Swansea University E-Theses

_________________________________________________________________________

Hot mill process parameters impacting on hot mill tertiary scale

formation.

Kennedy, Jonathan Ian

How to cite:
_________________________________________________________________________
Kennedy, Jonathan Ian (2012) Hot mill process parameters impacting on hot mill tertiary scale formation.. thesis,
Swansea University.
http://cronfa.swan.ac.uk/Record/cronfa42262

Use policy:
_________________________________________________________________________
This item is brought to you by Swansea University. Any person downloading material is agreeing to abide by the terms
of the repository licence: copies of full text items may be used or reproduced in any format or medium, without prior
permission for personal research or study, educational or non-commercial purposes only. The copyright for any work
remains with the original author unless otherwise specified. The full-text must not be sold in any format or medium
without the formal permission of the copyright holder. Permission for multiple reproductions should be obtained from
the original author.

Authors are personally responsible for adhering to copyright and publisher restrictions when uploading content to the
repository.

Please link to the metadata record in the Swansea University repository, Cronfa (link given in the citation reference
above.)

http://www.swansea.ac.uk/library/researchsupport/ris-support/
 EPSRC Engineering Doctorate Centre on Steel Technology

Hot Mill Process Parameters Impacting on Hot Mill Tertiary Scale Formation

By

Jonathan Ian Kennedy

■v

A'V'f

A thesis submitted to Swansea University

For the Degree o f Doctor o f Engineering

2012

Department o f Materials Engineering

Swansea University
 ProQuest Number: 10797970

All rights reserved

INFORMATION TO ALL USERS

The quality of this reproduction is d e p e n d e n t upon the quality of the copy subm itted.

In the unlikely e v e n t that the a u thor did not send a c o m p le te m anuscript

and there are missing pages, these will be noted. Also, if m aterial had to be rem oved,
a n o te will ind ica te the deletion.

uest
ProQuest 10797970

Published by ProQuest LLC(2018). C opyright of the Dissertation is held by the Author.

All rights reserved.

This work is protected against unauthorized copying under Title 17, United States C o d e
M icroform Edition © ProQuest LLC.

ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, Ml 4 8 1 0 6 - 1346
DECLARATION

This work has not previously been accepted in substance for any degree and is not being concurrently
submitted in candidature for any degree.

Signed (candidate)

Date...................

STATEMENT 1

This thesis is the result of my own investigations, except where otherwise stated. Where correction
services have been used, the extent and nature of the correction is clearly marked in a footnote(s).

Other sources are acknowledged by footnotes giving explicit references. A bibliography is appended.

Signed....................... (candidate)

Date

STATEMENT 2

I hereby give consent for my thesis, if accepted, to be available for photocopying and for inter-library
loan, and for the title and summary to be made available to outside organisations.

Signed (candidate)
THESIS SUMMARY

Candidate's Surname: Kennedy

Candidate's Forenames: Jonathan Ian
Candidate for the Degree of: Engineering Doctorate in Steel Technology
Full title of thesis: Hot Mill Process Parameters Impacting on Hot Mill Tertiary Scale
Formation

Summary:

For high end steel applications surface quality is paramount to deliver a suitable product. A
major cause o f surface quality issues is from the formation o f tertiary scale. The scale
formation depends on numerous factors such as thermo-mechanical processing routes,
chemical composition, thickness and rolls used. This thesis utilises a collection o f data
mining techniques to better understand the influence o f Hot Mill process parameters on scale
formation at Port Talbot Hot Strip Mill in South Wales.

The dataset to which these data mining techniques were applied was carefully chosen to
reduce process variation. There are several main factors that were considered to minimise this
variability including time period, grade and gauge investigated. The following data mining
techniques were chosen to investigate this dataset:
• Partial Least Squares (PLS)
• Logit Analysis
• Principle Component Analysis (PCA)
• Multinomial Logistical Regression (MLR)
• Adaptive Neuro Inference Fuzzy Systems (ANFIS)

The analysis indicated that the most significant variable for scale formation is the temperature
entering the finishing mill. If the temperature is controlled on entering the finishing mill scale
will not be formed. Values greater than 1070 °C for the average Roughing Mill and above
1050 °C for the average Crop Shear temperature are considered high, with values greater than
this increasing the chance o f scale formation. As the temperature increases more scale
suppression measures are required to limit scale formation, with high temperatures more
likely to generate a greater amount o f scale even with fully functional scale suppression
systems in place.

Chemistry is also a significant factor in scale formation, with Phosphorus being the most
significant o f the chemistry variables. It is recommended that the chemistry specification for
Phosphorus be limited to a maximum value o f 0.015 % rather than 0.020 % to limit scale
formation. Slabs with higher values should be treated with particular care when being
processed through the Hot Mill to limit scale formation.
Acknowledgments
This work would not have been possible if not for the assistance o f a number o f key people
and I would like to take a moment to acknowledge their contribution to this study.

First I would like to acknowledge the invaluable assistance o f Dr Fiona Robinson. Her depth
o f knowledge and contacts proved invaluable in running a successful project.

Dr Mark Evans for his advice and guidance on writing papers and interpreting data.

Mr Sid Petitte, whose decades o f experience in the steel industry were invaluable to
understanding which process parameters were important in influencing scale formation in
Port Talbot Hot Mill.

Dr John Albiston for giving me advice on suitable software available within Tata steel for the
project.

Dr Paul Thomas, for taking over the industrial supervisor position in the final year and aiding
me in structuring my thesis.

Dr James Sullivan for his support in the early stages o f the project.

I would also like to thank from the bottom o f my heart Natasha Marshall, for being so
supportive through the closing months of the write up. Also, for being the best person I know.

This research was funded by a grant from the Physical and Sciences Research Council
(EPSRC) and TATA Steel Strip Products, UK (TSSP-UK).

And last (but not least) my family who have offered moral support and advice throughout the
doctorate.
List of Publications
The following is a list o f publications I have written or been involved in during my doctorate.

Kennedy, J., Evans, M., Robinson, F., Identification for control o f the process parameters
influencing tertiary scale formation at the hot strip mill using a binary choice model, Journal
o f Materials Processing Technology, Volume 212, Issue 7, July 2012, Pages 1622-1630

Evans, M., Kennedy, J., Thomas, P., Process parameters influencing tertiary scale formation
at a hot strip mill using a multinomial logit model, Journal o f Manufacturing Science and
Engineering

Evans, M., Kennedy, J., Thomas, P., Non-linear partial least squares and its application to
minimising scale formation at a steel hot mill, Materials Science and Technology, Volume
28, Number 12, December 2012, pp. 1513-1522(10)

Bright, G., Kennedy, J., Robinson, F., Evans, M., Whittaker, M.T., Sullivan, J., Gao
Y.,Variability in the mechanical properties and processing conditions o f a High Strength Low
Alloy steel, Procedia Engineering, Volume 10, 2011, Pages 106-111

Sullivan, J., Weirman, C., Kennedy, J., Penney, D., Influence o f steel gauge on the
microstructure and corrosion performance o f zinc alloy coated steels, Corrosion Science,
Volume 52, Issue 5, May 2010, Pages 1853-1862
Contents

1 Chapter 1. Literature Review 2

1.1 Processes 2
1.1.1 Hot Mill Overview 2
1.1.2 Slab Yard 3
1.1.2.1 Slab Defects 3
1.1.2.2 Round 4
1.1.3 Furnace 4
1.1.4 Scale Breaker and Reversing Rougher Mill 5
1.1.5 Coil Box 5
1.1.6 Finishing Mill 6
1.1.7 Finishing Mill Entry 6
1.1.7.1 Crop Shear 6
1.1.7.2 Roll Coolant 7
1.1.7.3 Finishing Scale Breaker (FSB) 7
1.1.7.4 Water Containment 10
1.1.7.5 F5 Pre-Sprays 11
1.1.7.6 Finishing Mill Scrubbers and Edge Sprays 12
1.1.7.7 Interstand Cooling 13
1.1.8 Parsytec 13
1.1.8.1 Detection 14
1.1.8.2 Classification 14
1.1.8.3 Action 16
1.1.9 Run Out Table (ROT) and Coiling 16
1.1.10 Conclusions for Process 16
1.2 Scale 17
1.2.1 What is iron oxide made of? 17
1.2.2 Influences on Scale 19
1.2.3 Mill layout 20
1.2.4 Temperature 20
1.2.5 Dimensions 21
1.2.6 Time 21
1.2.7 Chemistry 21
1.2.7.1 Silicon 22
1.2.7.2 Phosphorus 22
1.2.8 Descaler sprays 23
1.2.9 Rolling 23
1.2.10 Downstream Scale 24
1.2.11 Conclusions to Scale section 25
1.2.12 References 25
1.3 Data Mining 27
1.3.1 Directed and Undirected Techniques 28
1.3.2 Cluster Detection 29
1.3.3 Binomial Regression 29
1.3.4 Classification and Regression Trees (CART) 29
1.3.4.1 Pruning 30
1.3.5 Artificial Neural Networks (ANN) 30
1.3.6 Multivariate Adaptive Regression Splines (MARS) 31
1.3.6.1 Hinge Function 31
1.3.7 Principle Component Analysis (PCA) 32
1.3.8 Partial Least Squares (PLS) 32
1.3.9 Genetic Algorithms 32
1.3.10 Multinomial Logistical Regression (MLR) 32
1.3.11 Adaptive Neural Fuzzy Inference System (ANFIS) 33
1.3.12 Conclusions to Data Mining 33

2 Chapter 2. Aims o f Research Project 34

3 Chapter 3. Experimental Techniques 35

3.1 Data Collection 35
3.1.1 Considerations 35
3.2 Statistical Methods 38
3.2.1 Binary Statistical methodology 39

4 Chapter 4. Identification, for control, o f the process parameters

influencing tertiary scale formation at the hot strip mill using a
binary choice model 40
4.1 Principal component analysis 40
4.2 Binary choice models 41
4.2.1 A generalised framework 41
4.3 Parameter estimation 46
4.4 Model identification 46
4.5 Results and evaluation 47
4.5.1 Principal component testing procedure 47
4.5.2 Distribution analysis 51
4.6 Logit model results 52
4.7 Logit Conclusions 57
4.8 References 57

5 Chapter 5. A Partial Least Squares Generalized Linear

Regression Algorithm 59
5.1 The Generalised Linear Model 59
5.2 Partial Least Squares within a Generalised Linear Model 61
5.2.1 Determination o f the first PLS component //,: 61
5.2.2 Determination o f the second PLS component t2i: 62
5.2.3 Determination o f the other PLS components: 62
5.3 PLS Results and discussion 63
5.3.1 The PLS Components and the PLS Logit Model 63
5.4 Analysis and Interpretation 68
5.5 PLS Conclusions 73
5.6 References 74

6 Chapter 6. Process parameters influencing tertiary scale

formation at a hot strip mill using a multinomial logit Model 75
6.1 Introduction 75
6.2 MLR data classification 75

viii ]
6.3 Principle Component Analysis 76
6.4 MLR model 77
6.5 Estimating a Multinominal Logit Model 78
6.6 MLR Results and Discussion 78
6.6.1 Principle component testing procedure 78
6.6.2 Multinominal logistic model results 81
6.7 MLR Conclusions 89
6.8 References 90

7 Chapter 7. Development of a multi-layer ANFIS model for the

prediction of tertiary scale formation 91
7.1 ANFIS model 91
7.2 ANFIS Results and Discussion 96
7.3 ANFIS Conclusions 101
7.4 References 102

8 Chapter 8. General Discussion 103

8.1 Binary vs. Multinomial vs. Continuous 103
8.1.1 Binary 103
8.1.2 Multinomial 104
8.1.3 Continuous 104
8.2 Model Comparisons 105
8.3 Model Use 107
8.4 Scale Formation 107

9 Chapter 9. General Conclusions 109

10 Chapter 10. Recommendations 111

List of Tables

Table 1. Slab Specifications (Dimensions)......................................................................... 3

Table 2. Slab Specifications (Bending)...............................................................................4
Table 3. Parsytec Recognition Rates................................................................................. 15
Table 4. Typical mechanical properties for scale phases (Bolt (2003))........................19
Table 5. Process Variables and their Sample Means and Standard Deviations.......... 37
Table 6 . Correlation matrix for the temperature process variables............................... 48
Table 7. Principle component results for the temperature process variables...............50
Table 8 . Logit model results............................................................................................... 53
Table 9. Quasi-elasticities................................................................................................... 54
Table 10. Estimated values for the pij weights for allalloying elements...................63

Table 11. Estimated values for the Py weights for various temperatures.................... 64
Table 12. Examples o f using the PLS logit model for process control.........................72
Table 13. Classification of splits for the MLR model.....................................................76
Table 14. Variables used in the construction o f the principle components................. 79
Table 15. Principal component results for the temperature process variables............ 80
Table 16. The estimated values for the parameters in Eq. (6.4).................................... 82
Table 17. Table o f actual and predicted values for scale count..................................... 83
Table 18. Elasticities for each o f the statistically significant process variables..........85
Table 19. Maximum temperature settings to achieve a 10% or less chance o f very low scale
forming...................................................................................................................................89
Table 20. Process Variables and their Sample Means and Standard Deviations 91
Table 21. Table showing the sum of the square errors for the different variables used.96
Table 22. Table showing the regression parameters for Var 1,2 ,3 . Bold numbers represent the
most significant parameters for that function................................................................. 1 0 0
Table 23. Table showing the calculated k and a values at Vari,2 , 3 ............................ 100
Table 24. Table showing selected coils at different scale counts................................101
Table 25. Lead Monitor for Tinplate grades................................................................... 110
List of Figures

Figure 1. Overview o f the Hot M ill..................................................................................... 2

Figure 2. Finishing Mill Entry.............................................................................................. 6
Figure 3. Lead Test...............................................................................................................10
Figure 4. Composition o f scale........................................................................................... 17
Figure 5. Scale as observed from the galvanising line....................................................25
Figure 6. The maximised log likelihood obtained at different values for p and q ...... 52
Figure 7. Probability o f scale formation with variations in the significant process variables.
................................................................................................................................................ 55
Figure 8. Predicted Probabilities o f Scale as a Function o f the PLS Components.....69
Figure 9. Cross Plot o f the PLS Components...................................................................70
Figure 10. Calculated Elasties for Each Process Variable..............................................71
Figure 11. Cumulative Frequency Plot for the Average Rougher Mill Exit Temperature. 72
Figure 12. Cumulative Frequency Plot for the Average Crop Shear Temperature. ...73
Figure 13. Probability o f scale formation with variations in some o f the process variables. 87
Figure 14. A typical binary regression tree with two inputs (xi and X2) and one output y. 92
Figure 15. ANFIS architecture corresponding to the representation shown in Figure 14.
Values for w given by Eq. (7.3)......................................................................................... 95
Figure 16. Observed tertiary scale versus predicted tertiary scale for Vari (AvgRM).97
Figure 17. Observed tertiary scale versus predicted tertiary scale for V a r^ (AvgRM, PerP).
................................................................................................................................................ 98
Figure 18. Observed tertiary scale versus predicted tertiary scale for V a r i^ (AvgRM, PerP,
AvgCS)...................................................................................................................................99
C h ap ter 1. L iterature Review

1.1 Processes

This section will introduce the different processes that occur in the Hot Mill and will
concentrate on the areas that have the greatest influence over scale formation.

1.1.1 Hot Mill Overview

The steel works in Port Talbot is a fully integrated steel works, producing steel from raw
materials with limited scrap additions. Steel is continuously cast at Port Talbot using a
continuous caster.

After being processed through the continuous caster the material is transported to the Hot
Mill (see figure 1) for further processing. The purpose o f the Hot Mill is to reduce the gauge
of the slab to form a hot rolled coil that can either be dispatched directly to a customer or
proceed for further processing. The aspects o f the processing that have an effect on upstream
surface quality are investigated. The Hot Mill is split into several sections including;
• Slab Yard
• Reheat furnace
• Scale breaker
• Reversing roughing mill
• Coil box
• Crop shear
• Finishing mill
• Run Out Table
• Coiler

Reversing Crop
R eheat R o u g h i n g Mill Shears F i n i s h i n g Mill
R u n o u t T a b le
F urnace
Coil B o x

,#AAAAAAA
A orB
OQOQC^OO

Figure 1. Overview of the Hot Mill.

n
1.1.2 Slab Yard

There are several purposes o f the slab yard, including the following;
• To receive slabs from the continuous caster
• To store slabs before they are processed
• To provide the slabs in the correct order for the Reheat Furnace
• To identify and rectify quality issues

When a slab arrives in the slab yard one o f two things will happen;
1. If the slab is part of a hot connect round it will be sent straight to the Hot Mill for
processing. Hot connect slabs have several advantages including;
a. Saving money on reheating because they start at a higher temperature when
entering the Furnace.
b. Reduce time between placed order and delivery.
c. Saving space. Under normal conditions a slab would sit in the slab yard for
several days to cool down before being processed for safety reasons.
2. Most slabs will be left to cool for several days so they can be checked and put into a
round for processing in the Hot Mill.

1.1.2.1 Slab Defects

It is vital that the slabs dimensions are tightly controlled before entering the Hot Mill (see
table 1). The Port Talbot reheat furnace is a walking beam furnace. This means that there are
a number o f stationary beams to support the slab and a number o f powered beams to move
the slab through the furnace. If the processed slabs are too long for the Hot Mill it can
damage the Furnace and if it is too short it could get stuck in the Furnace. The dimensions are
dependent on which mill the slab is going to be rolled at. Slabs that are too short (< 7 m) to
be processed at Port Talbot can be processed via Llanwem Hot Mill rather than be scrapped.

Table 1. Slab Specifications (Dimensions).

Length (m) Width (mm) Thickness (mm)

o
O
T-^
o

Port Talbot 6 4 0 - 1875 1 9 0 -2 5 4

1

Llanwem 5 .5 -8 .7 5 7 0 0 - 1550 1 9 0 -2 5 4
The shape o f the bar is also very important, because poor shape could lead to difficulties in
charging and reheating in the Furnaces. Or in the worst case the slabs are scrapped to protect
the Furnace and Hot Mill.

Table 2. Slab Specifications (Bending).

Max. bend off Max. bend off straight Max. bend o ff straight
horizontal (mm) (single width) (mm) (slit width) (mm)
Port Talbot 70 25 50
Llanwem 70 25 50

To stop slabs from bending during cooling, stacks should be arranged so that the longest slabs
are on the bottom or give a maximum overhang o f 1 metre. If the slabs deviate by more than
the values in table 2 then the slab will be placed between 2 hot slabs. This is undertaken to
bend the slab back to the correct shape. If this is not possible then the slab can be cut back
until the bend is below 70 mm out o f flatness. However, care should be taken to make sure
that slabs do not drop below their minimum specification for processing (see table 1).

1.1.2.2 Round

The round is the order to which the slab will be processed through the Hot Mill between
planned roll changes. Rounds attempt to maximize the surface and dimensional quality whilst
reducing the number o f work roll changes in the rolling mills to maximize profit.

The slabs in the round have rules governing how they can be positioned in the round
depending on width, gauge and grade considerations. The slabs planned for the round will be
checked for quality issues prior to being ordered and sent to the Furnace.

1.1,3 Furnace

After leaving the slab yard, slabs are placed on the slab charging roller tables and will be
given final visual checks o f their dimensions prior to entering the Furnace. The slabs are
moved to the charge end o f the Furnace and enter the Furnace when appropriate space is
available in Furnace A or B.

The Furnaces are long insulated rectangular metal boxes, with burners to control temperature.
The slabs are moved through the Furnace using the walking beam. The Furnace contains a
mixture o f fixed beams (to hold the slab) and walking beams (to move the slab). The walking
beams are lifted using hydraulic cylinders and rollers. The most important aspect o f slab
Furnace movement is that the slabs are moved in a controlled way. If control over the slab is
lost then slabs could interfere with each other, damage the Furnace wall, or in extreme cases
get stuck in the Furnace.

Slabs are heated to the required temperature (1180 - 1240 °C) depending on the dimensions
and chemistry. The temperature o f the slab is not measured in the Furnace due to the thick
layer o f scale preventing accurate measurements from being taken. Therefore the Furnace
temperature is back calculated from the Roughing Mill after the Furnace scale has been
removed and an accurate reading can be taken.

1.1.4 Scale Breaker and Reversing Rougher Mill

After leaving the Reheat Furnace the slab proceeds into the scale breaker which will remove
the scale created in the Furnace. Without a scale breaker there would be a dramatic increase
in the number o f primary scale defects on the strip, and would also impact on surface quality
by damaging rolls.

The Roughing Mill reduces the gauge o f the incoming slab to between 35 mm and 45 mm,
which is known as a transfer bar. The transfer bar is spread or squeezed by the edger to the
required width o f the coil and some o f the scale formed in the Reheat Furnaces is removed.
The Hot Mill in Port Talbot has a Reversing Rougher hence several passes are necessary to
achieve the required transfer bar thickness.

1.1.5 Coil Box

After the Reversing Rougher Mill the slab is then held at the coil box, the purpose o f which is
to reduce the size o f the mill and to increase the temperature o f the transfer bar before it
enters the Finishing Mill. When the coiling is complete the outer lap o f the coil is fed into the
Crop Shears. After the coil box the top/bottom and head/tail ends are switched.

1.1.6 Finishing Mill

The purpose o f the Finishing Mill (FM) is to reduce the dimensions o f the transfer bar to the
thickness that is required by the customer. This is achieved by passing the transfer bar
through a series o f close-coupled mill stands that have a reducing roll gap.

1.1.7 Finishing Mill Entry

The Finishing Mill is a vital area o f the Hot Mill when considering tertiary scale formation
(also called D-type scale). An overview o f the Finishing Mill entry is in figure 2. The purpose
o f the entire scale suppression system in the front section o f the Finishing Mill is to condition
the transfer bar so that scale is not formed.

