drifts - are also
caught by the fault detection method.

Resultant the applied multiply step technique


combining data pre-processing, PCA method and
fuzzy expert know how - Fig. 3 provides high
hitting rates on LFC signal patterns having a very
well acceptable false alarm rate.

3. Application
The following application examples of the
developed MMS function focus on the usability of
the detection scheme in every day process operation
of a thin slab caster.
3.1 LFC defect map on operator human machine
interface (HMI)
MMSs can generate an overwhelming flood
of information and process views. Most information
are of minor importance to the operators typical
temperature-time charts without LFC analysis are
open to operators interpretations. The MMS-LFC
uses a defect map focussing on the location of the
cracks at the broad faces during the casting
operation. The example shown in Fig. 6
demonstrates the appearance of an off-corner LFC
close to the narrow face position. The severe
accumulation of LFC alarm clearly indicates the
operator the place of defect appearance and
counteraction feed back like check of narrow face
taper or mould powder rim formation.

Prediction

Measurement

LFC

Fig. 5

LFC

3.2 Casting powder development


The performance of new casting powders
can be validated concerning process stability in
comparison to the established standard product using
the LFC algorithm showing the tendency of LFC
appearance. Fig. 7 shows the direct comparison of a)
the standard casting powder and b) the new casting
powder tested.

PCA method and its prediction error

Step 3 and 4: Defect diagnosis, consisting of


identification and isolation is realised using an
expert system. The expert knowledge of the defect
behaviour is used to validate fuzzy parameter, to
distinguish different defect types.
Fuzzy parameters are based on column wise
thermocouple gradients to extract the LFC signal
patterns only.

Broad Face Positions

a) Normal casting conditions

b) Abnormal situation

-637
-425

Narrow Face
Position

-212
0
212

Predicted
LFC

425
637
1200

1400

1600

1800

Fig. 6 Online MMS-LFC defect map

2000
2200
Cast Length [m]

2400

2600

2800

3000

a) Current casting powder

b) Tested casting powder

Fig. 7 Comparison of different casting powder regarding LFC occurrence

Even after relatively short casting trial it is possible


to compare the tendency in behaviour regarding the
LFC occurrence using different casting powders. As
all other important process information are stored
together with LFC detections in a long term database also statistic correlations between several
process variable are possible.
3.3 Statistical long term evaluation
Using the MMS-LFC function in the offline
mode large number of data sets are evaluated in an
over night task. A six month statistics shows the
influence of copper plate thickness in front the water
channel in correlation to the LFC index. The broad
face copper plates are divided into three groups, i.e.
new copper plates (first third of life time), best age
copper plates (second third of life time) and thin
copper plates (last third of life time). Each group
entails a similar amount of data. The results in Fig. 8
are in well accordance with thin slab casting
experiences, that as thinner the copper thickness
becomes as more LFC defect appear on the product.

mould onto the LFC index has been evaluated. The


existing funnel layout have been predicted leading to
an average LFC index of 4.5 whereas the redesigned
mould design designs reduces the LFC index to 1.7,
see Fig. 9.

LFC index
4.5
4
3.5
3
2.5
2
1.5
1
0.5

100

% of all sequences covered

90

Existing funnel

80

Redesigned
funnel

70

Fig. 10

60

4. Conclusion

50
40

thin copper

30

medium copper

20

thick copper

10
0
0

LFC index before and after redesigned funnel

6
LFC index

10

Fig. 8
Correlation of copper thickness in front of water
channel to LFC occurrence

In a second statistical evaluation over 1 year


production the impact of a redesigned funnel type

A new robust LFC detection technique for thin slab


casting has been successfully developed and
integrated into a mould monitoring system. System
robustness against malfunction of thermocouples is
given at high level of precision regarding LFC
detection. The presented tool is powerful regarding
future product quality improvements and process
stability enhancements. The limitation in system
performance is linked to the given horizontal
distance between thermocouple columns.

References:
[1] Rosenthal D, Henning W. CSP the trend setting
technology for more than 15 years. International symposium
on thin slab casting and rolling, p 9-17, 2006.
[2] Stewart D, Hewitt P N, Peeters L. The prediction of
longitudinal cracking during slab casting. 79th Steelmaking
Conference Proceedings, p. 207-214, 1996.
[3] Hemy P, Smylie R, Srinivasan C. Analyzing casting
problems by the on-line monitoring of continuous casting
mold
temperatures.
JOM-e,
http://:www.tms.org/pubs/journals/JOM/0201/Hemy/Hemy0201.hml, 2002.
[4] Thomas B G. On-line detection of quality problems in
continuous casting of steel. Modeling, Control and
Optimization in Ferrous and Nonferrous Industry, Chicago,
IL, TMS, Warrendale, PA, 2003.
[5] Chiang L H, Russell E L, Braatz R D. Fault detection and
diagnosis in industrial systems, advanced textbooks in control
and signal processing. Springer Verlag, 2001.

You might also like