Finishing Mill
Coil Box Interstand cooling

Crop
Shear

Figure 2. Finishing Mill Entry.

1.1.7.1 Crop Shear

After leaving the coil box the transfer bar enters the Crop Shear. The Crop Shear is a large
shear spanning the width o f the transfer bar, which will crop the end o f the bar to leave a
straight edge. The purpose o f the Crop Shear is to ensure that the transfer bar is fed correctly
into the Finishing Mill by cutting off the end o f the bar so a flat edge is left.
I
!
|

1.1.7.2 Roll Coolant

Roll coolant is added to the rolls in the Finishing Mill, for several important reasons
including;
1. To protect the work rolls from thermal failure due to the high temperatures and loads
associated with rolling.
2. To control the thermal expansion (thermal crown) o f the work rolls in a parabolic
manner so that a parabolic strip crown and flatness are obtained.
3. To minimise the surface degradation o f the work rolls. This is achieved by
maintaining a stable protective oxide layer on the surface o f the rolls in the early
stands to minimise roll wear.

The incorrect use and control o f the roll coolant can result in a number o f failure modes
including;
• Too much or too little water affecting roll gap profile - poor interstand shape leading
to manual bending interventions - target exit shape and/or crown/profile not obtained.
• Uneven application o f cooling water across the roll barrel length - asymmetric shape
or a ridge on the profile.
• Inadequate volume o f roll coolant - surface temperature o f the roll gets too hot during
rolling generating scale on the roll which can imprint on the strip.

1.1.7.3 Finishing Scale Breaker (FSB)

After the Crop Shears the transfer bar passes through the conditioning systems o f the
Finishing Scale Breakers (FSB). The FSB are used to remove secondary scale formed after
the Roughing Mill and condition the transfer bar to prevent tertiary scale formation. This is
achieved by using high pressure water from the descaling system. The FSB consists o f the
following:
• Two top and two bottom descaling headers, referred to as A and B legs.
• A descaling hood
o Collects and channels rebound water away from the bar.
• FSB top and bottom pinch roll - situated at the entry to the FSB
 o These provide traction through the FSB. The top and bottom rolls need to
make good contact with the transfer bar to prevent water from running back to
the Crop Shear.
■ Reducing temperature at the Crop Shear will increase the resistance to
shearing and increase wear on the Crop Shear.
• Damming roll - situated at the exit o f the FSB
o The top roll is fitted with a compressed air supply thus provide a downward
force to ensure good contact with the strip. This prevents excess water from
pooling on the top surface in the F5 guide area.
• A centrally located under bar driven roller - maintains the pass-line through the FSB.
• External deflector
o Situated between the pinch roll and the first descaling header. This is a series
o f narrow steel fingers attached to a horizontal pole again preventing excess
water escaping.
• A set o f carry-over fingers.

The two pairs o f headers (top and bottom), are all equipped with nozzles at approximately
100 mm apart. The second pair is offset by approximately 50 mm relative to the first, this
evens out the cooling effects and to give full width descaling coverage should a nozzle in one
o f the headers become blocked. Each header is split into a central zone (approximately 1400
mm wide); this is used for the majority o f material processed. For material with a width
greater than 1400 mm a smaller outer section o f the headers is used (approximately 300 mm
at each end). Water reserves need to be carefully managed for high width material to avoid
significant loss o f pressure and flow.

For full width coverage, it is important that the nozzles are aligned correctly to each other.
The correct descaling coverage is only achieved when the nozzles are aligned correctly
allowing full overlap o f sprays. This type o f installation is in danger o f not providing full
descaling and therefore increasing the risk o f defects such as scale.

To ensure that conditioning is achieved each nozzle is set at an angle to the rolling direction.
This is done to ensure that the spray pattern from one nozzle does not interfere with the
adjacent nozzle. The nozzle angle is also important because some angles are more likely to
suffer from clogging o f the nozzle.

It should be noted that if there is excessive undulation o f the transfer bar as it passes through
the FSB, then the change in relative heights/distances will give a different descaling footprint.
Incomplete descaling coverage is possible on the top surface as the standoff distance
decreases. On the bottom surface, the standoff distance increases and more nozzle overlap
results. This is another reason why the headers are offset by 50 mm.

As the nozzles are worn down through use the flow rate will increase, this will produce the
following effects;
• The descaling impact pressure will drop.
o This will reduce the amount o f scale removal.
• The increased flow rates will cause additional cooling.
o Additional cooling could aid in conditioning o f the transfer bar but can cause
other defects such as transverse cuts plus possibly low finishing temperatures.
• The increased overlap o f the water jets will result in additional “double” cooling in
the overlap area which can be seen as dark stripes on the surface exiting the FSB. This
is detrimental for transverse temperature distribution.

In order to control and minimize the effects o f nozzle wear, maximum flow limits are set for
initiating nozzle changes.

During maintenance, a lead test (see figure 3) is undertaken as a check for blocked or
misaligned nozzles and also to ensure that the carryover fingers do not interfere with the
descaling spray pattern. The lead test is a board covered on both sides with lead at the same
thickness as the transfer bar (45 mm) spanning the full width o f the FSB. The FSB is turned
on and then the lead blank is examined to identify misaligned or blocked nozzles. The test
should show that the sprays overlap with each other so that the entire width is covered.
Figure 3. Lead Test, To test the effectiveness of the sprays a wooden plank covered in
iead is inserted and the sprays are activated. The lines im printed on the lead are the
im pact of the sprays. It is desirable to observ e an overlapping pattern.

1.1. 7.4 Water Containment

The FSB headers have various pieces of equipment on either side o f them preventing the
escape of excess descaling water. The presence o f excess water pooling on the transfer bar in
this area can cause:
• An overall loss of temperature.
• Localised temperature variation resulting in gauge spikes.
o Different temperatures will affect how the transfer bar will react when it enters
the Finishing Mill and give different gauges.
• Increases the risk o f transverse cuts.
o Transverse cuts are caused by localized areas of cooling. These areas will have
different mechanical properties than the surrounding area and during further
 processing will damage the surface which is unsuitable for many applications
and can split the coil.

The purpose of the FSB hood is not only to cover the area but to stop descaling rebound
water from falling onto the strip which will impede the impact o f the sprays. To ensure
descaling occurs water is contained within the hood. The hood contains channels to collect
the rebound water and then allow it to flow away at the operator and drive side ends o f the
hood. During the rewinding o f a cobble, the hood should be raised to prevent internal
damage.

1.1.7.5 F5 Pre-Sprays

After the FSB there are the F5 pre-sprays (see figure 2), the purpose o f these are to control
the temperature o f the transfer bar to minimize tertiary scale formation. The conditioning o f
the transfer bar by the F5 pre-sprays can influence a number o f failures including the
following;
• Excess water applied to the strip. This can result in the following;
o Incorrect gauge.
■ Excess water can change the cooling pattern o f the steel which will
influence the steels reaction to the Finishing Mill rolls leading to an
incorrect gauge.
o Low Finishing Mill exit temperature.
■ Due to increased cooling mechanical property failure can occur,
o T-cuts.
■ Caused by localised cooling,
o Head or tail end pinching.
■ Caused by excess water added to the strip.
• Insufficient water applied to the strip, this can result in the following;
o High Finishing Mill exit temperature.
■ Due to decreased cooling mechanical property failure can occur,
o Scale.
■ Insufficient water applied to the strip, causes high temperatures which
leads to scale formation.
After the F5 pre-sprays the transfer bar enters the Finishing Mill. The Finishing Mill consists
o f a series o f high powered mill stands which will reduce the transfer bar into a flat
continuous strip o f the specified thickness. The Finishing Mill also controls the shape and
temperature o f the strip that in turn controls the metallurgical properties achieved by the strip.
The phase transformation will occur when the correct temperature for the chemistry is
reached, this will usually occur on the Run Out Table where the cooling is controlled with a
water cooling system. However, with low temperatures phase transformation can also occur
in the Finishing Mill.

Following the Finishing Mill is the Run Out Table (ROT). The ROT controls the mechanical
properties o f the strip via manipulation o f the phase transformation. This is achieved by
controlling the cooling pattern o f the strip which will affect the microstructure present. After
the Run Out Table the steel is coiled and dispatched to the customer.

7.7.7.6 Finishing Mill Scrubbers and Edge Sprays

The Finishing Mill scrubber system is situated within the guides at the entry to the first three
Finishing Mill stands. There are two main purposes o f the Scrubber system:
1. To condition the surface temperature o f the bar to avoid reaching temperatures where
tertiary scale could be formed which would then get rolled into the strip surface.
2. To provide a degree o f scale removal before each roll bite.

Like the scrubbers the edge sprays are situated at the leading edge o f the entry guides to the
first 3 stands. The purpose o f the edge cooling sprays is to cool the strip surface, along the
strip edges, with the aim o f preventing the surface temperature o f the bar from reaching a
temperature where tertiary scale could be formed.

The edge sprays are required due to the edge section (100 - 150mm in from the edge) having
a higher temperature than the centre surface temperature. A 50 °C increase in edge
temperature is observed compared to the centre line temperature. This temperature difference
is due to heating in the Furnace.
There are a number o f potential failures that are associated with the scrubbers and edge
sprays, which include the following;
• Leak in header or missing nozzle
o Uneven cooling across the strip width causes localised temperature variations.
This can cause a ridge across the profile structure. After cold reduction this
will appear as a shape defect,
o Surface temperature too high in early stands. This can lead to tertiary scale on
the strip surface.
• Incorrect use on Heavy Gauge products.
o Surface temperature is too low due to too much water being applied. This can
result in transverse cuts on the strip surface.

For the edge cooling sprays to work effectively, the surface o f the bar needs to be free o f
water coming from the previous interstand cooling curtains or from badly fitting top exit
strippers. Ineffective cooling will result. This also applies to the effectiveness o f the top
Scrubbers. The air blow-offs, situated on the side o f the mill housing, need to work
effectively to clear the water off the surface o f the strip.

1.1.7.7 Interstand Cooling

The interstand cooling system is located in the top and bottom exit furniture o f the first 4
Finishing stands (F5-F8). The purpose o f the interstand cooling system is to cool the strip as
it is being rolled. It comprises of a series o f continuous full width cooling water curtains
designed to cool the strip as it is being rolled. This is an automatic control system (but can be
put into manual mode) done as part o f the Finishing Mill exit temperature control. It also
conditions the bar so that the formation o f tertiary scale is suppressed, especially at the exit o f
the first 2 stands when the surface temperature o f the product is at its highest.

1.1.8 Parsytec

Parsytec is a computer aided Surface Inspection System (SIS) that is in use in more than 25
rolling mills worldwide. The defect catalogue has been built using experience o f inspection
personnel together with information gathered from the Parsytec system. In the Hot Mill in
Port Talbot the top surface cameras are at the start o f the ROT and the bottom surface is at
the end o f the Hot Mill.

1.1.8.1 Detection

All heterogeneous sections o f the strip are detected by the Parsytec system. Issues with
correct classification can occur due to environmental conditions. The main environmental
conditions in the Hot Mill include water droplets, Parsytec utilises a suppression algorithm to
remove these environmental factors. However, using these suppression algorithms inevitably
has the potential to also suppress real defects which are covered by pseudo defects like water
drops. Furthermore, high levels o f pseudo defects can consume the system’s computing at
high rates. This can result in real defects going unrecognized while the system is busy
processing pseudo defects. Disturbances created by the physical environment can never be
completely eliminated even by the most sophisticated software that Parsytec has to offer. It is
therefore highly beneficial to take measures to protect the surface inspection system and the
inspected strip surface from adverse environmental influences.

It should also be noted that the system does not scan the surface o f the strip in a uniform way.
The system gives preferential treatment to areas that are likely to have roll type defects. This
can result in a periodic pattern for scale appearing on the strip due to this more intensive scan.

1.1.8.2 Classification

Detected sections are classified into features and these features are compared to the Parsytec
catalogue to determine which defect the feature best represents. Problems with classification
arise when defects overlap each other or environmental conditions (water/steam) obscure the
defect. These can make it harder for an accurate judgement to be made and more likely that
the detected area will be classified as a non-classified defect.
Table 3. Parsytec Recognition Rates.

Commissioned System
Initial Performance
Performance
Area Defects Hit Rate Alarm Rate Hit Rate Alarm Rate
Heavy Scale
80% 5%
Medium Scale 90% 3%
Red Scale not included
Table 3 shows the expected accuracy of the Parsytec system from initial performance to when
the system is working fully. The hit rate is the percentage o f coils with a scale defect that are
caught by the Parsytec software. The alarm rate is the percentage o f coils that are incorrectly
classified by the system.

1.1.8.3 Action

The operator uses the Parsytec results to determine an action for the coil. The data is also
communicated to the Finishing Mill pulpit to allow corrections to be made to improve future
runs.

1.1.9 Run Out Table (ROT) and Coiling

There are two functions o f the ROT, the first is to connect the Finishing Mill to the coilers.
The second function o f the ROT is to provide an area for the correct degree o f cooling to
occur so that the required mechanical properties can be achieved.

After the ROT section is the coiling section. The purpose o f the coiling is to coil the strip so
that it can be sent to the customer or for further processing.

1.1.10 Conclusions for Process

The Hot Mill process at Port Talbot has a long history o f suffering from scale related quality
issues, mostly due to its layout and the power o f the Finishing Mill stands. This lack o f power
means that harder grades need to be exposed to higher temperatures to allow them to be
rolled, which results in a greater chance o f scale formation. To be able to discern the impact
o f scale formation on the process, the process o f scale formation needs to be understood.
1.2 Scale

Scale is an oxide layer that builds up on the surface o f the steel when it is exposed to
environmental conditions. Iron oxide can form at any point during the life cycle o f the
product; however it is particularly likely to arise at the Hot Mill. Higher temperatures
increase the overall reaction rate due to greater kinetic energy, this causes a faster chemical
reaction. This is why it is important to gain an understanding o f the scale creation at the Hot
Mill. However, this section will concentrate on understanding scale formation at the Hot Mill
and what factors contribute to its formation.

1.2.1 What is iron oxide made of?

The scale that forms closest to the steel substrate is Wustite (FeO) followed by Magnetite
(Fe3C>4) and Hematite (Fe 2 0 3 ) (see figure 4). The ratio o f the phases can vary considerably
depending on the temperature, time and steel chemistry; a ratio o f 95:4:1 is to be expected
(Wustite:Magnetite:Hematite). Bolt (2003) states that the important properties to consider for
the oxide layer are thickness, composition, adhesion and structure o f the scale formed.

Hematite

Magnetite

Wustite

Steel Substrate

Figure 4. Composition of scale.

The thickness o f the oxide layer will also be determined by its processing conditions. Bolt
(2003) states that the most important parameters for thickness are the finishing temperature
and gauge o f the coil. Depending on the temperature o f the strip, oxide could be formed at the
ROT and also during coiling but that scale formation at these stages is rather modest. This is
supported in the work by Yang et a l (2008), whereby they discuss the influence o f Silicon on
thickness. Yang et a l (2008) investigated steels with Silicon contents ranging from 0.01 -
1.91 wt. % and concluded that oxidation rates and scale thickness decreased with increasing
Silicon. They also concluded that different oxides are formed at different percentages o f
Silicon. This work is important because it demonstrates the importance o f limiting the
chemistry range to avoid scales with different properties forming. Taniguchi et al. (2001)
further demonstrated the importance o f limiting the chemistry range. They found that when
Silicon is at high percentages (1.14 wt. % and above), the penetration depth o f the scale
increases with increasing Silicon. The work o f Munther and Lenard (1999) concluded that
thick scale provides a degree of lubrication to the work rolls, while a thin scale has a higher
coefficient o f friction and is more adherent to the metal substrate, which makes it harder to
remove. Therefore from the work o f Yang et a l (2008) and Taniguchi et a l (2001) it can be
concluded that increasing the Silicon percentage creates a thin adherent layer o f scale which
can a high deree o f penetration into the steel substrate. However, Silicon is not likely to play
an important part in this analysis due to the percentage and limited range present in the steel
investigated.

The composition o f these formed oxides influences the properties o f the scale. The classical
three-layer scale model, as observed in figure 4, reveals that when pure iron is oxidised under
normal conditions, Wustite (FeO) is formed closest to the steel substrate followed by
Magnetite (Fe 3 0 4 ) and Hematite (Fe 2 C>3). Each o f these oxides have unique properties. Bolt
(2003) states that at hot rolling temperatures, Wustite has a high plasticity, but is extremely
brittle at room temperature, which means that a large amount o f Wustite will result in poor
scale cohesion at room temperatures. Hematite is extremely hard and brittle and is highly
detrimental during all stages o f processing due to the increased work roll wear it causes and
the poor strip quality it creates. Magnetite is the phase that is least brittle at room
temperature. This phase can be tolerated in small quantities, because it is not as hard as
Hematite and shows some plasticity at hot rolling temperatures. Thinner scales are more
adherent than thicker scales and the reason for their increased adherence is the higher
amounts o f Hematite and Magnetite, especially closer to the steel substrate. This causes
problems for further processing because Magnetite and Hematite are more adherent than
Wustite, which makes it more difficult to remove.

The adherence o f the scale to the steel is proportional to the composition o f the scale. This is
because Magnetite and Hematite are a lot harder than Wustite and are more difficult to
remove by pickling/scale breakers. This increased tensile strength (see table 4) o f the
Magnetite and Hematite contributes to increased work roll wear and rolled in scale.

Table 4. Typical mechanical properties for scale phases (Bolt (2003)).

Tensile Strength at room temperature

Wustite 0.4 M Pa
Magnetite 40 M P a
Hematite 10 M P a

1.2.2 Influences on Scale

There is no single variable that can be attributed to scale formation in the Hot Mill. To truly
comprehend complex relationships between the variables affecting scale formation the
relationships need to be understood. There are many potentially important variables
including;
1. Mill layout

2. Temperature

3. Dimensions

4. Time between different processes

5. Chemistry

6. Water systems

7. Rolling
The mill layout is o f vital importance to how the different variables will be able to interact
with each other, leading to scale formation. Blazevic (1995, 1996) stated that the distance
between the descaler and Finishing Mill entry has a significant impact on tertiary scale
formation. The shorter the gap between the sprays and the Finishing Mill entry allows less
opportunity for scale formation to occur. Even though the mill layout is o f vital importance
and needs to be considered when reading literature on scale, the aims o f this project are not to
recommend a new mill layout for Port Talbot Hot Mill rather to better understand how the
variables are influencing scale formation. Due to this a more detailed discussion on the
advantages o f different mill layout will not be performed in this review. Instead the Port
Talbot Hot Mill will be discussed.

1.2.4 Temperature

The thickness o f the scale is proportional to temperature and time. Sun et a l (2004) observed
that as temperature and oxidation time increases the scale thickness also increases. It was
observed that for the first 20 seconds the scale increased in thickness according to linear
growth, and then by parabolic growth after that. They explain this observation by the
transport mechanism o f oxidation. Initially the scale is very thin which allows rapid diffusion
o f oxygen with the metal at the iron-air interface. As the scale layer increases in thickness the
transport rate will reduce and the diffusion will obey parabolic laws.

The composition o f the scale that develops is also highly temperature dependent. At
temperatures greater than 570 °C the main composition o f scale is Wustite, which is generally
the thickest layer. This is then followed by Magnetite and Hematite (see figure 4). The
literature on what composition o f scale is observed at different temperatures is highly
variable, mostly due to different Hot Mill layouts, different types o f steel, different
experimental methods, and different temperatures. The literature review o f Bolt (2000)
concludes that as the temperature is increased the more likely it becomes that the scale
formed will be composed of a higher percentage o f Magnetite and Hematite at the expense o f
Wustite.
It should also be noted that the surface temperature o f the transfer bar can be vastly different
from the core temperature o f the coil, especially when the coil interacts with interstand sprays
and work rolls. The surface temperature will rise due to heat flow from the core o f the coil.
The amount o f scale formed depends on the surface temperature at the time, but the surface
temperature is also dependant on the scale thickness. The temperature recorded by the
pyrometers gives a good indicator o f the temperature o f the steel; however it can be
influenced environmental factors. If the scale formed across the coil is not consistent then the
surface temperature will not be consistent.

Thermal stresses caused by thermal expansion due to cooling/heating and metallographic

changes can help break up the scale. For example when austenite transforms to ferrite a 1 %
volume expansion occurs, which can aid in breaking up the scale.

1.2.5 Dimensions

The main influence o f the transfer bar and strip dimensions is the effect this has on
temperature. Thinner gauges are more difficult to keep at the required temperature and
thicker gauges have a large difference in temperature between the centre and surface o f the
coil. The percentage o f gauge reduction will also affect the amount o f temperature in the slab
and how much scale will form.

1.2.6 Time

At a constant temperature the growth o f scale is linear with time until the scale becomes
sufficiently thick. This initial linear growth is when the Wustite is formed. After this
thickness is reached the growth is determined by the diffusion o f iron ions, this is when
Magnetite and Hematite are formed.

1.2.7 Chemistry

The chemistry o f the steel substrate has a strong influence on the type o f scale that will form
in the Hot Mill. The two elements that have the biggest influence on scale properties are
Silicon and Phosphorus. If Silicon and Phosphorus are at levels and temperatures where they
will influence scale then Silicon will have the larger impact.
The distribution of elements in the scale is not uniform. Elements that can be found
throughout the scale include Manganese, Aluminium (to an extent) and Silicon (to an extent).
However, the majority o f elements have a high oxygen affinity and will form oxides that will
accumulate in the scale at the metal/scale interface; these elements include Aluminium,
Silicon, Phosphorus, Boron, Chromium, Molydbenum, Titanium, Niobium, and Zirconium.
More noble elements such as Copper, Nickel and Tin will become enriched in the metal
increasing the adhesion o f the scale.

1.2.7.1 Silicon

The work done by Taniguchi et al. (2001) and by Yang et al. (2008) has discussed some o f
the influences of Silicon on scale. They found that the Silicon decreases the amount o f
oxidation that occurs which in turn reduces the scale thickness, but at high percentages o f
Silicon (> 1.14 wt. %) an oxide enriched with Silicon increases the oxides adhesion to the
steel, making it harder to remove. Increased levels o f Silicon have the effect o f vastly
increasing the oxidation rate especially above 1177 °C. This is due to the temperature o f 1177
°C being the eutectic temperature o f Fe0-Fe2Si04. This causes a catastrophic amount o f
scale to be formed due to the presence o f the liquid phase at the interface and at the grain
boundaries. Overall the scale will be very adherent to the metal. Silicon can also cause red
scale/tiger stripes to form due to an increased metal/scale entanglement. This type o f scale
can be extremely difficult to remove.

Silicon has a large influence on scale, however, the grades under investigation have low
levels o f Silicon and a limited range. Therefore, Silicon is not likely to be observed as a
significant variable.

1.2.7.2 Phosphorus

For tertiary scale formation, Bolt (2000) states that Phosphorus is o f particular importance.
Phosphorus reduces the overall amount o f scale that is able to form on the strip. This is
because Phosphorus becomes enriched at the metal/scale interface which suppresses the
diffusion o f Fe ions. This lack o f Fe ions will not stop the oxidation from occurring, but will
promote the formation o f Magnetite and Hematite over Wustite.
These observations are corroborated by operator’s experiences with re-phosphorised steels
being very sensitive to oxide defects; increased work roll wear and increased frequency o f
work roll changes. This increased wear is due to increased friction caused by the harder scale
(Magnetite and Hematite) being present.

1.2.8 Descaler sprays

After the coil box is the Crop Shear followed by the FSB (see figure 2). The purpose o f the
FSB is to break up scale on the strip before it enters the Finishing Mill. There are descaling
sprays situated prior to F5 (the first Finishing Mill stand) and are used to suppress the
formation o f scale. There are also interstand cooling sprays between each pair o f work rolls.

Silk (2001) investigates descaler effectiveness and determines that the main components o f
successful descaling include the impact pressures used and the angle o f the descaler
conditioning sprays. The slabs then proceed to the Interstand Spray System for further scale
to be removed by this spray. This system is either operated manually or run in automatic
mode. There are many practical problems when using the descalers. However, the most
important one is the higher the flow the more descaling can be achieved but the faster the
temperature is reduced. This loss o f temperature is a significant concern because if the strip
loses too much heat it will not be within its finishing temperature specification and will not
achieve the required mechanical properties. If the temperature o f the strip is lowered further
then the strip might become too cold to be hot rolled and could damage the mill. This is a
particular problem for thin gauges because they have less internal heat to maintain a useable
working temperature.

A more accurate term for descaler sprays is conditioning sprays. As their purpose in reducing
scale formation is to condition the surface o f the strip to not generate scale.

1.2.9 Rolling

M unther et a l investigated the effect o f different scale thicknesses on the friction within the
Hot Mill work rolls. It was discussed that a thicker scale produces a lower degree o f friction
and adds a degree o f lubricity to the rolls; and that a thin scale was more adherent, imparting
a higher friction and was more difficult to remove than the thicker scale. It was determined
that the thickness of scale had a greater influence on friction than the amount o f reduction,
the rolling temperature, and the roll velocity.

There are two types o f rolls that are used in the Finishing Mill; these are high chrome (HiCr)
and high strength steel (HSS) rolls. The HSS rolls are cheaper and can last longer than HiCr
rolls. However, due to an uneven build up o f scale on the HSS rolls there will be an impact
on the scale on the strip. This will result in a poor surface quality o f the strip when HSS rolls
are used instead o f HiCr rolls. Due to high surface demands, some products (such as tinplate)
will only be rolled with HiCr rolls rather than HSS rolls.

1.2.10 Downstream Scale

If scale is left to form and not removed before it reaches its end application it can have a
detrimental effect on surface quality. Figure 5 indicates how scale can appear at the
galvanising line. If scale is left on the surface o f the coil at the coating stage then scale could
cause quality issues including;
• Fleck - Worn work roll surfaces on finishing stands pick up scale from strip and
impress pattern back onto strip.
• Jet - Scale formed in Reheat Furnaces or between Roughing stands - not removed by
water sprays due to inadequate water pressure or blocked jets.
• Pits - Small piece(s) o f scale not washed off between finishing stands that pickle out
leaving deep pit(s).
• Rolled in - Single pieces o f scale, rolled into the strip during hot rolling and not
pickled out.
Figure 5. Scale as observed from the galvanising line.

1.2.11 Conclusions to Scale section

Scale has a detrimental effect on surface quality especially for downstream applications. The
formation o f scale is influenced by a number o f factors including temperature, time and
chemistry. These factors and those that influence them need to be considered when selecting
the dataset and performing the analysis.

1.2.12 References

D.T. Blazevic, Rough surface - The major problem on thin cast hot strip mills, Spring
Convention Assoc. Iron & Steel Eng., Salt Lake City, USA (1995).

D.T. Blazevic, Tertiary rolled in scale - The hot strip mill surface problem o f the 1990's, in:
37th Mechanical Working and Steel Processing Conf. Proc., Vol. 33, Hamilton (Canada)
1995, Iron & Steel Society, Warrendale 1996, p. 33-38.
Bolt, H., 2000. The properties o f oxide scales on hot rolled steels: A literature review,. Corns
Research, Development & Technology, Ijmuiden Technology Centre.

Bolt, H., June 2003. Understanding the properties o f oxide scale on hot rolled strip, in
METEC Congress 03/3rd European Rolling Conference (ERC), Dusseldorf.

Munther, P. and Lenard, J., 1999. The effect o f scaling on interfacial friction in hot rolling o f
steels. Journal o f Materials Processing Technology, 88, 105-113.

Silk, N., October 2001. The practical aspects o f hydraulic de-scaling. Steel Times
International, 38-44.

Sun, W., Tieu, A. K., Jiang, Z., Zhu, H., Lu, C., 2004. Oxide scales growth o f low-carbon
steel at high tmperatures. Journal o f Materials Processing Technology, 155-156, 1300-1306.

Taniguchi, S. Yamamoto, K., Megumi, D., Shibata, T., 2001. Characteristics o f

scale/substrate interface area o f Si-cntaining low-carbon steels at high temperatures.
Materials Science and Engineering A, 308, 250-257.

Yang, Y., Yang, C., Lin, S., Chen, C., and Tsai, W., 2008. Effects o f Si and its content on the
scale formation on hot-rolled steel strips. Materials Chemistry and Physics, 112, 566-571.
1.3 D ata M ining

Data mining is a valuable set o f techniques which are utilised for improving the
understanding of a system. This is achieved by understanding the system through expert
knowledge, with the aim o f collecting relevant and usable data sources so that they can be
interrogated to give a greater understanding o f the process. Data mining techniques are used
in many different sectors, including the steel industry.

Data mining is a powerful method o f getting useful information from a dataset. There are
however several factors which need to be considered when performing any data mining
technique.
• Badly formatted data
o This can give unexplained or inaccurate results.
• Non standard classification terms and methods
o An example o f this is the location o f defects in a coil. Some systems will mark
the defect at its centre while others will log the edge position. If the difference
is not noted then similarities and differences in the data can be missed, and
inaccurate results will be acted on.
• Lack of detailed information on the process.
o Without the having the important variables for the system, links can be made
between data that cannot be explained. Not understanding the relationships
between how the data interacts with other data will result in poor results.
• Black-box techniques.
o Some techniques are difficult to interrogate and understand how decisions
have been made. If the results come with conclusions that are expensive to
implement or go against past practices, then having an explainable reason will
help the argument. An example o f a black-box technique is an artificial neural
network.
• Accuracy of the data.
o If the accuracy o f the data is not known then the accuracy o f the results can be
called into question.
1.3.1 Directed and Undirected Techniques

One method o f classifying data mining techniques is to classify them into directed or
undirected methods. Directed and undirected are also referred to as hypothesis testing and
knowledge discovery. Hypothesis testing is a top-down approach, with the aim being to
prove/disprove a given theory. Knowledge discovery is a bottom-up approach with the
intention being to interrogate the data to find out something unknown.

There is a specific process for both directed and undirected methods. First the undirected
method will be explained.
• Identify data availability and reliability, to do this attention needs to be given to the
following
o Understanding o f how data is acquired (sensors, probes, manual/automatic
recording systems)
o Find out how the data is stored and the best method o f extracting it
o Investigate if errors in data alignment can occur when the data is retrieved
• Format data for analysis
o Ensure the data can be used for the chosen data mining technique
• Create a model and use a training data set to train the model
• Use an evaluation data set to evaluate the accuracy o f the model
• Use the model on live data to show similarities in the data
• Decide which similarities require further investigation
• Perform the required direct method to get further information from the data
• Investigate another hypothesis

The directed method is as follows;

• Arrange the data so that it is clustered into relevant classes
• Having the data clustered by an undirected method is standard
• An example o f preclassified data is splitting process data from the galvanising line
into classes for annealing code or grade
• Format data for analysis
• Ensure the data can be used for the chosen data mining technique
• Create a model and use a training data set to train the model
• Use an evaluation data set to evaluate the accuracy o f the model
 • Use the model on live data to show similarities in the data
A more detailed review is provided for specific data mining techniques.

1.3.2 Cluster Detection

Clustering is used to determine similarities between the data; it does this by assigning groups
to the data depending on their variables. This is an example o f an undirected data mining
technique.

1.3.3 Binomial Regression

This type o f regression can take a variety o f input values and outputs them as 1 which equals
success or 0 which equals failure. There are many different types o f binomial regression,
including logistic regression.

By fitting data to a logistic curve the probability o f an event occuring can be predicted, this is
called Logistic Regression. The predictor variables used can be either numerical or
categorical. This model’s strength comes from the fact that any value for the input can be fed
into the model while the outputed value is between 0 and 1. The input value represents the
factors that influence the probability o f the data.

The probit model is used in binomial regression as a fast way o f determining the maximum
likely estimates for a model.

1.3.4 Classification and Regression Trees (CART)

CART uses a decision tree method for defining data into sets. The main advantage o f
decision trees is that it is easy to explain the results and work out which parameters affect the
process. Due to the ease in explaining the rules, they can be expressed easily in a SQL search,
which makes it easy to add new records. CART is an example o f a directed data mining
technique and is particularly good at classification o f data.

CART partitions the inputted data using a series o f rules. These rules usually consist o f
interrogating the data to split it on the grounds o f if it is greater than or less than a given
value.

1.3.4.1 Pruning

Pruning is performed to improve the efficiency o f the model. Pruning takes into account the
accuracy and size o f the model, with the intention o f reducing the size but keeping as much o f
the accuracy as possible.

As the data is split generally the probability o f misclassification1 (R*(T)) will reduce. When a
2 • • • • ♦ •
node contains only one value, it is classified as that value and will give a misclassification
result as a zero. However, in practice there are an optimum number o f nodes to be used.
Having a large number of nodes will result in a decrease in the R" value. Increasing the
number o f nodes beyond the optimum value will result in a detrimental impact on the
prediction o f any new datasets. This is due to overtraining o f the training dataset, which
means that the model is too specific to the dataset to give an accurate prediction o f unseen
datasets. This can result in the tree with a large number o f nodes having a lower R" value than
a tree with the optimum nodes. A tree with fewer nodes will not usually have enough
classification power to be a useful as a predicting tool.

The optimum number o f nodes can be reached by one o f two ways. The first method is to
grow a tree that is far too big and then prune it back to reach the optimum value. The second
method is to work out a better method o f predicting the R*(T) value. A good value o f R*(T)
can be obtained by using a test sample of a larger sample.

1.3.5 Artificial Neural Networks (ANN)

Artificial neural networks are based on the neural connections present in the brain. The neural
network uses a training set to generalize relationships for the data, and then it is used to
predict general data sets. Neural networks have two weaknesses. The first is their sensitivity
to the given format o f the data it is used on. A change in format can give vastly different

1When R*(T)=1=100 % misclassification, R*(T)=0=100 % classified

2A node is the result of a split.
results; so a vast amount o f time is spent formatting the data and removing anomalous results
prior to analysis. The other weakness is the lack o f description in the results output by the
neural network. Neural networks can be applied to undirected data mining techniques in the
form o f Self Organising Maps (SOM's).

SOM's output their results in the form o f a 2D map. Further analysis o f the map can be
carried out using clustering, to attain information on the different properties.

1.3.6 M ultivariate Adaptive Regression Splines (MARS)

MARS is a non-parametric regression analysis that is used to improve the accuracy o f linear
models. There are many benefits and disadvantages to the MARS technique including the
improved flexibility over the linear regression technique; however it is more difficult to prove
that the MARS model is more accurate than it is for the linear model. Explanation o f results
is easier for MARS models than neural networks, because they are represented graphically
and the knots can be explained using the rules that created them. Another advantage over
neural networks is the data does not have to be prepared before it is used, due to the hinge
functions automatically partitioning the data.

1.3.6.1 Hinge Function

The addition o f a hinge function allows the MARS model to partition data in a similar way to
CART models. This partitioned data can be treated independently from each other which is
beneficial for inputs that act differently under changing conditions. This is achieved within
the hinge function with the addition o f a constant that is usually referred to as a knot. The
hinge function is usually set to zero for part o f its range and will change when the conditions
are met. A simple hinge function will often look like this:

max(0,x-c) or max(0,c-x)

c=constant/knot

When the hinge function achieves its knot value it is activated and the data is partitioned.
MARS is used for both categorical and numerical data, but the hinges often give improved
accuracy for numerical data. The use o f large datasets is common with MARS models, but
some other techniques can obtain the required information faster, such as CART.

1.3.7 Principle Component Analysis (PCA)

The purpose o f PCA is to turn variables into linararity uncorrelated variables called principle
components. This is useful if the dimensionality o f the dataset needs to be reduced but the
variability needs to be maintained.

1.3.8 Partial Least Squares (PLS)

The PLS method has many elements in common with PCA. The main difference is that rather
than trying to reduce the variance it finds a linear regression of the data by projecting the
observed and predicted values in to a new space.

1.3.9 Genetic Algorithms

Genetic algorithms were developed around the idea o f using the principles o f natural
selection to identify the optimum set of parameters to work out the best result. The model is
run several times and the least predictive results are removed each time, so that the best set o f
parameters is converged on as the program is run. Genetic algorithms are often used to
improve other techniques such as neural networks. This is an example o f a directed data
mining technique.

1.3.10 Multinomial Logistical Regression (MLR)

Instead o f a binary model that has two possible outcomes, 1 and 0. A multinomial model can
have multiple outcomes, which can (in some cases) be used to more accurately represent the
dataset.
 1.3.11 Adaptive Neural Fuzzy Inference System (ANFIS)

The ANFIS model combines elements from neural networks and fuzzy logic principles.

1.3.12 Conclusions to Data Mining

There are many different techniques used in data mining, all with their own strengths and
weaknesses. The key factor in successful data mining is to understand the data which is
available, and to be able to perform a successful analysis from that data.

Whatever statistical approach is used to model the formation o f scale at the Hot Mill, it must
address the issues associated with co-linearity and measurement error. One approach is to
directly model the count data collected by the Parsytec system, i.e. the recorded number o f
defects on the top (and or bottom) surface o f the coil. Traditionally, this is done using a
Poisson distribution with the mean o f the distribution being made a linear function o f the
process variables. For the data set collected for this thesis this is not straightforward because
this scale count data has many “missing” values - for example none o f the coils in the data set
had scale counts in the range 991 to 1002 (there are no values within this range). Indeed there
are many other such missing value counts in this data set (over the recorded range o f 0 - 4000
counts). To use the Poisson model would therefore require a form o f truncation and
consequent loss of information. A nonlinear variation o f this approach would model the mean
o f the distribution using, for example, a neural network.

The following chapters will explain the methodology behind several different data mining
techniques. The methods that will be included encompass the following;
• Partial Least Squares (PLS)
• Logit
• Principle Component Analysis (PCA)
• Multinomial Logistical Regression (MLR)
• Adaptive Neuro Inference Fuzzy Systems (ANFIS)

Due to the inherent strengths and weaknesses o f different data mining techniques it is likely
that multiple techniques will need to be employed to achieve the best results.

Trmnr'iFrimiriimm-i ~mnr rji r mm t«—■'•m rTirriD-ni m i TniriiM ir < in ~~riwrrirmm tuttth ■ i r r m~r< ~~<niin i »iiinvnT»ibTOim— n w ~ f m ir n r tTiimiiTirtTTi'iriTT»r-TTn i~nwrtw w n i > v nn r nnir ~fiirrrri~miT limn—t i'THi~nrTifiTri rr v r m mwim n*i li t

33 1
Chapter 2. Aims of Research Project

This project aims to use a collection of data mining techniques to better understand the
influence o f Hot Mill process parameters on scale formation at Port Talbot Hot Strip Mill in
South Wales. This understanding will then be used to recommend operating conditions
designed to minimise the formation o f scale at the Hot Mill.

The dataset to which these data mining techniques were applied was carefully chosen to
reduce process variation. There are several main factors that were considered to minimise this
variability including time period, grade and gauge investigated. The following data mining
techniques were chosen to investigate this dataset;
• Partial Least Squares (PLS)
• Logit Analysis
• Principle Component Analysis (PCA)
• Multinomial Logistical Regression (MLR)
• Adaptive Neuro Inference Fuzzy Systems (ANFIS)

These data mining techniques were chosen because they represent different ways o f
investigating the dataset. Some o f the differences include diverse ways to represent the
dependant variable used in the analysis. Three different representations were used binary,
multinomial and continuous.
Chapter 3. Experimental Techniques

3.1 Data Collection

The purpose of this section was to understand how data was collected from the available
process parameters, with the intention o f better understanding their influence on scale
formation by using data mining techniques.

3.1.1 Considerations

To be able to achieve a successful data mining investigation, the data sampled needs to be
carefully selected to be able to accurately represent the process being investigated. The main
factors that need to be considered for scale formation include;
• A large dataset
o Many o f the data mining techniques require large datasets to be effective
• A comparable dataset
o The cases in the dataset investigated need to be comparable to each other
• A clean dataset
o Data recording anomalies need to be removed, these include;
■ Probe recording errors
• Comparable time period
o Due to changes and improvements being implemented to the process o f the
Hot Mill, longer time periods may not be comparable for the duration o f the
time period.
• Representative independent variables
o The variables selected to represent the process
• Representative dependent variables
o A variable will need to be chosen to represent the effect o f the process
variables on scale.
• Processing
o The material needs to be processed the same way, with particular attention to;
■ Gauge and grade
• Different gauges and grades will react and be reacted to
differently when being processed in the Hot Mill
 • HSS/HiCr rolls will react differently to rolling over time and
could give different results.
• Schedule
o It has been noticed that the coils processed in the Hot Mill are less stable when
the gauge or grade is changed
• Application
o The Dataset needs to represent material for the same application so the results
are meaningful

The above problems were overcome by using data associated with coils processed for
Tinplate applications. The tinplate products have several advantages over other products
including;
• The tinplate product is intended for surface critical application so scale levels must be
low.
• All the coils in the dataset have the same chemistry requirements.
• To minimise the effect o f different heating/cooling cycles, only a single gauge o f 2.1
mm has been investigated.
• This is a high quantity product type so a lot o f material is processed, and they are
processed in their own rounds.

Coil data was collected between September 2009 and March 2010 for all the process
variables described above and table 5 gives some descriptive statistics for this sample o f data,
together with the full name o f each process variable and its short hand descriptor - Xj. The
sample size was n = 1577.
Table 5. Process Variables and their Sample Means and Standard Deviations.

Process Variable xj Mean Standard Deviation

Wt. % Carbon xl 0.069 0.0042
Wt.% Manganese x2 0.485 0.0226
Wt. % Phosphorus x3 0.014 0.0031
Wt. % Sulphur x4 0.014 0.003
Wt. % Silicon x5 0.005 0.0032
Wt. % Copper x6 0.016 0.0059
W t.% Nickel x7 0.011 0.003
Wt. % Chromium x8 0.018 0.004
Wt. % Aluminium x9 0.034 0.0047
Wt. % Tin xlO 0.003 0.0018
Wt. % Nitrogen xl 1 0.0125 0.0009
Wt. % solAl x l2 0.031 0.0043
Maximum RM Exit Temperature, °C x l3 1121.75 17.87
Minimum RM Exit Temperature, °C x l4 918.34 11.97
Average RM Exit Temperature, °C xl5 1099.43 16.18
Maximum CS Temperature, °C x l6 1099.88 19.89
Minimum CS Temperature, °C x l7 951.69 48.59
Average CS Temperature, °C x l8 1069.55 20.8
Maximum FM Exit Temperature, °C x l9 907.16 7.18
Minimum FM Exit Temperature, °C x20 829.87 28.96
Average FM Exit Temperature, °C x21 881.05 5.75
ROT Temperature, °C x22 731.35 25.13
Maximum Coiling Temperature, °C x23 693.133 28.3786
Minimum Coiling Temperature, °C x24 438.354 53.8953
Average Coiling Temperature, °C x25 561.627 2.7173
Furnace Aim Temperature - Average
Temperature, °C x26 4.77933 14.587
Furnace Centre/Surface Difference, °C x27 44.9575 24.2802
Furnace Surface Temperature, °C x28 1227.13 19.5941
Furnace Centre Temperature, °C x29 1182.48 17.7393
Scrubber Pressure (set) x30 1.748 0.6603
 ISS (set) x31 -0.692 0.4616
Thread Speed x32 11.370 0.1772
Slab Length (m) x33 9.677154 0.470653
Slab Width (m) x34 0.95601 0.047791
Slab Gauge (m) x35 0.235172 0.003213
Slab Weight (tonnes) x36 17.05425 1.163756
Transfer Length (m) x37 61.01447 4.611045
Transfer Gauge (m) x38 0.040309 0.002432
Coil Width (m) x39 906.4052 52.28183
Top Scale Count x40 126.200 219.9389
Bottom Scale Count x41 250.008 402.6597
Indictor Variable Y 2.48193 1.09386
Scale Count Indicator (Binary) YB 0.327 0.4693
RM = Roughing Mill, CS = Crop Shear, FM = Finishing Mill, ROT = Run Out
Table, ISS = Interstand Spray System (-1 = Manual and 0 = Automatic), Srubber
pressure (0 = low pressure, 1 = medium pressure, 2 = high pressure). YB = 1 when
the scale count on the bottom o f the sample exceeds 200 and zero otherwise.

The data collected for this investigation was obtained from Port Talbot Hot Mill. The
temperature data was collected from temperature sensors at various points along the Hot Mill.
The temperature sensor is positioned in the centre o f each section o f the hot mill and records
the temperature along the entire length o f the coil. The temperature data is converted into
maximum, average, and minimum for each coil. The chemistries o f the coil are measured at
the secondary steel making process stage. The scale count is the absolute bottom scale count
from the Parsytec inspection system.

3.2 Statistical Methods

This chapter will lay out the methodology used in the Logit, PCA, PLS, MLR and ANFIS
models used in the analysis o f the dataset described in the data collection section.
3.2.1 Binary Statistical methodology

The Logit, PCA and PLS all use a binary split to describe the dependent variable o f the
bottom scale count. The dependent variable takes on a value o f unity when the scale count on
the bottom o f table 5 exceeds 200 and zero otherwise. The cut o f 200 was chosen for two
main reasons. The first reason was that scale counts lower than this are unlikely to be a
problem for further processing but as the scale count increases above 200 it becomes more
difficult to determine if the scale will cause a problem for further processing without
understanding how adherent the scale is. The second reason is that a split o f 200 makes the
population o f bottom surface affected coils represent more than 30 % o f the dataset which
makes it easier to be confident about the output o f the models, i.e. it creates a more balanced
data set from which to apply non-linear Partial Least Squares.
Chapter 4. Identification, for control, of the process parameters influencing tertiary
scale formation at the hot strip mill using a binary choice model

4.1 Principal component analysis

Principal component analysis (PCA), see Jolliffe (2002) for a good review o f this technique,
is an extensively used technique o f multivariate linear data analysis. The purpose of PCA is
to reduce the dimensionality o f the data but still maintain its variability and to remove the
correlation amongst the process variables used to control the Hot Mill process.

The principal components (PC) are calculated as follows. The p process variables (xi to xp)
are standardised to have a zero mean and unit variance to avoid the problems associated with
units o f measurement. Then p linear combinations o f these variables can be formed as
follows

PC1 — T &12^2 "T *“ CL-±pZp

PC2 = &2lz l T a 22Z2 T a 2pZp

PCp CLpiZ-i + CLp2Z2 T- CLppZp

Eq. (4.1a)

where the z variables are the standardised values o f the Xj (j = p) process variables shown in
table 5. The so called loadings, an to aip, are then chosen so as to maximise the variance o f
PCi subject to the normalising condition (the purpose o f the normalising condition is to
maintain the same sum o f loadings squared across all o f the principle components):

ah + a\ 2 + ••• + = 1
Eq. (4.1b)

PCj is then stated to be the first principal component and is a linear function o f the z ’s (and
thus the process variables) that has the highest variance. PC 2 then has the next highest
variance and so on until all the variation in the z ’s is picked up by the p principal
components. That is
variance(PCi) + variance(PC 2 ) + .... + variance(PCp) = variance(zi) + v a ria n c e ^ ) + .... +
variance(zp) = p

with

variance(PCi) > variance(PC 2 ) > ... > variance(PCp)

Various different algorithms can be used to find values for these loadings, ranging from the
simple summation method put forward by Burt (1945) (of which a good review can be found
in Childs (1970)) to algorithms that calculate a spectral decomposition o f the correlation
matrix amongst all the x variables (see the text by Mardia et al. (1979)). The Eigen values
from this decomposition measure the variation in all the process variables explained by each
principal component. Thus the first Eigen values from this decomposition measures the
variation in all the process variables explained by the first principal component. The Eigen
Vector from this decomposition contains the required loadings for each principal component.
By construction these principal components are also all orthogonal and hence completely
uncorrelated with each other. The spectral decomposition method will be used for this
Principle Component Analysis.

Further, Cattell (1952) has suggested that only those principal components having an Eigen
value greater than 1 should be considered as essential and therefore retained in the analysis.
This simple rule allows for a substantial reduction in the number o f variables included in the
final statistical model.

4.2 Binary choice models

4.2,1 A generalised framework

A binary choice model views the outcome o f a discrete incident (such as the occurrence o f a
large amount o f scale) as a reflection of an underlying regression. The cost associated with
any additional processing or downgrading o f a heavily scaled coil (defined above as 200+
observations o f scale) is modelled using an unobserved variable yj* such that
 V

y \ = A) P j X y + £ = x p + £i
7= 1

Eq. (4.2a)

tfi
where xy is the i measurement on process variable j associated with the Hot Mill and the pj
are the parameters that are to be estimated from the data set. P is a vector containing the p+1
parameters o f this model and x is the data matrix with the columns o f this matrix containing
each o f the process variables (including an initial column o f l ’s) and each row contains the i*
measurement on these variables. e\ is assumed to follow some standardised distribution with
known variance - to be discussed further below. The cost o f having large amounts o f scale is
not directly observed, just whether the coil is heavily scaled or not. Therefore observations at
the Hot Mill take the form

yj = l if y ; > 0 Eq. (4.2b)

y rO if y f < 0

Provided that a constant (po) is included in Eq. (4.2a) the assumption o f a zero threshold for
yi* is an innocent one. The process variables are then taken to determine the probability o f y**
>0

Prob(yi* > 0 | x) =Prob(yi = 1 J x) = F(x, p) Eq. (4.2c)

Prob(yj* < 0 | x) =Prob(yj = 0 | x) = 1 - F(x, p)

where F(x, P) is a function, dependent on x and p, that determines the probability that yj = 1
given the values in x, with the natural requirement that F(x, p) should tend to 1 as x;p tends to
infinity (and F(x, p) tends to zero as x;p tends to minus infinity). There are many cumulative
density functions that fit these criteria, but most applications o f this type o f model are
confined to the use o f the normal distribution (giving rise to the Probit model - see Maddala
(1991) for a good introduction to this model) and the logistic distribution (giving rise to the
Logit model - see Maddala (1991) for a god introduction to Logit model).
The following model can be seen as a generalisation o f these two binary choice models that
enables discrimination between these and other functions through use o f standard likelihood
ratio tests. From Eq. (4.2a-c) it follows that

f V p
\

■*3
1

I-
Prob(yi = 1) = Pt = Prob< S i > - Po T ^ ' Pj%ij

II
Po "F ^ ' Pj%ij
< j =1 j =1 J

'
Eq. (4.3a)

Eq. (4.3a) makes it clear that F is the cumulative distribution function for e. One very general
specification can be derived from the generalised F distribution. Prentice (1975) has
suggested a non-degenerate and standardised version o f this distribution whereby the
probability density function is given by

/v \ (m1/m
IXlJIi / 2)m
F i exp(m
CAU 1111,,w
W )I
f(w) = J , r/ , w m>) Eq. (4.3b)
{T(m, )T(m 2)}/r(m, + m 2) (i + {ffl| / m i J
where

r
Wi = — Po + ^ t Pj- ij
j =i
Eq. (4.3c)

and where f() is the derivative o f F() with respect to w and T() is the gamma function. In Eq.
(4.3a-c), the parameters mi and m 2 determine the cumulative distribution function for Wj and
thus 8j. That is, mj and m 2 determine the form o f the binary choice model such as the Logit
and Probit model. Unfortunately, there is no closed form expression for F() within this setting
except for some special cases. To take some well-known examples, first consider the case
where mj = m 2 = 1. Then

exp(Wi )
1 - F[Wi] =
1 + exp(w i)
Eq. (4.4a)

[M'Wfc'TiffiMi'nirTi'iwnwjJi ■inmT"n» nnrHrrys«j«MJ*r,ftr(Tii-in-rr‘irfifftb<'(r-wrBn—m~ii w i w— < II < ■IIn iiin n m n n in i w m i i r t r v m m n itmmmnronrfiTrTHTrr m i

431
which is the cumulative (standardised) logistic function giving rise to the Logit model which
can also be expressed as the log odds model

p
Pi 1
Ln Po "b / f Pj%ij
ll - p j
j =1
Eq. (4.4b)

For finite mi and m2 but with mj = m 2 this logistic function becomes the generalised logistic.
Then when mi = 1 and m 2 = 0

1 - F(wi ) = 1 - exp[—exp(wi)]
Eq. (4.5a)

which is the cumulative (standardised) extreme value function giving rise to the proportional
hazards model which can also be expressed as

L n [-L n ( 1 - P*)] = -
J= l

Eq. (4.5b)

Other special cases include the probit model when mi = 00 and m 2 = 00 and the generalised
gamma cumulative function when just m 2 = 00 (see Lawless (2003) for a descriptions o f this
function).

Whilst there are no closed form expression for the integral o f Eq. (4.3b) when mi and m 2 are
not restricted to the special cases outlined above, values for this integral at a given value for
Wj can be obtained numerically. If mi and m 2 are both finite, values for this integral at a given
value for Wj are found by calculating areas under the Beta distribution with mi and m 2
degrees o f freedom. More precisely,
 t'
Pi = 1 - F Po + ^ Pj- ij
;= i
Eq. (4.6a)

where B(sj,m 1,1112) is the area under the Beta distribution with mi and m 2 degrees o f freedom
to the left o f Sj, where

m2

Eq. (4.6b)

For finite mi and m 2 = 00, values for this integral at a given value for Wj are found by
calculating areas under the Chi distribution with 2mj degrees o f freedom. More precisely,

r'

Po + Pj- ij = x(2v it 2m i)
J= 1

Eq. (4.6c)

where x(2vi,2mi) is the area under the Chi distribution with (2mi) degrees o f freedom to the
left o f 2 v j , where

Wi
Vi = m ^exp

Eq. (4.6d)

The integrals calculated using Eq. (4.6c, 4.6d) tend to those obtained from the standard
normal distribution (and hence yields the Probit model) as mi tends to 00. For finite m 2 and mi
= 00, the procedure is to replace Vj with -Vj and to replace mi with m 2 in Eq. (4.6c, 4.6d). Most
statistical packages have routines capable o f calculating such areas and as demonstrated by
Prentice (1975) this is best achieved by reparametrising mi and m 2 as

p = (m1+ m 2) 1 and q = (m -1
] 1+, m 2l)(m11+. m 2')
-I x -l
Eq. (4.6e)

45 !
4.3 Parameter estimation

Since the y\ are realisations of a binomial process with probabilities given by Eq. (4.6a or
4.6c) we can write the likelihood function as

yi=l y t= 0

Eq. (4.7a)

so that the log likelihood is given by

n n
log[L] = Y y p o g i P t ) + ^ ( l - y d l o g O - - Pi)
i=l i=l

Eq. (4.7b)

where n is the sample size o f the data set and Pj is calculated from Eq. (4.6a or 4.6c) for given
values o f mi and m 2 . Given these values for mi and m 2 , the p parameters in Eq. (4.6a, 4.6c)
are chosen so as to maximise the log likelihood in Eq. (4.7b).

4.4 Model identification

Taking a general to specific approach to model identification, all the process variables shown
in table 5 are initially included in the model (either individually or as part o f a principal
component). Then a grid search is carried out at different values for mi and m 2 using this
general model. The log likelihood defined by Eq. (4.7b) at each (mi, m 2 ) combination is
stored and the largest o f them recorded. Call this statistic log[L]max. Clearly the (mi, m 2 )
value combination associated with this value determines the distribution most supported by
the data. However, any (mi, m 2 ) combination that produces

Ar 2 [ lo g ( L ) max l o g ( L ) r\
Eq. (4.8a)
with a value less than 3.84 is also supported by the data at the 5 % significance level. log[L]r
th
is the log likelihood associated with the r (mi, m 2 ) combination. This follows from the fact
that Ar, under the null hypothesis that the rth (mi, m 2) combination is the true combination, is
asymptotically Chi square distributed with 1 degree o f freedom.

Once the correct values for mi and m 2 are determined in the way described above, a data
based simplification o f the model then takes place. The model associated with log(L)max is
called the unrestricted model. If the j th explanatory variable is then removed from this model,
a restricted model is obtained with a log likelihood given by log(L)R. This is achieved by
removing the Xj with the smallest student t statistic. The final simplified model is obtained by
removing all those variables for which

Ar = ~2 \l o g ( L ) max
Eq. (4.8b)

has a value in below o f 3.14. Finally, the adequacy o f this parsimonious model can be
assessed in one o f two ways. First the proportion o f correct predictions can be calculated,
where a prediction is considered correct if Eq. (4.6a or 4.6c) yields a value above 0.5 when yj
= 1 or alternatively when Eq. (4.6a or 4.6c) yields a value equal to or below 0.5 when yj = 0.
The count R defined in Eq. (4.8c) below then lies between 0 and 1 with better performing
models having a value closer to 1.

2 number o f correct predictions

Count =
total number o f observations
Eq. (4.8c)

4.5 Results and evaluation

4.5,1 Principal component testing procedure

With the use o f principal component analysis the correlations existing between the process
variables were investigated. The analysis was able to reduce both the number o f process
variables and remove the correlations among the process variables used in the model. This
allowed the model to predict the scale count and from this to gain an understanding o f how to
run the Hot Mill so as to minimise scale formation. The maximum correlation coefficient
between all the chemistry variables was 0.47 and this was between Copper and Nickel.
Further, most of the correlation coefficients were less than 0.1 in absolute value. These
relatively low correlations mean that there is little benefit to be gained from forming the
principal components o f these chemistries. However, table 6 shows that all the temperature
variables are highly correlated - especially for example between the maximum and average
Rougher Mill temperatures ( x b and X 15) . In general, it can be seen from table 6 that the most
highly correlated relationships occurred between the maximum and average temperatures at
the Rougher Mill and at the Crop Shear.

Table 6. Correlation matrix for the temperature process variables. A value of 1

represents a perfect correlation and a value of 0 represents zero correlation. A negative
value shows an interaction in the opposite direction.

X]3 Xl4 Xl 5 Xl6 Xl 7 Xi8 Xl9 X20 X21 X22

Xl 3 1
Xj 4 0.04 1

Xl 5 0.92 0.13 1

Xi6 0.76 0.73 -0.01 1

Xl 7 0.3 0.26 0.04 0.38 1

Xl8 0.53 0.57 0.01 0.69 0.3 1

X]9 0.46 0.42 -0.03 0.61 0.46 0.23 1

X20 0.16 0.12 0.03 0.18 0.12 0.13 0.16 1

X2 l 0.48 0.44 -0.04 0.63 0.51 0.26 0.7 0.25 1

X22 0.06 -0.03 0.02 0.02 0.06 -0.01 -0.01 0.08 0.02 1

Given these high levels o f dependency between the temperature process variables, principal
components o f these temperatures need to be constructed to remove these correlations.
Table 7 shows how much o f the variability in all the temperature variables shown in table 5
are accounted for by the principal components.

49 j
Table 7. Principle component results for the temperature process variables.

Components Eigen values Variation, % Cumulative Variation, %

PC, 4.201 42.01 42.01
PC2 1.106 11.06 53.07
PC3 1.043 10.43 63.51
PC4 0.9741 9.74 73.25
PC5 0.8418 8.42 81.67
PC6 0.7769 7.77 89.44

PC7 0.4999 5.00 94.44

PC8 0.2805 2.81 97.24

PC9 0.2040 2.04 99.28
PC ,0 0.0718 0.72 100.00

The first three principal components accounted (see table 7) for nearly 64 % o f the variability
present in all ten temperature process variables, whilst the first four principal components
accounted for nearly three quarters o f the variability present in all ten temperature process
variables. Using Cattell’s rule (1952) o f only needing to use those principal components
having an Eigen value o f around unity or more, the ten temperature process variables can be
replaced by the first three or four principal components. These principal components had the
form

PCi = -0.0233(x13 - 1121.8) - 0.0016(x14 - 918.3) - 0.0251(x15 - 1099.4)

- 0.0223(x16 - 1099.9) - 0 .0045(x17 - 951.7) - 0.0179(*18 - 1069.6)
- 0.0493(x19 - 907.2) - 0 .0045(x20 - 829.9) - 0.0644(x21 ~ 881.1)
- 0.0003(^:22 - 731.4)
Eq. (4.9a)

PC2 = 0.0156(x13 - 1121.8) + 0.043 1(jc14 - 918.3) + 0 .0214(x15 - 1099.4)

+ 0.0008(x16 - 1099.9) + 0 .0008(x17 - 951.7) - 0.0007(x18 - 1069.6)
- 0.0384(x19 - 907.2) - 0.0 1 1 6 (x 20 - 829.9) - 0.0622(x21 - 881.1)
- 0.0 185(x22 - 731.4)
Eq. (4.9b)
 PC3 = —0.0037(x13 - 1121.8) + 0.0597(x14 - 918.3) + 0.0005(x15 - 1099.4)
- 0.0027( x16 - 1099.9) + 0.0022(x17 - 951.7) + 0.0003(*18 - 1069.6)
- 0.0149( x19 - 907.2) + 0.0116feo “ 829.9) - 0.0007(*2i - 881.1)
+ 0.0235(>22 - 731.4)
Eq. (4.9c)

PC4 = 0.0043Ois ~ H 21.8) + 0.0085O14 - 918.3) + 0.0080O15 - 1099.4)

+ 0.0040O16 - 1099.9) - 0.0054O17 - 951.7) + 0.0078O18 - 1069.6)
- 0.0047O19 - 907.2) - 0.0249O20 - 829.9) - 0.00 86 O 2i “ 881.1)
+ 0.0233O22 - 731.4)
Eq. (4.9d)

The loading coefficients in front o f each variable show how heavily weighted each variable is
within each principle component. If the loading coefficients are similar , then instead o f
creating principle components a simpler alternative would be to take their sums (if the
loadings are positive) or differences (if the loadings are negative). Thus in the first principal
component, the loadings in front of the maximum Rougher Mill (x,3), the average Rougher
Mill (x, 5) and the maximum Crop Shear (xi6) temperatures are very similar, suggesting that
an appropriate dimensionality reduction would be to look at the differences between these
three temperatures. The loadings in front o f the minimum Rougher Mill (x,4) and the average
Crop Shear (x,8) temperatures are very similar, suggesting that an appropriate dimensionality
reduction would be to look at the differences between these two temperatures. The loadings
in front of the minimum Crop Shear (x,7) and the minimum Finishing Mill (x2o) temperatures
are very similar, suggesting that an appropriate dimensionality reduction would be to look at
the differences between these two temperatures. The loadings in front o f the maximum (x,9)
and average Finishing Mill (x2,) temperatures are also very similar, suggesting that an
appropriate dimensionality reduction would be to look at the differences between these two
temperatures.

4.5.2 Distribution analysis

The p parameters in Eq. (4.6a and 4.6c) were estimated by maximising the log likelihood in
Eq. (4.7b) at given values for p and q (with p and q defined by Eq. (4.6e)). All the chemistry
variables in the table 5 were included in Eq. (4.6a and 4.6c), and all the temperature variables
were replaced by the first three principal components specified in the previous section. Figure
6 shows the maximised log likelihoods associated with different values for p and q. It can be
seen from figure 6 that the largest maximised log likelihood occurs where p = 1 and q = 0.5.
The area o f the graph below the dashed line in figure 6 shows which maximised log
likelihoods are significantly different from this one at the 5 % significance level using the test
statistic in Eq. (4.8a). It is clear from figure 6 that both the Logit and Probit models are
therefore supported by the data and have log likelihoods very close to the largest log
likelihood. Consequently, and to keep the computational aspects o f the work simpler, all
further results shown below are based on the Logit model.

-558
Binary choice model most spported by the data
Probit model
-560

'O
§ -562 5 % significance level for values o f p
and q supported by the data Logit
model
b£ *564

a«

-568

-570
0.25 0.5 0.75
p (as defined in Eq. (4.6e))

Figure 6. The maximised log likelihood obtained at different values for p and q.

4.6 Logit model results

Using the data based simplification procedure described in Section 4.4 above, a simplified
Logit model was derived and this model is summarised in table 8. Only a few chemistries
were statistically significant at the 1 % significance level, with these being Phosphorous,
Silicon, Copper, Nickel and Chromium. Further, only 2 o f the above derived principal

52
components were statistically significant at this level. Further, o f the n = 1577 observation o f
y, 1140 o f these were predicted correctly by the model. This is an overall correct prediction
rate o f 73 %. However, the model was significantly better at predicting the Hot Mill
conditions leading to no significant scale formation (i.e. yj = 0), as the model predicts when yj
= 0, some 98 % o f the time.

Table 8. Logit model results.

Process Variable Coefficient Std.Error t-value

Constant -2.49 0.46 -5.38
Percentage o f Phosphorus (X 3 ) 228.28 25.13 9.08
Percentage o f Silicon (X 5 ) -88.87 22.25 -3.99
Percentage o f Copper (X 6 ) 72.66 13.33 5.45
Percentage o f Nickel (X 7 ) -184.66 28.79 -6.41
Percentage o f Chromium (xg) -64.12 19.74 -3.25
PC, -0.391 0.04 -10.3
PC2 -0.232 0.07 -3.32

From table 8 it can be observed that each principal component has a negative impact on the
log odds ratio as does the amount of Silicon, Nickel and Chromium. Phosphorous and Copper
on the other hand have a positive impact on this ratio. For an indication o f which variables
have the biggest effect on the probability o f scale forming at the Hot Mill it is better to
convert the coefficient values in table 8 into quasi - elasticities. These elasticities, contained
in table 9, show the percentage change in the probability o f scale forming following a
percentage change in each o f the process variables. These elasticities are calculated around
the mean values - shown in table 5 for each process variable as the elasticity will vary with
the magnitude o f each process variable.

53
Table 9. Quasi-elasticities.

Process Variable Quasi - Elasticity

Percentage o f Phosphorus (X 3 ) 0.62
Percentage o f Silicon (X 5 ) -0.09
Percentage o f Copper (X 6 ) 0.24
Percentage o f Nickel (X 7 ) -0.39
Percentage o f Chromium (xg) -0.22
PC, -0.08
PC2 -0.05

From table 8 it can be seen that when the Phosphorous content is at 0.014 % wt (i.e. at its
mean value shown in table 5) a further 1 % increase (0.00014 % wt) in the Phosphorous
content will increase the probability o f scale formation by just over 0.6 percentage points. It
can also be observed from table 9 that when the Nickel content is at 0.01 % weight (i.e. at its
mean value shown in table 5) a further 1 % increase in the Nickel content will decrease the
probability o f scale formation by just under 0.4 percentage points. Given the size o f the other
elasticities in table 9, these are therefore the most important chemistries for controlling scale
formation at the Hot Mill, in that changes in these two chemistries will bring forth the biggest
change in the probability o f scale forming when the temperature variables are at their mean
values shown in table 5.

As for the temperature variables, it can be seen from table 9 that when the first principal
component is at its mean value a further 1 % increase in this component will decrease the
probability o f scale formation by just under 0.1 percentage points. The temperatures in the
first principle component with negative loadings (see Eq. (4.9a) above for PC,), have a
positive impact o f the probability o f scale forming. This relationship is looked at in more
detail below. Prior to this it is important to look at some probability plots as the above
elasticities only apply at the mean values for each process variable. To have a full
understanding, the complete range o f values must be looked at and this is shown in the
probability plots o f figure 7.
 2 0.9 m 0.9
0.8 I 0.8
<S
« 0.7 0.7
| 0.6 0.6
| 0.5 8 0.5
I 0.4
«3
0.3 ■S 0.3
0.2 £ 0.2
2 0.1

0 0.005 0.01 0.015 0.02 0.025

0.005 0.01 0.015 0.02
Phosphorous content, %wt Silicon content, %wt

2P09 0.9
1 0.8
« 0.7 <2 0.7
U
0.6 8 0.6
8 0.5 s- 0.5
8
S 0.4
I °'4
£0 0.3
I*
a 0.2
« 0.1
A A
A
m
1 o £
0 0.01 0.02 0.03
( 0.04 cl,
0 0.005 0.01 0.015 0.02 0.025
Copper content, %wt
Nickel content, %wt

u 0.9 .S 09
| 0.8 5 u .o
1 08
c:
| 0.7 a 0.7
3
1 0.6 “ 0.6
B S
3 0.5
IB
s 05
0.4 .1
Iss 09 0.4
V- 0.3 o 0.3
e
sJ&O
0.2 0.2
A

A 01 A
o*

£
£ CL,
a.
0.005 0.01 -10 ■5 0 5 10 15 20 25
Chromium content, % wt First principal component, PCI

1100
m 0.9
s 1000
I 0.8
| 0.7 900
§ 0.6 800
1 0.5
| 0.4 700

£ 0.3 600
e
& 0.2
3A 500

400
-4 ■2 0 2 4 6
Second principal component, PC2 Temperature variables at the hot mill

Figure 7. Probability of scale formation with variations in the significant process

variables. The vertical line of a - e indicates the mean value for that variable.
In figure 7 the probabilities shown on the vertical axis are calculated from Eq. (4.6a and 4.6c)
by varying one o f the process variables whilst holding the other process variables at their
mean values. These probabilities can also be interpreted as the proportion o f coils
manufactured that have a significant number o f flaws - so preventing them being used for the
high value added markets. Plots like those shown in figure 7 can therefore be used for
optimising the process, i.e. finding how to operate the Hot Mill so that only a small
proportion o f manufactured coils have scale counts that prevent them from being used in high
end applications. For example, to keep the percentage o f defective coils below 10 %, figure
7a reveals the Phosphorous content must be below 0.011 % wt, when running the Hot Mill at
the mean values for all the other process variables (see table 5 for these mean values). Again,
to keep the percentage o f defective coils below 10 %, figure 7d reveals the Nickel content
must be higher than 0.013 % wt., when running the Hot Mill at the mean values for all the
other process variables. As a final example, figure 7 f shows that to keep the percentage o f
defective coils below 10 %, the first principal component must be below 1.9 when running
the Hot Mill at the mean values for all the other process variables.

Using Eq. (4.9a) and PCi = 1.9, a particular Hot Mill temperature can be calculated assuming
the process runs at the mean values for all the other temperature variables at the Hot Mill.
These calculated temperatures are shown in figure 7h. It can be seen that to keep the
percentage o f defective coils below 10 %, the maximum Rougher Mill temperature must be
kept at 1040 °C or less. When looking at the average Rougher Mill temperature the
temperature should be 1025 °C or less. The corresponding temperatures for the maximum,
average and minimum Crop Shear temperatures are 1020 °C, 970 °C and 540 °C respectively.
The corresponding temperatures for the maximum, average and minimum Finishing Mill
temperatures are 870 °C, 852 °C and 410 °C respectively. No conditions could be found for
the minimum Rougher Mill temperatures or the Run Out Table temperature when all other
temperatures are at the mean values shown in table 5.

This type o f calculation can be done for any PCi value, and therefore for any frequency o f
significant scale formation to indicate what temperatures to run the Hot Mill at to achieve this
required frequency o f scale formation. The final impression to be gleamed from figure 7 is
that the variables to keep under strict control, so as to remove the possibility o f very high
proportions o f coils being produced with a significant number o f defects (probabilities above
50 % for example), are Phosphorous, Copper and the various temperatures as summarised in
the first principal component.

4.7 Logit Conclusions

The principal component analysis above showed that the temperatures at the Hot Mill were
all highly correlated. However, nearly 75 % o f the variation in the ten Hot Mill temperatures
could be accounted for by just four principal components that were completely uncorrelated
with each other. Both the Logit and Probit models were supported by the collected data set
and when a Logit model was estimated, only the first two principal components were
statistically significant at the 1 % significance level, together with just five chemistries -
Phosphorous, Silicon, Copper, Nickel and Chromium. The Logit model also had a correct
prediction rate o f 73 %,

It appears from this analysis that the easiest way to reduce scale formation is to lower the
temperature entering the Finishing Mill. It was observed that average Crop Shear
temperatures greater than 1050 °C have a significantly higher percentage chance o f forming
scale at levels greater than a count o f 200 on the bottom surface. Therefore, reducing the
average temperature entering the Finishing Mill will reduce the amount o f scale formed on
the surface o f the coil.

The most practical way o f achieving this is to reduce the amount o f variability o f the
temperature. This reduction o f the standard deviation avoids the issue o f failure to achieve
finishing temperature requirements for mechanical properties. With increased temperature
control the average temperature could be reduced with no detrimental impact.

4.8 References

Burt, C.L., 1945. How the Mind Works. Allen & Unwin, London.
Cattell, R.B., 1952. Factor Analysis. Harper, New York.
Childs, D., 1970. The Essentials o f Factor Analysis. Holt, Rinehart and Winston, London.
Jolliffe, I.T., 2002. Principal Component Analysis, 2nd edition. Springer-Verlag, New York.
Lawless, J.F., 2003. Statistical Models and Methods for Lifetime Data, 2nd edition.
John Wiley & Sons, New Jersey.

57
Maddala, G.S., 1991. Introduction to Econometrics, 3rd edition. John Wiley & Sons
Ltd., New York, p. 322.
Mardia, K.V., Kent, J.T., Bibby, M., 1979. Multivariate Analysis. Academic Press,
London.
Prentice, R.L., 1975. Discrimination among some parametric models. Biometrika 62,
607.
Chapter 5. A Partial Least Squares Generalized Linear Regression Algorithm

The Partial Least Squares (PLS) regression method in its basic form applies for one single
response variable y and is non-iterative. It is particularly useful when the Xj variables are
closely correlated with each other. Marx (1996) proposed a generalization o f the PLS
algorithm to generalized linear models (see McCullagh and Nelder (1989) for an excellent
account o f generalized linear models). This approach is very useful because the binary logit
model can be given a generalised linear regression representation. Marx (1996) used the fact
that, in the context o f the exponential family, maximum likelihood estimates are obtained by
an iterative weighted least squares procedure. M arx’s approach consisted o f replacing the
iterative weighted least squares step by a sequence o f PLS regressions.

5.1 The Generalised Linear Model

Suppose that a sample o f i = 1 to n specimens have been scanned for scale formation on the
ROT o f the Hot Mill. Then let yj be the binary response that equals unity when significant
scale (defined as a scale count for the Parsytec system in excess o f 200) is detected on the
bottom o f the tested specimen and zero otherwise. Next let nt be the probability o f significant
scale formation and also let xii,..,xPj be all the explanatory or process variables described in
table 5. Finally, let g(7ij) be the link function. Then the generalized linear regression model
has the following form:-

Eq. (5.1a)

where

Eq. (5.1b)

59
and where there are p process variables. The model parameters, pj, can be estimated using the
iterative weighted least squares procedure put forward by McCullagh and Nelder (1989). This
procedure involves the creation o f an adjusted dependent variable Zi with components

'dr)i
zt = m + ( y i - .djii.

Eq. (5.2a)

If then, the following (nxn) diagonal matrix o f weights is calculated

(dm) 2
\d n j)
W = diag

Eq. (5.2b)

then the following standard weighted least squares formula can be used to estimate the Pj

p = (.x Tw x y 1x Tw z
Eq. (5.2c)

where X is a (nxp) matrix o f the process variables x i; ...., xp such that the first column o f X
contains the n values for xj and the last column o f X contains the n values for xp. XT is the
transpose o f X and Z is an (nxl) vector containing the n values for Zj as defined by Eq. (5.2a).
P is a (pxl) vector containing the estimates for pi to Pj. This has to be an iterative solution

where initial guesses for pi to pj are inserted into Eq. (5.1a-b) to enable values for z\ and W to

be calculated from Eq. (5.2a-b). Eq. (5.2c) then gives revised or improved estimates for pi to

Pj. These revised values are then used to recalculate values for z\ and W from which further

revised estimates for pi to Pj are obtained. This iterative process continues until the changes

in all the Pj values are very small.

60 j
5.2 Partial Least Squares within a Generalised Linear Model

Within this setting, the following Partial Least Squares (PLS) generalized linear regression
algorithm can be stated as follows. This algorithm follows closely that given by Bastien and
Tenenhaus (2005). Let Xo be the matrix containing the n values for the standardized input
variables xou,...., xop

Xji Xj
j = l ,p and i = 1, n

Eq. (5.3)

where xj is the sample average for the n values on variable Xj and Sj the corresponding sample
standard deviation (the last two columns o f table 5 show these two statistics for each process
variable). Then, the first column o f Xo contains the n values for xoi, and the last column o f Xo
contains the n values for xop. Within this setting the PLS components can be derived as
follows:

5.2.1 Determination o f the first PLS component tn:

1. For each j = 1 to p, compute the regression coefficient pij o f xoj in the generalized
linear regression model given by Eq. (5.1a-b) o f yj on xojj using the iteratively weighted least
squares procedure described above.
2. For each j = 1 to p, compute the standard error, seq, o f the regression coefficient Py
of xoj in the generalized linear regression model given by Eq. (5.1a-b) o f y\ on xoji using the
iteratively weighted least squares procedure described above. Then compute the student t
statistic vjj = Pij/ seij.

3. Form the vector Pi = (P n ,...,P ip) , and a (pxp) diagonal matrix o f weights \ \ =

diag(vij) and the (pxl) vector V]pi.

4. Compute the vector ti = X oVi P j. ti is a vector made up o f the n values o f tu and

constitutes the first PLS component.

This PLS component is essentially a weighted average o f the p predictors o f yj, namely a
weighted average o f the PijjXqji series. A cruder but simpler version o f this is to remove from
pi those coefficients that are found to be statistically insignificant at the 1 % significance
level using the student t test, and o f course the corresponding process variable from Xo, when
forming the vector ti. (The justification being that vij will be small if the student t value for
pij = 0 is less than the 1 % significance level).

5.2.2 Determination o f the second PLS component t2 i:

1. Compute the residuals from the set o f linear regressions o f each column o f Xo on ti
and store these residuals in the matrix Xj. (For example, xij would be the first column in Xi
and would contain the n residuals derived from the linear regression o f xon on the tu.)
2. For each j = 1 to p, compute the regression coefficient P2j o f xjj in the generalized
linear regression model o f yi on tu and xijj using the iteratively weighted least squares
procedure described above.
3. For each j = 1 to p, compute the standard error, se2j, o f the regression coefficient p 2j
o f xjj in the generalized linear regression model o f Eq. (5.1a-b) o f yj on xyi using the
iteratively weighted least squares procedure described above. Then compute the student t

statistic V2j = p 2j/ se2j.

4. Form the vector P2 = (P 21 ,---,p2P)T , a (pxp) diagonal matrix o f weights V2 =
diag(v 2j) and the (pxl) vector V2 P2 .

5. Compute the vector t 2 = X 1V2 P2 . t 2 is a vector made up o f the n values o f t2 i. It is

essentially a weighted average o f the p predictors o f yj given in step 1. Again a cruder but
simpler version o f this is to simply delete from P2 those coefficients that are found to be
statistically insignificant at the 1 % significance level using the standard student t test, and the
corresponding transformed process variable from Xj, when forming the vector t2 .

5.2.3 Determination o f the other PLS components:

This procedure is iterated for the other PLS components thi- At each step the generalized
linear regression o f yj on components tij,..., thi is carried out. The procedure is stopped, and
the component tm+ij, not included in the model if that component is not statistically significant
at the 5 % significance level using the standard student t test. The final regression equation is
obtained by expressing the generalized linear regression o f yj on tn ,..., tmj and this can also be
written as a function o f the original variables by making use o f the standardisation procedure
given in Eq. (5.3) and the means and standard deviations given in table 5. In the case o f
ordinary multiple regression, this algorithm gives the usual PLS regression when there is no
missing data.

5.3 PLS Results and discussion

5.3.1 The PLS Components and the PLS Logit Model

Table 10 shows the estimated values for the pij weights o f the first PLS components
associated with all the different alloying elements. (See table 5 for a description o f the
shorthand used for each o f the process variables listed in table 10).

Table 10. Estimated values for the Pn weights for all alloying elements.
Process Logit slope Elasticity (%) Count R (%)
variables coefficient, pij

Xl -0.0298 [-0.59] -0.24 49.08 (54.26, 46.56)

X2 -0.0807 [-1.60] -0.87 50.35 (57.56, 46.84)
X3 0.6371 [11.15]# 1.43 65.57 (56.98, 69.75)
X4 0.1065 [2.11] 0.25 57.64 (46.32, 63.18)
X5 0.1713 [3.31]* 0.14 60.88 (35.47, 73.23)
X6 -0.0585 [-1.16] -0.08 47.69 (58.91,42.22)
X7 -0.2122 [-4.14]" -0.39 53.52 (45.54, 57.40)
Xg -0.0937 [-1.85] -0.21 50.86 (53.88, 49.39)
x9 -0.1892 [-3.71]* -0.68 56.50 (54.84, 57.30)

XlO -0.2360 [-4.50]* -0.19 48.19(75.58, 34.87)

Xll -0.0507 [-1.00] -0.35 48.19(50.39, 47.13)

Xl2 -0.1935 [-3.80]* -0.7 54.85 (57.56, 53.53)

Student t value for testing the null hypothesis that Pij = 0 shown in parenthesis.
# Coefficient is significantly different from zero at the 1 % significance level.
Elasticity measures the percentage change in the probability o f observing scale
resulting from a one percentage point change in the value o f process variable Xj.

This elasticity is calculated using n = 0.5 and using the values for Xj at this
probability. Count R shows the percentage number o f correct predictions which
is further broken down, in brackets, into the percentage number o f y\ = 1 and y\
= 0 correct predictions respectively.
Only the slope coefficients in front o f the elements P, Si, Ni, Al, Sn and solAl (X3 , xs, X7, x9)
xio, X12) are statistically significant at the 1 % significance level. O f these elements, increases
in the amount o f P, S and Si ( X 3 s X4 5 X 5 ) lead to an increase in the probability o f scale
formation, whilst increases in the other elements had a negative impact on scale formation.
Further, o f these statistically significant elements only P (and possible Al and solAl) seems to
<■%

have predictive power for both yj = 0 and yi = 1 (as measured by Count R ) above that given
by purely random predictions. The estimated elasticities are however quite small (especially
when compared to the temperature variables in table 11 below), with for example a 1 %
increase in Phosphorous content bringing forth a 1.43 % increase in the chances o f significant
scale forming (when this probability is already at 0.5).

Table 11. Estimated values for the Pij weights for various temperatures.
Process Logit slope Elasticity
Count R2 (%)
variables coefficient, pij (%)
Xl3 0.7274 [11.84]" 22.83 63.28 (72.67, 58.72)
X 14 0.0299 [0.59] 1.18 55.49 (38.57, 63.72)
Xl5 0.8002 [12.80]* 27.18 64.24 (72.67, 60.14)

Xl6 0.7175 [11.74]* 19.84 64.55 (71.51,61.17)

Xl7 0.2140 [3.98]* 2.1 53.77 (68.99, 46.37)
X]g 0.7465 [10.76]* 19.19 64.23 (75.00, 59.00)
X J9 0.5033 [8.85]* 31.78 61.00 (55.81,63.52)

X20 0.2205 [4.29]* 3.16 49.78 (71.51,39.21)

X21 0.5851 [11.44]* 44.85 65.38 (66.67, 64.75)

X22 -0.2130 [-3.88]* -3.1 54.60 (41.28,61.07)

Student t value for testing the null hypothesis that P ij = 0 shown in parenthesis.
# Coefficient is significantly different from zero at the 1 % significance level.
Elasticity measures the percentage change in the probability o f observing scale
resulting from a one percentage point change in the value o f process variable Xj.

This elasticity is calculated using n = 0.5 and using the values for Xj at this
probability. Count R shows the percentage number o f correct predictions which
is further broken down, in brackets, to the percentage number o f yj = 1 and yj =
0 correct predictions respectively.
Table 11 shows the estimated values for the pij weights o f the first PLS components
associated with all the different temperatures. Only the minimum Rougher Mill temperature
appears to be statistically insignificant at the 1 % significance level and all but the average
Run Out Table temperature has the expected sign - with increases in temperature bringing
forth an increase in the probability o f significant scale forming. The average Finishing Mill
temperature appears to have the biggest effect, with a 1 % increase in this temperature
bringing forth a 44.85 % increase in the chances o f significant scale forming (when this
probability is already at 0.5). In terms o f the Count R values, the average Rougher and
Finishing Mill temperatures have the best predictive capabilities for both yj = 1 and yj = 0.

Using the algorithm described in Section 5.2 above and the values shown in table 10 and
table 11, the first PLS component, written out in terms o f the standardised values for the
process variables, can be computed as follows

tu = —0.0176xoli —0.1291^O2i 4- 7.1037*031 + 0.2247*O4i + 0.5670*Osi “ 0.0679*O6i

—0.8785*o7i —0.1734*o8i —0.70 19*09i ~ l-062*0iof “ 0.0507*On i
—0.7353*oi2i + 8.6124*oi3i + 0.0176*0i4i + 10.2427*0isi + 8.4235*0i6i
4- 0.8517*oi7i + 8.0323*oi8i 4- 4.4542*0i9i + 0.9459*O2oi + 6.69 3 5*02ii
— 0.6603*o22i

Eq. (5.4a)

The first PLS component, written out in terms o f the process variables in their original units,
can also be computed as follows

tu = —3743.73 —4.09*!i—5.71*2i + 2258.17*3j + 75.07*4i + 175.02*5; —11.51*6i

—293.87*7i — 43.21*8i “ 149.65*9; - 575.44*10j —55.94*11[
—170.32*12j + 0.4819*13i0.0015*14i + 0.6329*15j + 0.4235*16i
4* 0.0175*i7j + 0.3862*18j + 0.6201* i 9( + 0.0327*2oi 4* 1.1647*2ij
— 0 .0 2 6 3 * 2 2 i

Eq. (5.4b)
Then when tu is used instead o f all the process variables, the following PLS logit model is

obtained using iterativly weighted least squares

Ui
rji = In 0 .0 2 8 6 ^ [15.69] Count R 2 = 70.32 %
Ll - 7T|.

Eq. (5.4c)

The student t value given in parenthesis shows that this PLS component is significantly

different from zero at the 1 % significance level and so is an important predictor variable o f

significant scale formation. This PLS model predicts the formation o f scale or lack o f scale

with approximately 70 % accuracy, which given the very large variability in the data is a

good outcome. Further, the Count R2 associated with yj = 1 suggests that the model can

predict the processing conditions that lead to excessive scale with 80.43 % accuracy (but the

Count R2 associated with yj = 0 is lower at 65.41 %).

Using the above algorithm, the next PLS component using the standardised process variables
was identified as

t2i = —0.1729*1:li —0.0207*12i + 6.29*13j + 0.2287*14i + 0.1464*15j —0.1043*16;

- 0.4650*17j —0.2017*18j —0.4807*19i —0.5822*11Oj - 0.0011*1:L1j
- 0.4740*112j —3.1391*113i + 0.00001*114i —0.7965*11Si - 5.1901*116i
- 0.3543*117i —0.3668*118i - 0.8308*119j + 0.03 5 7*120i - 0.0562*121i
- 0.7789*122j

Eq. (5.5a)

or in terms o f the process variables in their original units this can be re-written as
 t2i = —74.36 —41.14*^-2.07*2* + 2458.89*3j + 91.61*4i + 80.62*5i —20.03*6i
- 215.05*7^ - 59.07*8j —132.79*^ - 431.94*10i - 12.51*ni
- 144.26*12i —0.0781*13i + 0.0003*14i + 0.0789*15i —0.1752*16i
- 0.0037*17i + 0.0605*18j + 0.0099*19i + 0.0078*2oi + 0.2260*21j
- 0.03 63*22i

Eq. (5.5b)

The extended PLS logit model encompassing this extra component is then estimated by

iterative weighted least squares to be

Ui
T]( = In = 0.0315U ; + 0 .0 8 8 6 t2i [15.69] [11.44] Count R 2 = 73.30 %
.1 — Ui.

Eq. (5.5c)

The student t values, given in parenthesis, shows that both these PLS components are

significantly different from zero at the 1 % significance level and so are an important

predictor variables o f scale formation. This PLS model predicts the formation o f scale or lack

o f scale with approximately 73 % accuracy and so the addition o f the second PLS provides an

additional 3 % to overall predictive accuracy o f the previous model given by Eq. (5.4c).

Further, the Count R2 associated with yj = 1 suggests that the model can predict the

processing conditions that lead to excessive scale formation with 84.69 % accuracy (but the

Count R2 associated with yj = 0 is lower at 67.77 %).

Using the above algorithm, the third PLS components was identified as

Ml ■■■■III 11fT| 1 1Mil LI I L 1 r 'l' r ^ r *. .................................. ..

67 j
t3i = —0.5834*21i ~ 0.05 06*22i — 2.1281*23l- + 0.0101*24* + 0.3162*25i + 0.0005*26i
— 0.3368*27* “ 1.0641*28i — 1.0751*29j — 0.1006*2i 0* — 0.0129*2u*
— 1.0759*2i2i — 0.0041*213* - 0.0049*2:L4j + 0.0886*2i 5* + 0.1006*2i 6*
— 0.0473*217* + 0.00001*218* — 0.3424*2i9* + 0.0761*22o* ” 0.0035*22i *
— 0.3305*222i

Eq. (5.6a)

with the extended PLS logit model

TCi
Tji = In = 0.0316 tu + 0.0888 t2i + 0 .0 3 2 1 t3i [15.98] [11.45] [1.61] Count R 2
Ll - n t\
= 73.50 %

Eq. (5.6b)

The student t values given in parenthesis shows that the first two PLS components are

significantly different from zero at the 1 % significance level, but the third component is only

statistically significant at the 10 % significance level. Further, this PLS logit model only has

an additional 0.2 % overall predictive accuracy compared to that given by Eq. (5.5c). Taken

in conjunction, these two facts suggest that the PLS logit model with just two components is

a valid parsimonious model, i.e. the model given by Eq. (5.5c).

5.4 Analysis and Interpretation

Figure 8 plots the scale indicator variable yj against the PLS component tu. Also shown are
the predicted probabilities o f scale formation given by the above parsimonious PLS logit
model. These predictions are consistent with the spread o f y\ values shown on figure 8. When
the second PLS component is zero, scale formation becomes more likely than not when the
first principle component exceeds zero in value. At this value for tu, a change in t 2 i has a big
impact on the probability o f scale formation.
 0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

-200 -150 -100 -50 100 150 200

PLS component ti;

Figure 8. Predicted Probabilities of Scale as a Function of the PLS Components.

Figure 9 plots the first two PLS components against each other with the data points shaded or
un-shaded depending upon whether the scale indicator variable is zero or unity. It is clear
from figure 9 that samples with significant scale formation tend to correspond to those with
higher t u values - more specifically with positive tn values. Variations in the values for t 2 i are
a far less indicative o f whether scale has formed or not. These positive tu values can be traced
back to the specific operating conditions. Clearly, many different conditions can be identified
to achieve a required rate o f scale formation, but figure 10 and table 12 give an indication o f
what can be achieved.
 30

o
>;
II
20

II
v:
o
10

c
a> 0 \
a
o
a
S
©
o -10
Ifl
Plh -20

-30

-40 H—r- i■ i i i i i i i—|—i—i—i—i—|—i—i—i i —i—i—i " t "■|— i— i i i— r~i— i— r—r

-200 -150 -100 -50 0 50 100 150
PLS component, tu

Figure 9. Cross Plot of the PLS Components.

Figure 10 shows the elasticity’s associated with each process variable as calculated around
the mean values for each process variables shown in table 12. These elasticity’s give the
percentage change in the probability o f scale formation following a 1 % change in that
process variable from its mean value. It can be seen from figure 10 that the most important
process variable appears to be the average Finishing Mill temperature. When this temperature
is increased by 1 % from its mean of 881 °C (i.e. from 881 °C to 890 °C), whilst keeping all
the other process variables at their mean values shown in table 5, the chances o f significant
scale forming as a result o f this change, increases by 25 %, i.e. from 0.5 to 0.63.
Subsequently, the most important variables appear to be the average Rougher Mill
temperature with an elasticity o f approximately 15 %. In importance, this is closely followed
by the average Crop Shear temperature and the maximum Finishing Mill temperature whose
elasticity’s are both approximately 10 %. The only alloying element with an elasticity above
2 % is Phosphorous. With an elasticity o f 2.04 %, when this element is increased by 1 %
from its mean o f 0.014 % (i.e. from 0.014 to 0.01414 %), whilst keeping all the other process
variables at their mean values shown in table 5, the chances o f significant scale forming as a
result of this change, increases by 2.04 %, i.e. from, 0.5 to 0.51. This is clearly a much
smaller effect than that associated with the above mentioned temperature variables.
 26
24
22
20
18
16
14
12

a93 10
w

I M.
x l x2 x3 x4 x5 x6 x7 x8 x9 x lO xll x l 2 x l 3 x l 4 x l 5 x n ) x l 7 x l 8 x l9 x 2 0 x 2 1 x&l

Process Variables

Figure 10. Calculated Elasties for Each Process Variable.

Next, table 12 shows what the process variables with the largest elasticity’s should be set at
so as to achieve a required low probability of scale formation when all the other process
variables are set at their average values shown in table 5. The low probabilities looked at in
table 12 are 0.3 down to 0.1. Looking for example at the Finishing Mill temperature (X 2 1 ) , to
reduce the probability o f significant scale forming from 0.5 (when all process variables are
set at the average values shown in table 5) to 0.1, this temperature must be set as low as 842
°C (compared to the sample average value o f 881 °C). The same low probability o f scale
formation can be achieved by lowering the Phosphorous content to 0.0065 % (from the
sample average o f 0.0141 %) whilst holding all the other process variables at the sample
average values shown in table 5. The values calculated in table 12 are theoretical values
calculated using the PLS model, and should be considered in conjunction with other
considerations; such as achieving the required mechanical properties and to avoid situations
such as ferritic rolling.
Table 12. Examples of using the PLS logit model for process control.
V ariable/Probability X3 (%) X13 (°C) X l5 ( ° C ) X 1 8 (°C ) X 1 9 (°C ) X 2 l(°C )

0.1 0.0065 855.8 1017.8 944.2 799.5 842.3

0.2 0.0093 953.9 1047.9 990.5 839.2 856.6
0.3 0.0112 1019.2 1067.9 1021.2 907.2 866.1
Sample averages 0.0141 1121.7 1099.4 1069.5 907.2 881

A final insight into the estimated PLS logit model can be obtained using cumulative
frequency plots. This cumulative frequency plot is constructed by sorting the sample data set,
from lowest to highest, by the variable o f interest and counting how many observations there
are at or below various values for that variable o f interest. This frequency count can then be
plotted alongside the frequency count implied by the probabilities predicted from Eq. (5.5c).
Figure 11 shows the cumulative frequency plot for the average Roughing Mill temperature. It
can be seen that the observed and predicted values show similar trends. The model correctly
predicts when scale starts to become an issue (around 1080 °C) and follows the observed
frequency count until approximately 1120 °C.

600

o Observed Cumulative Frequency

500
Predicted Cumulative Frequency
400

300

200

100

900 950 1000 1050 1100 1150 1200

Average Rougher Mill Exit Temperature, °C

Figure 11. Cumulative Frequency Plot for the Average Rougher Mill Exit Temperature.
Figure 12 shows the cumulative frequency plot for the average Crop Shear temperature. It
can be seen that the observed and predicted values do not correspond to each other in
comparison with the other temperature variables. The predicted results over predict when
scale starts to become an issue (1070 °C instead o f approximately 1060 °C) and follows the
observed line until about 1090 °C.

600
o Observed Cumulative Frequency
500 - — Predicted Cumulative Frequency

400 -
cr

o 300 -

200 -

100 -

800 850 900 950 1000 1050 1100 1150

Average Crop Shear Temperature, °C

Figure 12. Cumulative Frequency Plot for the Average Crop Shear Temperature.

5.5 PLS Conclusions

Using a non-linear PLS logit model, it was found that there exists only two linear PLS
components that were important and statistically significant in determining scale formation in
this data set obtained from the Hot Mill at Tata Steel Port Talbot. The estimated model was
capable o f correctly predicting when yj =1 (i.e. when significant scale formation had
occurred) 85 % o f the time. The good predictive capability was also confirmed by the
cumulative frequency plots, where it was observed that the predictions from the model with
respect to those temperatures with a high elasticity corresponded well with the observed
values. This was especially true for the temperature at which scale is first observed to occur.
The chances o f scale formation also appeared to increase with increasing values o f the first
PLS component. This trend was then translated into suggested best practice operating
conditions. For example, if the Finishing Mill temperature was set at 842 °C, with all other
process variables set at the average values shown in table 5, the model predicted the chances
o f significant scale forming can be kept as low as 10 %. The elasticities associated with the
non linear PLS logit model also showed that the main determinants o f scale formation were
the various temperatures. The elasticities derived from the model showed that the minimum
temperatures have significantly less predictive power than the maximum and average
temperatures at the various stages o f operation within the Hot Mill. The most significant
temperatures for scale formation were the average temperatures at the Rougher and Finishing
Mills, as well as the average Crop Shear temperature. It was also found that the only
chemistry element that had a substantial role to play in scale formation was the amount o f
Phosphorus present in the coil, which could be increasing the adhesion o f the scale to the coil
substrate.

5.6 References

B. D. Marx: “Iteratively Reweighted Partial Least Squares Estimation for Generalized Linear
Regression”. Technometrics, 1996, Vol. 38, No. 4, pp. 374-381.

P. McCullagh, P., and J. A. Nelder: Generalised Linear Models, 1989, 2nd Ed. Chapman &
Hall, London.

P. Bastien and M. Tenenhaus: PLS Generalized Linear Regression: Application to the

Analysis of Life Time Data”. Computational Statistics & Data Analysis, 2005, Vol. 48, Issue:
l,p p . 17-46.

74 f
Chapter 6. Process parameters influencing tertiary scale formation at a hot strip mill
using a multinomial logit model

6.1 Introduction

An alternative to using the scale count is to classify scale count into groups. Whilst this also
results in some loss o f information, it can often lead to more interpretable models and give
predictions in the form o f probabilities o f scale forming under different operating conditions -
which can be useful for process control due to large variability. A t one extreme is the binary
classification where an indicator variable replaces scale count with this variable equalling
unity when scale count exceeds the chosen value and zero otherwise. This is a simple scale -
no scale classification. This classification results in the largest amount o f lost information
relative to the count data itself but is appropriate if scale counts less than this value is
unlikely to be a problem for further processing (and all the costs associated with that).

Whilst the above binary approach can identify conditions most likely to lead to counts below
200 and therefore where no further processing is required, it cannot predict conditions leading
to other amounts o f scale forming. Where for example further processing is required but the
cost o f this is not too excessive. As the cost o f further processing tends to increase with an
increasing scale count, finer and more detailed process control often requires a broader
classification of scale count. For example, the scale count data could be categorised as low,
medium and high. Note that as the scale classification is broadened this type o f approach
tends to that obtained using a Poisson distribution for count data. This paper uses a broader
classification than the binary one which can be seen as a balanced approach between the
extremes of a binary classification and no classification - where the raw count data itself is
used.

6.2 MLR data classification

Table 13 summarises the categories o f scale count used for the dependent variable for the
MLR model. Essentially, a coil is classified as having either a very low scale count, a low
scale count, a moderate scale count or a high scale count. An indicator variable Yj is used to
summarise the category into which each o f the individual manufactured coils falls into. Y* = 0
when a coil has between 1 and 10 scale counts on its bottom surface, Yj = 1 when a coil has
between 11 and 75 scale counts on its bottom surface, Yj = 2 when a coil has between 76 and
350 counts o f scale on its bottom surface and Yj = 3 when a coil has over 350 scale counts on
its bottom surface (the subscript i denotes an individual coil). These boundaries were chosen
because they resulted in the most balanced distribution o f coils within each grouping - with
around 25 % o f coils making up each category.

Table 13. Classification of splits for the MLR model.

Number in % o f Dataset
Indicator Variable, Yj = k Boundaries for scale count
category in category

k = 0; Very Low Scale 0-10 367 23.30

k = 1; Low Scale 11-75 462 29.30
k = 2; Medium Scale 76 - 350 369 23.40
k = 3; High Scale 3514- 379 24.00

6.3 Principle Component Analysis

For an explanation of PCA, see Section 4.1.

Once all p principle components have been derived two additional questions require
answering. First, how many principal components are essential for representing all the
process variables? Cattell (1952) has suggested that only those principal components having
an Eigen value greater than 1 should be considered as essential and therefore retained in the
analysis. Secondly, do all the process variables need to go into making up each principle
component? Frequently, extraction communalities are used to answer this second question.
An extraction communality measures the variance in the process variables (once
standardised) that are accounted for by each principle component. As such, extraction
communalities are equal to the squares o f the loadings described above. A common rule o f
thumb is that the loadings should be 0.7 or higher to confirm that a process variable is well
represented by a particular principal component. This paper will therefore use a cut o ff o f 0.6
for the squared loadings and extracted communality values below this number are taken to
indicate variables that do not fit well with the spectral decomposition and should not be
included in the construction o f the principal components.
6.4 MLR model

The multinomial model can be written as

Prob(Yi = k) =
X L o *XP (pOk + £ y =1 Pjkx i j )

Eq. (6.2)

For the data set used in this paper, K = 3 (i.e. 4 scale categories), there are n = 1577 coils that
have undergone a scale count performed and there are p = 39 process variables. However,
there is an indeterminacy in the model as it is written in Eq. (6.2). This indeterminacy takes
the form o f the probabilities given by Eq. (6.2) being not unique because, with the
probabilities having to sum to unity over the K categories at a given set o f process variable
values, only K -l parameter vectors would be needed to determine the K probabilities
associated with each scale category. A convenient normalisation that solves this problem is to

set Pok = Pik = •••• = Ppk = 0 when k = 0, i.e. for the very low scale category. Then the
probabilities are

eXp{PQk + EjLi/JjfcXy)
ProbCYi = k ) = Pik fc = 0,1, 1

Eq. (6.3)

This multinomial model implies that K -1 log - odds can be computed as

Eq. (6.4)

Within this framework the binomial logit model used in chapter 4 results if K = 1.
6.5 Estimating a Multinominal Logit Model

From the point o f view o f estimation, it is useful that the odds ratio Pjm/Pik is not dependent
on the other scale categories. The log likelihood to be maximised when estimating the Pjm
parameters can be derived by defining for each coil, dik — 1 if coil (i) has a scale count
placing it in the kth scale category and zero otherwise for the K possible scale categories.
Then for each coil (i) one and only one of the djk’s is unity. Then the log likelihood is

InL=zz n K-

i=l k=0
1

dijln{Prob(Yi = k )}
Eq. (6.5)

The estimated values for the parameters are taken to be those that maximise Eq. (6.5) which
can be easily achieved using a standard nonlinear optimisation algorithm such as SOLVER in
Excel 2010.

6.6 MLR Results and Discussion

6.6.1 Principle component testing procedure

Using principal component analysis the correlations existing between all the process
variables was investigated. The analysis was performed to reduce both the number o f process
variables and to remove the correlations among the process variables used in the model to
predict scale count. With the intention o f gaining an understanding o f how to run the Hot Mill
so as to minimise scale formation. The temperature variables in table 5 were highly correlated
with each other and so it was decided to restrict the principal component analysis to these
variables. For example, the correlation coefficient between the average and maximum
Rougher Mill temperatures was 0.92.

It was decided to use different variables in the construction o f this PC, compared to the one
for Chapter 4. This was undertaken to gain a greater understanding o f the relationship
between the temperatures. The variables used in the construction o f the PC are listed in table
14.
Table 14. Variables used in the construction of the principle components.
Important Temperature Variables

X26 Furnace Aim Temperature Average Temperature, °C

X27 Furnace Centre/Surface Difference, °C

X28 Furnace Surface Temperature, °C

X29 Furnace Centre Temperature, °C

Xl3 Maximum RM Temperature, °C

Xl5 Average RM Temperature, °C

Xl4 Minimum RM Temperature, °C

Xl6 Maximum CS Temperature, °C

X 19 Maximum FM Temperature, °C

X21 Average FM Temperature, °C

An iterative procedure was adopted to identify the important principal components and the
temperature process variables present within them. First, all the temperature variables were
used and all principal components were extracted. If a variable had a squared loading below
0.6 in any o f the principal components having an Eigen value above 1, that process variable
was removed from the analysis. Second, with the reduced number o f temperature variables,
the principal component analysis was repeated and the same procedure as above used to
eliminate further temperature variables. This process was repeated until there were no longer
any squared factor loadings below 0.6 in value. Table 14 lists the ten important process
variables identified by this iterative procedure and table 15 shows the Eigen values associated
with these principal components. It can be seen from table 15 that the average and minimum
Crop Shear temperatures, the minimum Finishing Mill temperature, the average ROT
temperature and all the coiling temperatures appear to be unsuitable for inclusion in any o f
the principal components.
Table 15. Principal component results for the temperature process variables.
Components Eigen values Variation, % Cumulative Variation, %

PC, 3.977 39.771 39.771

PC 2 2.165 21.649 61.42

PC3 1.283 12.826 74.247

PC 4 1.058 10.583 84.829

PC 5 0.807 8.074 92.903

PC 6 0.289 2.889 95.792

PC 7 0.227 2.265 98.057

PC 8 0.098 0.979 99.036

PC 9 0.069 0.693 99.73

PC,o 0.027 0.27 100

Further, table 15 reveals that the first four principal components accounted for nearly 84 % o f
the variability present in all ten temperature process variables. Using Cattell’s rule (Cattell
(1952)) o f only needing to use those principal components having an Eigen value o f around
unity or more, the ten temperature process variables can therefore be replaced by the first four
principal components. These four principal components had the following form;

PC± = 0.0168 * ( * 2 6 - 4.78) + 0.0067 * ( x 27 - 44.96) + 0.0153 * (x28 - 1227.13)

+ 0.0065 * ( * 29 - 1182.48) + 0.0242 * ( * 13 - 1121.75) + 0.0259 * ( * 15

- 1099.43) + 0.0014 * ( * 1 4 - 918.34) + 0.0219 * ( * 16 - 1099.88)

+ 0.0489 * ( * 19 - 907.16) + 0.0636 * (x21 - 881.05)
Eq. ( 6 .6 a)

PC2 = -0 .0 3 6 1 * ( * 2 6 - 4.78) + 0.02 12 * ( x 27 - 44.96) + 0.00 29 * ( * 2 8 - 12 2 7.13)

- 0.0366 * ( * 29 - 1182.48) + 0.0059 * ( * 13 - 1121.75) + 0.0026 * ( * 15
- 1099.43) + 0.0021 * ( * 14 - 918.34) + 0.0059 * ( * 16 - 1099.88)
+ 0.0049 * ( * 19 - 907.16) - 0.0119 * (x21 - 881.05)
Eq. ( 6 .6 b)

80
PC3 = - 0 .0 2 17 * ( * 2 6 - 4.78) - 0.0 176 * (x27 - 44.96) - 0.03 31 * (x28 - 12 27.13)
- 0.0069 * ( * 29 - 1182.48) + 0.0025 * ( * 13 - 1121.75) + 0.0029 * ( * 15
- 1099.4337) - 0.0091 * ( * 14 - 918.34) + 0.0081 * ( * 16 - 1099.88)
+ 0.04807 * ( * 19 - 907.16) + 0.0602 * ( * 21 - 881.05)
Eq. (6 .6 c)

PC4 = 0.00 44 * ( * 2 6 - 4.78) + 0.00 58 * ( * 2 7 - 44.96) + 0.00 9 2 * ( * 28 - 12 2 7.13)

- 0.0008 * ( * 2 9 - 1182.48) - 0.0101 * ( * 13 - 1121.75) - 0.0184
* ( * 15 - 1099.44) - 0.0729 * ( * 14 - 918.34) - 7.63E - 006
* ( * 16 - 1099.88) + 0.0256 * ( * 19 - 907.16) + 0.0297 * ( * 21 - 881.05)
Eq. (6 .6 d)

6.6,2 Multinominal logistic model results

A simple general to specific procedure was next utilised to obtain the structure o f a
parsimonious Multinomial Logistic Model. First, all the variables in table 5 were included in
the model, but with all the temperature variables shown in table 14 replaced by the four
principal components shown in the previous sub section. Using all these process variables
and principal components the log likelihood given by Eq. (6.5) was maximised using
SOLVER in Excel. All process variables that were not significant at the 5 % significance
level were then removed from the model. This significance o f a process variable was judged
by calculating twice the ratio o f the log likelihoods calculated with and without this process
variable. Under the null that this process variable is unimportant, this statistic is Chi square
distributed with 3 degrees o f freedom (as each variable enters the equation for predicting the
log odds in the four different categories of scale). Using this reduced number o f process
variables, the log likelihood given by Eq. (6.5) was maximised again using Excel’s SOLVER.
This iterative procedure was repeated until all the process variables in the Multinomial Logit
Model were statistically significant at the 5 % significance level.
Table 16. The estimated values for the parameters in Eq. (6.4).
Scale category
Variable Low Scale; Yj= 1 Medium Scale; Yj = 2 3 High Scale; Yj = 3
Constant 8.98 [1.2]’ 43.02 [4.7] 30.89 [2.7]

X38 -119.76 [33.6] -191.43 [-5.0] -303.39 [-6.7]

X39 -0.0004 [-0.3]* -0.005 [-2.5] -0.009 [-4.7]
Xl -54.75 [-3.0] -42.79 [-2.0] -151.18 [-5.9]
X3 84.33 [2.8] 384.79 [10.1] 702.25 [15.4]

X6 27.52 [1.8] 50.09 [2.7] -8.80 [-0.4]*

X7 18.13 [0.6]* -72.52 [-2.0] -135.25 [-3.1]

X8 -48.69 [-2.4] -120.66 [-4.6] -162.81 [-4.9]

X9 568.61 [3.4] 532.48 [2.7] 375.19 [1.7]

X l2 -632.02 [-3.5] -680.09 [-3.2] -572.31 [-2.3]

X31 -0.229 [-1.3]* -0.888 [-4.2] -1.13 [-4.5]

X ,8 0.016 [3.1] 0.011 [1.8] 0.037 [4.6]

X22 -0.023 [-3.2] -0.054 [-7.3] -0.055 [-7.2]
PC, 0.116 [2.2] 0.625 [9.0] 1.03 [11.9]
PC2 0.276 [5.0] 0.401 [6.2] 0.303 [4.1]
PC 3 -0.001 [0.1]* 0.204 [2.5] 0.321 [3.1]
Ln(L) = -1598. Base Ln (L) = -2179. Student t values in parenthesis. Statistically
insignificant at the 10 % significance level.
Table 16 shows the estimates made for pjk parameters in the parsimonious version o f Eq.
(6.4) as identified using the above mentioned iterative procedure. Recall that for model
identification, normalisation with respect to the very low scale category was undertaken (i.e.
the parameters o f the model for this category are set to zero). In table 16, all variables are
statistically significant in the sense that they determine the probability o f scale formation in at
least one category o f scale severity. For example, the addition o f Nickel (X7) to the base steel
does not affect the probability o f low amounts o f scale forming, but it does significantly
determine the probability o f medium and high amounts o f scale forming. Exactly the same
can be stated for the operation o f the Interstand Spray System (X 3 1 ) in that the choice between
automatic and manual operation does determine the probability o f medium to large amounts
o f scale forming - but not the probability o f low amounts forming. On the other hand, the
addition o f Copper (x 6) is statistically significant in the determination o f the probability o f
low to medium amounts o f scale forming on the steel surface, but not large amounts.

Table 17. Table of observed and predicted values for scale count.
Predicted Category
Sum % o f correct
Yj = 0 Yj = 1 Yj = 2 Yj = 3
Observed predictions
0

182 136 19 30 367 50

II

T3 Yj = 1 122 242 71 27 462 52

<D b
'c
<D too
<L>
Yj = 2 28 103 139 99 369 38
Cfl
BS
O U Yj = 3 10 29 68 272 379 72
Sum Predicted 342 510 297 428 1577
Yj = 0 corresponds to a very low scale count, Yj = 1 corresponds to a low scale count, Yj
= 2 corresponds medium scale count, Yj = 3 corresponds to a high scale count.

The base log likelihood, is the log likelihood associated with a model containing only a
constant term. Thus the ratio o f this to the log likelihood when the process variables in table
16 are included provides a measure o f explanatory power. Using these criteria, the inclusion
o f these process variables improves the ability to predict various categories o f scale intensity
by almost 30 %. Table 17 provides another way o f visualising the explanatory power o f this
model as demonstrated by giving a cross plot o f the models predictions against the actual
severity of scale formation. Thus 379 of the 1577 coils had a high scale count and the model
correctly predicted 272 o f these coils to have a high scale count. (If the model gives a
probability o f more than 0.5 for a high scale count, then that coil is predicted by the model to
have high scale count formation). This is a correct prediction rate o f 72 %. The model has
mediocre power in predicting low scale counts, and poor predictive power in relation to
moderate amounts of scale formation. It thus appears that this type o f statistical model works
best at identifying the process conditions leading to large amounts o f scale forming compared
to low amounts o f scale forming. When using this model, more success in controlling scale
will be achieved by ensuring the Hot Mill is not run under those conditions most likely (as
predicted by the model) to lead to high scale counts, compared to ensuring the Hot Mill is run
under those conditions most likely to lead to low scale counts.

There are various ways to use this model to identify these optimal Hot Mill conditions. One is
to use an optimisation algorithm to search for those Hot Mill conditions the model identifies
as most likely to result in high scale counts (and then avoid running at these conditions).
There are likely to be many such conditions that could then be tabulated for the Hot Mill
operators to refer to. An alternative is to take a partial view, i.e. look at one process variable
at a time whilst running the Hot Mill at the average (or other predetermined set o f values)
values for all the other process variables. Using this approach there are two useful tools or
statistics to refer to - probability plots and elasticities.

Whilst the parameter values in table 16 are useful for identifying the statistically significant
variables that are present in scale formation, they demostrate very little about how important
each variable is for controlling the process - because these parameters are related to the log
odds ratio rather than the probabilities o f scale forming. Elasticities are superior in this
respect because they show the percentage change in the probability o f each category o f scale
forming that results from a one per cent change in the process variables. In table 18 below,
these elasticities are calculated at the average values for each process variable.

84 |
Table 18. Elasticities for each of the statistically significant process variables.
Process Variable Elasticity (% change)
Scale Categories
Yj = 0 Yj = 1 Yj = 2 Yj = 3
Transfer Gauge (x 3 g ) 1.06 0.36 -0.62 -0.81
Coil Width ( x 39) 0.45 0.80 -0.52 -0.75
% Carbon (x i) 0 .6 6 -0.06* 0 .2 2 * -0.83
% Phosphorous (X 3 ) -0.62 -0.82 0 .6 6 0.79
% Copper (X 6 ) -0.08 0 .0 1 * 0.13 -0.07
% Nickel (X 7 ) 0.07 0 .2 1 -0.14 -0.14
% Chromium (x g ) 0.25 0.19 -0.25 -0.19
% Aluminium (X 9 ) -2 . 6 6 1.75 1.13* -0 .2 2 '
% Solute Aluminium (X 1 2 ) 3.01 -1.26 -1.58* -0.18
Interstand Spray System (X 31) -0.06 -0.07 0.08 0.05
Average CS Temperature (x jg ) -2.70 0.73* -1.07* 3.04
Average ROT Temperature (X 2 2 ) 4.38 2.75 -5.11 -2 . 0 1
Yj = 0 corresponds to a very low scale count, Yj = 1 corresponds to a low scale
count, Yj = 2 corresponds medium scale count, Yj = 3 corresponds to a high
scale count. * Statistically insignificant at the 10 % significance level.

Ignoring the temperature variables in the three principal components for a moment, it can be
observed that irrespective o f the scale category, the largest elasticities are associated with the
Aluminium content o f the slab and various Hot Mill temperatures. Thus it can be seen that
the average Crop Shear (xig) and the average ROT (X22) temperatures have the highest impact
on the probability of observing high scale counts on the finished coil. For example, a 1 %
increase in the average Crop Shear temperature (xig) will bring forth a 3.04 % increase in the
probability o f high scale counts occurring on the finished coil when the Hot Mill runs at the
average values for all the other process variables. Then at the other end o f the scale
classification, it can be seen that the % Aluminium Solute (X12) and the average ROT
temperature (X 2 2 ) have the biggest impact on the probability o f observing very low scale
counts on the finished coil. For example, a 1 % increase in the average ROT temperature (X22 )
will bring forth a 4.38 % increase in the probability o f very low scale counts occurring on the
finished coil when the Hot Mill runs at the average values for all the other process variables.
When looking across the scale categories, the elasticities associated with each variable
change sign as you move from the lower scale count categories to the higher ones For
example, a 1 % increase in the % Phosphorous added to the base steel will bring forth a 0.62
% decrease in the probability o f very low scale counts being observed on the finished coil,
but a 0.79 % increase in the probability o f high scale counts being observed when the Hot
Mill runs at the average values for all the other process variables. Elasticities can also be
calculated using values other than the average for the other process variables.

In table 18 the elasticities for the principal components are not shown becausi by
construction they will equal zero. This is because the mean values for each component a*e by
construction zero. This does not imply that the process variables in each o f the compoients
are unimportant, but that under the special conditions where these variables are such thzt the
components are zero, their influence is minimal. Under all other conditions the temperature
variables in each component are significant variables as observed from the probability plots
in figure 13.
 - Yi = 0 ; Very low scale count
- Yi *= 1 ; Low scale count
■ Yi - 2 ; Medium scale count
Yi = 3; High scale count

M 0.7

£°6
e
* •0.5
*

Yi = 0 ; Very Low scale count

— Yi = 1; Low scale count
— - Yi * 2 , Medium scale count
Yi = 3, High scale count

0.004 0.006 0.008 0.01 0.012 0.014 0.016 O.OtS 0.02 0.022 0 OS 0.055 0.06 0.065 0.07 0.075 0.08

% Phosphorous % Carbon

§ 0.6
£
3 0.4

-Yi-6 Very tow scale Yi * 0 ; Very tow scale count

count
— — Yi = 1 Low scale count — — Yi * 1 , l«ow scale count
---------- Yi =2 Medium scale count - - - - - Yi =2 ; Medium scale count
---------Yi = 3 High scale count * • Yi = 3; High scale count

0.03 0.04 0 03 0 035

% Alumiium % Solute Aluminium

R< T n iq 'c t.r tm e V F iM Pi im ip .il <\>nqM>nci»t p< \

S e t on«l P i iik ij>al ( *«uiip<>ncnt P< l l l l l t l P lU K J p .il < * * 'm p i‘llC llt P< *<

Figure 13. Probability of scale formation with variations in some of the process
variables.

87
Figure 13 shows the effect of changing each process variable on the probability o f observing
the various categories o f scale whilst holding all the other variables at their mean values. The
three principal components o f the temperature variables are shown, together with some o f the
other process variables that have high elasticities.

To illustrate how to use such graphs for control purposes take the amount o f aluminium
added to the slab. At low levels o f Aluminium (0.024 % or less) it is virtually certain that the
processed coil will have a very low scale count when operating the Hot Mill at the average
values for all other variables shown in table 5. Contrastingly, when the incoming slab to the
Hot Mill has 0.04 % Aluminium and the Hot Mill operates at average values for the other
variables there is limited chance o f the processed slab having a very low scale count. Under
these conditions around 55 % o f processed coils will have low scale counts and a further 50
% medium scale counts. The other chemistries shown in figure 13 can be interpreted in
similar ways. Interestingly the effect o f % Solute Aluminium (X12) is similar to % Aluminium
(X 9 ) but in the opposing direction.

Next consider the first principal component in figure 13. When this first principal component
is below -3 in value (with all other process variables set at their average values), around 90 %
o f processed coils will have very low or low scale counts. Consequently, a small increase in
this principle components value above -3 will result in a dramatic change, for example over
75 % of coils being medium or high in scale count when this principal component value is 2.
Such probability plots can also be calculated using values other than the average for the other
process variables.

These principal component values can also be converted into actual temperature values at
various points in the Hot Mill. As an example take the 90 % or more o f coils being low or
very low in scale as being an acceptable outcome for the Hot Mill output. This corresponds to
PCj = -3 or less as shown above. Using Eq. ( 6 .6 ) and PC] = -3, a particular Hot Mill
temperature can be calculated assuming the process runs at the mean values for all the other
temperature variables at the Hot Mill. These calculated temperatures are shown in table 19. It
was not possible to achieve a PCi value o f -3 or less by manipulating process variables X26 ,
X27 and X]4 , when running the mill at average values for all the other variables shown in table
5. The centre o f the Reheat Furnace should be set at 875 °C. When observing at the average
Rougher Mill temperature the temperature should be 1020 °C or less. The corresponding
temperature for the maximum Crop Shear temperatures is 1010 °C or less. The corresponding
temperature for the maximum Finishing Mill temperatures is 865 °C or less. Finally, for the
average Finishing Mill temperature the temperature is 850 °C or less. This type o f calculation
can be done for any PCi value, and therefore for any frequency o f significant scale formation
to determine what temperatures to run the Hot Mill at to achieve this required frequency. It
should be noted that the values in table 19 are unlikely to be used because the material would
struggle to be effectively processed at these temperatures or achieve the required mechanical
properties.

Table 19. Maximum temperature settings to achieve a 10 % or less chance of very low
scale forming.

X28 X29 X13 X15 X]6 X19 X21

Temperature, °C 1095 875 1040 1020 1010 865 850

6.7 MLR Conclusions

The principal component analysis above indicated that the recorded temperatures at the Hot
Mill were highly correlated. However, nearly 85 % o f the variation in the ten Hot Mill
temperatures could be accounted for by just four principal components that were completely
uncorrelated with each other.

The above analysis showed that to be able to successfully predict scale formation many
variables need to be taken into account including chemistry (most notably the addition o f
Phosphorous and Aluminium), temperature (that at the Run Out Table and those temperature
constituting each principal component), and the descaler sprays. The Multinomial Logistical
Regression model also had an overall correct prediction rate o f 50 %. However, a correct
prediction rate o f 72 % was obtained for the high scale category. This large difference in
prediction shows that there is a large inherent variability within different sections o f the
dataset, which implies that some scale count ranges have a significant influence from
variables not being utilised within this dataset.
The main aim for estimating this multinomial model was not just to find the most important
determinants o f scale formation as the above identified variables are broadly consistent with
what is already known. The main aim was to provide quantitative information on how to
control the Hot Mill, i.e. where to set the process variables so as to minimise the number of
coils produced with significant scale present on their surfaces. The model suggested that this
can be done by choosing either the right chemistry for the incoming steel or through setting
the correct temperatures at the Hot Mill. It was found, for example, that to keep the rate o f
coils containing medium or high scale counts below 10 %, the average Rougher Mill
temperature should be kept at 1020 °C or more, when running the Hot Mill at the average
values (shown in table 5) for all the other process variables. For chemistry it was found for
example, that to keep the rate o f coils containing medium or high scale counts below 10 %,
the Aluminium content must be below 0.024 % wt. when running the Hot Mill at the mem
values for all the other process variables.

It should be noted that the model could also be used to identify the optimum temperatures to
minimise scale formation if the chemistry was required to be changed.

6 .8 References

Cattell, R. B., 1952. Factor Analysis. New York, Harper.

Chapter 7. Development of a multi-layer ANFIS model for the prediction of tertiary
scale formation

The Adaptive Neuro fuzzy inference system (ANFIS) model integrates neural network and
fuzzy logic principles and uses them within a single framework. The model developed within
this body o f work is constructed in 3 layers. The aim is then to use the knowledge gained
from this technique to better understand scale formation and how it can be reduced at the
point o f manufacture. Only those variables identified as being significant (see table 20) in the
previous 3 chapters are considered for inclusion in this ANFIS model.

Table 20. Process Variables and their Sample Means and Standard Deviations.

Process Variable xj Mean Standard Deviation

% Phosphorus X3 0.014 0.0031
Maximum RM Exit Temperature, °C X l3 1 1 2 1 .8 17.87
Average RM Exit Temperature, °C X ]5 1099.4 16.18
Maximum CS Temperature, °C X i6 1099.9 19.89
Average CS Temperature, °C X ]8 1069.6 2 0 .8

Maximum FM Exit Temperature, °C X i9 907.16 7.18

Average FM Exit Temperature, °C X21 881.05 5.75
Scale Count Bottom Surface Y 250.1 402.6
RM = Rougher Mill, CS = Crop Shear, FM = Finishing Mill and ROT = Run Out
Table.

7.1 ANFIS model

The basic philosophy o f the ANFIS model is to split the data up into subsets and then to fit a
standard second order response surface model to the data in each subset. However, where the
splits in the data occur are fuzzyfied. Figure 16 describes the structure for a simplified ANFIS
model where the experimental data is split up into four sub sections.
 Yes
No

Yes No
Yes No

f (x,, x2) = S0 + Sj. Xj + £ / , (x, ,x 2) = S, + 6 , £ Xt + £ / , (x,, x2) = S6 + SJ]T .1

/= ! . ;= 4

Figure 14. A t> pical binar> regression tree with two inputs (xi and X2 ) and one ou tp u t y.

The decision tree partitions the input space into a number o f non overlapping rectangular
regions, each o f which is assigned a label fj() to represent a predicted output value. In this
illustration the tree identifies four simple if - then rules. Note that each terminal node has a
unique path that starts with the root node and ends with the terminal node; the path
corresponds to a decision rule that is a conjunction o f various tests or conditions. Typically,
each fj() will be a linear function of X] and X2 and for each region there is a separate linear
equation. The functional form for fj() does not have to be linear however. More generally,
and as used in this paper, it can be a second order response surface model given by

2 2 2 2

f i x 1 , x 2) = ft) + Pj Xj + Xjv ^ XjXv + (pj £ x f

j=1 y=l i;=y' +l 7 =1

Eq. (7.1)

92
The problem with this approach is that the resulting modelled response surface is highly
discontinuous in that it changes abruptly at the decision rules. This problem is overcome by
fuzzyfying the decision rules. For example, the crisp decision rule associated with the left
most branch o f the binary tree in figure 14 is

If xi < a and X2 < b then y = fi()

It is established that fuzzy expert systems use membership functions to quantify possibilities
(Jang et a l (1997)). Possibility is a fuzzy measure indicating the degree o f evidence or belief
that a certain value xi belongs to a set. A membership function has a value between 0 and 1
such that xi values further and further below a, have membership values closer and closer to
one. A common functional form used for the membership function is the sigmoidal function

1
HxKa - 1 - X + e xp[ - K( Xl - a)]
Eq. (7.2a)

where k is a parameter requiring estimation. The value for k determines the steepness o f the
membership function at a. So the further xi is below a, the greater will be the value for pxi < a,

indicating a stronger belief that that value for xi belongs to the set xi < a. pxi < a varies over
the range 0 to 1 , with 1 indicating the strongest possible belief.

Finally, the parameters o f the model are fine tuned using a particular type o f neural network.
This network is called an ANFIS - adaptive network - based fuzzy interference system
(Tsoukalas and Uhrig (1997)). There are various ANFIS architectures, but one using a first
order Sugeno (1995) fuzzy model is the most common. The layers o f this ANFIS network are
shown in figure 15 for the decision rules shown in figure 14.

Rule 1: If xj < a and X2 < b, then y = fi()

Rule 2: If xi < a and X2> b, then y = f2 ()
Rule 3: If xi > a and X2 <c, then y = fj()
Rule 4: If xi > a and X2> c, then y = f*()
In the first layer o f the ANFIS network, each o f the i values for xi and X2 are given
membership quantities using the following sigmoidal functions

= ! - : --------- r — 7 ------ 77 with inverse (INV) n xXia =

1 + expj-K-, (jCj - a)\ 1 + exp [-a:, (x. - a ) ]

Eq. (7.2b)

1
M,2<b = 1 “ 7-------- 7 —,------— w ithinverse (INV) n n>h =
l + exp[-x*2 (jc, - 0 )] l + exp[-A*2 (jc, ■
-* )]
Eq. (7.2c)

1
Mr?<r = 1--------------- with inverse (INV) u ^ >r -
1 + exp[-x -3 (x, - c)\ 1+ exp[-/r 3 (jCj - c)]

Eq. (7.2d)

where Ki to K3 are parameters requiring estimation. In layer 2 weights are determined that
represent the possibility that each pairing for the i values o f X] and X2 belong to one o f the
four sets given by the decision rules above. These weights are given by

wi = (pxi<a)(px 2 <b)
Eq. (7.3a)

w 2 = (pxi<a)(px2>b)
Eq. (7.3b)

w 3 = (pxi>a)(px 2 <c)
Eq. (7.3c)

w 4 = (jj,xi>a)(px2>c)
Eq. (7.3d)

These are examples o f a T-norm operator for working out the possibility, for example, that xi
is less than a AND X2 is less than b. In figure 15, n stands for the use o f this T-Norm. In layer
3 each value for fj() is multiplied by its Wj value so that more emphasis is placed on the linear
function corresponding to the rule most likely to describe the x\ and X2 pairing (the form for
fj(0 is as shown in figure 14). Finally, in layer 4 these weighted functions are added up to
give the prediction for y coming out o f the ANFIS network. Conjugate gradient methods are
then used to optimise the values for Ki to K3 and all the parameters o f fi() to f4 ().

Layer 1: Memberships Layer 2 Layer 3: Rule

Figure 15. ANFIS architecture corresponding to the representation shown in

Figure 14. Values for w given by Eq. (7.3).

The only remaining question is how to decide which variables to split by and how many
variables to split. In this paper a simple binary procedure is used. The first o f the p process
variables shown in table 20 is selected and the above ANFIS model fitted for just this one
variable. The predictions from this ANFIS model with p = 1 are then compared to the
observed values and the sum squared error is computed

Sum Square E rror = 'Zi=i(Predictedi — Yt) A2 Eq. (7.4)

for all different values o f Ki. The value for K] is then taken to be that value producing the
smallest sum squared error. This is repeated for all p variables and the variable chosen for the
first split is the one producing the smallest Sum Squared Error. The variable chosen for the
second split is found in the same way and any number o f splits can be made provided each
sub section o f the data has enough observations in it to estimate Eq. (7.1).

7.2 ANFIS Results and Discussion

Table 21 shows the sum square error of each o f the stages described above. Var is a vector
containing the variables used to split the data. Thus Vari contains the variable used to create
the first split and Var2 , contains the variable used to create the second split in the data and so
on. Table 21 shows that the best predictor for Vari was the AvgRM. For, Var2 , PerP is the
best predictor o f the variables investigated. While for Vara AvgCS was added to the ANFIS
model.

Table 21. Table showing the sum of the square errors for the different variables used.

Sum Square Error Sum Square Error Sum Square Error for
for Vari for Vari ,2 (AvgRM, Vari 52,3 (AvgRM, PerP,
(AvgRM) PerP) AvgCS)
PerP 1363.95 935.99 N/A
MaxRM 1264.09 1174.43 825.16
AvgRM 1207.79 N/A N/A
MaxCS 1373.91 1150.69 811.82
AvgCS 1375.78 1072.16 758.16
MaxFM 1466.22 1103.87 809.35
AvgFM 1409.06 1139.55 826.95
The numbers in bold are the variables that have the lowest sum o f squared error
for that Var. The numbers in Italics have the highest sum o f square error for the
Var.

Consider first the ANFIS model with just the first variable identified in table 21. Figure 16 is
a graph o f observed versus predicted scale count for a Vari (AvgRM) ANFIS model. The
values for a and k in Eq. (7.2a) were 1.69583 and 1132.70 respectively and f(Vari) was
estimated to be

f(Vari) = (3,97)+(-3.08)*AvgRM+( 1.31 )*AvgRM 2 when Var, > a

f(V ar,)= (-0.23)+(0.17)*AvgRM+(0.01)*AvgRM2 when Var, < a

When a linear trend line is added to the data in figure 16, the equation for the trend line has a
slope value o f near 1 and a very low intercept term. The R value is 0.2336 which means that
23 % o f the observed values for scale count can be explained by this ANFIS model. The line
is close to a 45 degree line so there is very little systematic prediction error.

4000

3500

3000

2500
y = 1.004x- 1.0985
R 2 = 0.2336
2000

1500

1000

500 oo

500 1000 1500 2000 2500 3000 3500 4000

Predicted

Figure 16. Observed tertiary scale versus predicted tertiary scale for Vari (AvgRM).

Figure 17 shows the observed versus predicted results for the Vari ;2 ANFIS model. It can be
seen that in comparison to Vari that the slope o f the trend line has improved (converged to
one) by 0.0024 and the intercept has improved (converged to zero) by 0.2899. This combined
'7 • • •
with an improved R value o f 0.4104, indicates that the prediction obtained by this model has
less systematic errors in prediction and a higher predictive capability. This ANFIS models
prediction now describes 41 % o f the observed dataset’s results.
 5000

4500

4000

3500

T5 3000 y = 1.0016x- 0.8086

R 2 = 0.4104
S3 2500
-Q
o 2000

1500

1000
500

1000 2000 3000 4000 5000

Predicted

Figure 17. Observed tertiary scale versus predicted tertiary scale for V an ,2 (AvgRM,
PerP).

Figure 18 shows the observed versus predicted results for the Vari,2,3 ANFIS model (i.e. with
all 3 variables identified in table 21 included). When compared to the Varj^ model its trend
line equation has its slope value decreased by 0 .0 0 2 1 and the intercept has decreased by
0.9902. However, this is not a significant change. By using the 3 best calculated predictors
for the ANFIS model, the predictions explain 52 % o f the observed dataset.
 4500

4000

3500

3000 y = 1.0037x- 1.7988

R 2 = 0.5208
2500

£ 2000
1500

1000 o 0

500

1000 2000 3000 4000 5000

Predicted

Figure 18. Observed tertiary scale versus predicted tertiary scale for Vari,2,3 (AvgRM,
PerP, AvgCS).

Table 22 illustrates the regression parameters for this last ANFIS model, with the most
significant coefficients for each of the interactions highlighted in bold. Observations from
table 22 indicate the value with the highest regression parameter is X 2 (PerP) with 3 out o f 8

values highlighted. This is followed by Xi (AvgRM) and the constant with 2 out 8 each,
followed by X 3 (AvgCS) with 1 out o f 8 . This indicates that the variable that has the greatest
influence on scale count within a sub region o f the data is X 2 (PerP), and the general trend is
for variables to have greater significance than the interactions between the variables or the
squared values o f the variables.
Table 22. Table showing the regression parameters for Vari,2 3 . Bold numbers represent
the most significant parameters for that function.
Regression parameters

fl f2 f3 U f5 f6 f8
const. -216.27 67.83 -89.15 67.83 13.43 -4.47 14.74 -0.36
X, 168.70 -45.54 55.90 -45.54 10.24 -2.15 8.12 -0.19
x2 334.64 -18.57 97.07 -18.57 -28.31 1.58 -1.38 0.18

X3 -229.92 -14.60 85.77 -14.60 9.68 0.45 -21.42 0.30

X2 1 -23.04 8.61 -13.55 8.60 0.38 -0.73 0.51 0.02

X2 2 -64.08 4.72 19.47 4.72 8.11 -0.30 0.10 0.02

X 23 69.61 -2.43 -22.29 -2.43 -5.56 0.29 7.89 0.01

X 12 -44.87 4.26 -17.87 4.26 -3.56 -0.63 -4.21 -0.15

X,3 -6.65 5.70 -2.39 5.70 2.61 0.51 -5.95 -0.03

X 23 -25.81 -2.31 2.12 -2.31 3.73 1.15 0.25 0.24

Table 23 shows the calculated k and a values for Vari ,2 ,3 . The k represents the steepness o f
the membership function, a low value will give a shallow slope while a higher value will give
a steeper slope. A steeper slope will tend to yield an ANFIS model very similar to a CART
model. The data shows that the k value is lowest for AvgRM and highest for AvgCS. This
demonstrates that the steepness o f the membership function improves with every new
variable added to the model.

The a value (from table 23) represents where the fuzzy split in the membership function will
occur relative to the variable. The split for AvgRM is over two standard deviations away
from the mean value, so is at the higher end o f the range. The a value for PerP is within half a
standard deviation o f the mean, so is comparable to the average value o f the dataset. For
AvgCS the a value is greater than one standard deviation from the mean. These are the three
most important determinants of scale count at the Hot Mill.

Table 23. Table showing the calculated k and a values at Vari ,2 3 .

AvgRM PerP AvgCS

K 1.69583 3.2322981 4.3772184

a 1132.70 0.0155 1094.03
Table 24 illustrates coils at different scale counts with the predicted scale counts and the three
variables (AvgRM, PerP, and AvgCS). The variables show an overall increase with
increasing predicted values as one would expect. The difference in the observed and
predicted count in the sampled coils varies from being off by 20 to 200 counts. In general the
scale count prediction decreases in accuracy the higher the scale count becomes (excluding
the highest values). This is likely to be due to the influence o f conditioning sprays in the
Finishing Mill.

Table 24. Table showing selected coils at different scale counts

Observed Bottom Predicted Bottom Scale

Scale Count Count AvgRM PerP AvgCS
0 87.3 1094 0.014 1066
199 175.8 1109 0 .0 1 2 1077
1003 712.4 1084 0 .0 2 1072
3806 3924.2 1156 0.017 1093

7.3 ANFIS Conclusions

From the variables that were selected it can be concluded that scale formation is mostly
governed by the temperature o f the steel as it enters the early stands o f the Finishing Mill
(AvgRM & AvgCS). With 23 % o f the dataset being predicted by AvgRM and 11 %
(difference in R between V ari 52 and Var 1,2 ,3 ) o f the dataset being explained by AvgCS. This
gives a combined value o f 34 % o f the dataset explained by pre-Finishing Mill temperatures.

There is also an element o f chemistry that impacts upon the scale count. With PerP being the
best predictor for Vari s2 accounting for 18 % o f the observed scale counts.

The ANFIS model gives good predicted scale counts at low and high scale. This is most
likely a product o f the process, because if the temperature going through the Finishing M ill is
low, then a low amount o f scale will be generated. While if the coil is very hot coming into
the Finishing Mill then a high volume o f scale will be generated. However, for mid scale
counts (approximately > 750) the effectiveness o f the conditioning sprays might have a

101
significant influence. These sprays are situated on the entry into the Finishing Mill and will
suppress scale formation. If these conditioning sprays are running less effectively than
normal for a period o f time, then this could explain why the predicted values usually
underestimate the scale count. Also, if the sprays do not come on/off at the correct time then
the head/tail o f the coil will not be conditioned which will lead to increased scale on the
missed section. The other reason for this weakness in the model could be having average
temperature values across the coil.

Overall the 3 level ANFIS models prediction rate was 52 % o f a high variability dataset (i.e.
variation in scale count).

7.4 References

L. H. TSOUKALAS and R. E. UHRIG. “Fuzzy and Neural Approaches in Engineering”

(John Wiley & Sons Inc, New York, Chapter 2, 1997).

J. S. R. JANG, C. T. SUN and E. MIZUTANI. “Neuro-Fuzzy and Soft Computing: A

Computational Approach to Learning and Machine Intelligence” (Prentice Hall, Upper
Saddle River, Chapter 12, 1997).

T. TAKAGI and M. SUGENO. “Fuzzy Identification o f Systems and its Applications to

Modelling and Control”, IEEE Transactions on Systems, Man and Cybernetics. 15 (1995)
pp. 116-132.

muiW ni'ii ~ iiHiwaiw i rn r> rr> i — n i i i i i h i in an n m »p iin iiii w h i i nwi in i >mi m»i n m i n r i n nuni M T ir»Tnnanrw<w w im wii Mi y ir Miom wigtfn m irm iiifm n m iMru w rnmnnnr »>M» n 'm ynm im«ri«nirnnTi»-iM iiniiam»ii

102 |
Chapter 8. General Discussion

This general discussion section will bring together the main conclusion obtained from the
application o f the statistical techniques employed within this research.

8.1 Binary vs. Multinomial vs. Continuous

The dependent variable has been described in a number o f ways for the different data mining
techniques. The three ways that have been described are as follows;
• Binary
o Changes the continuous scale count to a binary classifier
• Multinomial
o Changes the continuous scale count to a multinomial classifier
• Continuous
o The dependent variable is the Parsytec scale count.

Each way o f classifying the data has its own strengths and weaknesses. Each o f these
strengths and weaknesses need to be considered before the models can be compared to each
other.

8.1.1 Binary

Models that use a binary split tend to have higher predictive capability. This is because it is
easier to predict a binary split compared to a continuous dependent variable. The relevance o f
these results are dependent on the choice o f the split.

It is also important to consider the placement o f the split as it needs to be relevant to the
process that is being investigated. The problem with using the scale count is that there is no
strict scale count that indicates scale problems downstream, but values above the selected
value for the split are more likely to have downstream issues.

The main strength o f this method is in classifying coils that have very low or very high scale
counts. If the split is in the correct location then the binary split should give good results for
low and high values o f the scale count.
The main weakness o f this method is the coils with scale counts around the split are going to
be harder to correctly predict. With the split being at 200, scale counts o f 199 will be classed
as 0 while values o f 201 will be classed as 1. In reality there is usually less difference
between 199 and 201, than 199 and 0 or 201 and 401+. This can cause errors and incorrect
predictions, which could lead to the wrong variables being selected.

Splitting the dataset into low quantity and high quantity scale has several advantages and
disadvantages. The main purpose o f this investigation is to identify the process parameters
related to scale formation and it could be argued that splitting the dataset in a binary way will
achieve this better than other methods. On the other hand the split gives low prediction for
values around the binary split.

8.1.2 Multinomial

Splitting the continuous dependent variable into sets has several advantages and
disadvantages over other techniques. The main advantage is that there are multiple sets that
the coil can be classed into so you can better understand the severity o f the scale present.

Like the binary dependent variable, problems with prediction occur around the splits.
However there are more splits present and they cover less range than in the binary split so
they should have less overall error associated with them.

The problem comes with the scale count not representing an increase in adherence but an
increase in quantity o f scale. This means that the different groups are not necessary
representative o f increasing scale problem. However scale count is an indicator o f potential
problems, so the different groups do represent increasing potential for scale problems to
occur.

8.1.3 Continuous

Keeping the dependent variable as a continuous variable has several strengths and
weaknesses when compared to other ways o f representing the dependent variable.
The main strength is that it directly represents the observed dependent variable, unlike other
methods which simplify the data. This avoids the problem associated with splitting the
dataset into various groups.

The main weakness from using a continuous dependent variable is that the scale count being
predicted is not a measure o f adherence o f the scale but frequency. As there is limited data at
high scale counts this could lead to the model incorrectly predicted significant variables
based on anomalous results. Also the variability is greater in the raw scale count data.

8.2 Model Comparisons

All the different models used in this investigation show slightly different significant
variables. This is due to differences in how the models work and how the dependent variable
is expressed. However there are some variables that keep occurring through the investigation.

In the different analysis undertaken temperature proved to be significant. As there is a high

correlation between the different temperature variables, as seen by the two PCA performed in
Chapter 4 and Chapter 6 . It can also be observed from the PC results that there is a greater
correlation between the pre-Finishing Mill temperatures than the post-Finishing Mill
temperatures.

PCI from Chapter 6 (see figure 12f) is comparable to the PCI from Chapter 4 (see figure 6 f).
When Chapter 4 ’s PCI has all its components at their mean values (zero on the horizontal
axis) there is a 15 % chance o f the bottom scale count being greater than 200. While Chapter
6 ’s PCI (at its zero value) shows a 13 % chance o f the bottom scale count being greater than
350, a 35 % chance o f it being between 76 - 350, a 32 % chance o f it being between 11 - 75,
and a 20 % chance o f the scale count being between 0 - 1 0 . These values are very similar to
each other and both show that as the temperature is increased above the mean values there is
a significant increase in the chance o f bottom scale being produced.

The temperature entering the Finishing Mill is the most significant variable for scale
formation. This makes logical sense as tertiary scale is formed in the early stands o f the
Finishing Mill (see 1.2.4).
The main chemistry component that seems to contribute to scale formation is Phosphorus.
Other elements have limited significance in some o f the models but are not in all o f them
because the tinplate grade investigated has a tight chemistry specification. Phosphorus was
found to be significant throughout the different analysis used, with values greater than 0.015
causing greater amounts o f scale formation. Chapter 4 shows that at 0.015 % there is a 20 %
chance that the coil will have a scale count greater than 200 on the bottom surface, which will
increase to a 45 % chance at 0.020 % Phosphorus (see figure 6 a). A similar pattern is
observed in Chapter 6 in figure 12a; at 0.015 % Phosphorus there is a 19 % chance o f the
bottom scale count being greater than 350, a 31 % chance o f it being between 76 - 350, a 35
% chance of it being between 11 - 75, and a 15 % chance o f the scale count being between 0 -
10. While at 0.020 % Phosphorus there is a 66 % chance o f the bottom scale count being
greater than 350, a 28 % chance o f it being between 76 - 350, a 4 % chance o f it being
between 11 - 75, and a 2 % chance o f the scale count being between 0 - 1 0 . Chapter 7 shows
that 18 % of the dataset is predicted by Phosphorus, and is the second strongest variable used
in the ANFIS analysis. These models both show that at Phosphorus levels greater than 0.015
% there is an increased chance o f scale being a problem.

The Logit and PLS models have several similarities for comparison. The first is that they both
use a binary dependent variable, which makes it easier for comparisons to be drawn. The PLS
model has a greater prediction rate than the Logit model (85 % compared to 73 %), which
means that the PLS model has a 12 % improved prediction rate on the dataset than the Logit
model. Which could be interpreted that the PLS gives more reliable results than the Logit
model. Both models show that temperature is a significant variable in scale formation. The
PLS model predicts a 10 % scale rate if the average Finishing Mill is below 842 °C. The
Finishing Mill temperature is dependent on the pre-Finishing Mill temperatures and the
conditioning sprays present. This implies that if the conditioning sprays are working and the
pre-Finishing Mill temperature is limited then better scale results will be achieved.

Even though the MLR and ANFIS models are very different there are several similarities that
need to be discussed. Both models show that some areas o f the dataset are more difficult to
predict than others. The MLR model had an overall prediction rate o f 50 % but the prediction
rate for individual sets varies from 38 % to 72 %. The ANFIS model gave improved
prediction at the extremes o f the dataset, but had decreased performance at predicting mid­
range values. This shows that there are scale counts that are difficult to predict with the
process data that is available. Both o f these models show that the percentage o f Phosphorus
and the temperature o f the strip is significant for scale formation.

8.3 M odel Use

To be able to perform successful analysis a number o f stages need to be undertaken. The first
stage is to identify the reason the analysis is being carried out, this will influence how the
problem will be structured and aid in interpreting the results. The second stage is to
understand what data is available and if the data will allow the question to be answered. This
should be carried out until the data collected is capable o f answering the required question.

Stage 3 is the preparation o f data. This is done by removing anomalous values and presenting
the data in a way that is suitable for whichever models have been selected. Stage 4 is
commencing the initial modelling, which should allow for a greater understanding o f the
data. This increased understanding will allow for improved modelling to be performed. This
process will be repeated until usable results are achieved.

Stage 5 is to evaluate the results to understand if it is has met the requirements o f stage 1. If
the results have not answered the original problem then, the analysis should return to stage 1 .
If the analysis has answered the original question then the results should be deployed to
obtain value from them.

8.4 Scale F orm ation

The temperature is the most significant variable for scale formation, particuly the temperature
entering the Finishing Mill. Due to data limitations it cannot be determined the exact location
o f the scale formation but the data analysis and literature review confirm that the entry into
the Finishing Mill is the critical location. This is highlighted in Chapter 6 and it can be
observed that there is marked increase in scale formation when the average Roughing Mill
temperature is greater than 1070 °C (see figure 10). With a similar increase in scale
occurrence occuring when the average Crop Shear temperature increases above 1050 °C (see
figure 1 0 ).
At lower temperatures little scale is generally formed. This confirms that scale formation is a
temperature dependent reaction. While at high temperatures scale is generated in large
amounts. Mid-way temperatures are influenced by other variables more than temperature,
such as chemistry and descaling practice.

The chemistry present in the slab also has an influence on scale formation. The chemical
element that is most significant for scale formation is Phosphorus. The literature states that
Phosphorus becomes enriched at the metal/scale interface which suppresses the diffusion o f
Fe ions. This lack o f Fe ions will not prevent the oxidation from arising, but will promote the
formation o f Magnetite and Hematite over Wustite. This will cause scale to be slightly more
adherent which can lead to higher counts present on the coil.
Chapter 9. General Conclusions

It can be concluded that temperature is the most significant variable for scale formation. With
the temperature entering the Finishing Mill being the most significant region to control to be
able to effectively manage scale formation. Low temperatures will result in low levels o f
scale formation while high temperatures will result in high levels o f scale formation. Mid
level temperatures are more difficult to predict because it seems that other factors are
influencing scale formation more significantly than temperature. In particular Phosphorus
seems to have a significant impact on scale formation especially at values greater than 0.015
%.

Models should not be used in isolation, as every analysis method has advantages and
disadvantages and they are often difficult to identify without several other methods to
compare against.

After discussion with process specialists it was decided to implement monitors to pass the
significant variables in a more visible way downstream o f the Hot Mill. The plan is to
continue to use the monitors to communicate between units in a productive way and to
continue to evolve the monitors over time to manage changes in the system. The initial
monitors are featured on table 25. The values and boundaries have been decided by
comparing the results from this thesis with ongoing investigations into scale formation.

Each o f the significant variables selected will be classed as RED (major risk o f scale),
AMBER (risk o f scale), GREEN (low risk o f scale). The class is selected depending on the
value o f the variable and the boundary conditions for the process condition. This simple
method o f visualising the data was selected to improve communication and improve the flow
o f information. There will also be an overall colour designation for the coil. The colour
selected will be the highest colour (ranked; RED, AMBER then GREEN) that the coil had.
 o o O
nt
oo

s
r-
o 0 OO
C /3
II C/3
l-H
A Sh
§ x Z z £
cd
U
w
X

f § 1 8
o
a>
o
X
o
Gh

N® o S® O x® a
X a3
"3
d «H
ox
d ox
<D NT)
a
ox o no
d u
x <L> m 0) NO <d
73
X
ON X ON B ON X
H->
pH d X X
A A A *5
o S-H >■^ <«
X
X u
O X u
o
<4-4 u
NO
PH A U o < X

o o u
o
d
00 t-~ <u
o o o r-
no
<u
Sh
r-H nt Ph
a> 00
oo -*-< Sh
o
^ X A m • O

hO PQ B00 w
'B W X P-H is no
0Q
Sb C*H Sh 73 <0
Sh
d d ^ o <D
Vh
o> <l> d w +H -*->
C /3 O
X® < O
X 00 £
C l''
o O '" o
^ O O o 0s t"-
no N® nd
<N (N -6 70 X <d Ox CL)
o 73 O X
A A A cd
X o d
<4-4 O X u X X CQ
Ph

cd
Q X
A A Q TO
A <N GO
So X C /3
d a O So K
<u a
a> c
0 <D
C /3
x
v
X Sh
H-»
1 C/3
nO
<0
C /3
t- H
o
n \ N CVN C/3 TO X
73 C /3
u
o
u
o 0s CD
Sh d oN
O
H-> C/3 X
U o
o o
oo
(N O h <d
o
O
'B
X Sh
o
A ON A A no
£ o 73
X o x o m o A a>
X
oo £ X

O
o o
m O
O X
OO
O

fi
U
o Oh o O h
O S
0) O B U
co CN 4-* CM CD
H-J o
-H -H B O
(N
a U
o < o < -fl

00 00 73
d O
Gft JO §
0) 13 13 "O
a)
C/3
73
2 (U
■o Oh 73
O h
C/3 *
O h
73
O h
'573 CD 'C a> S h
-*-> 0
8 73
u
d GO jd C/3 O h a> cd
00 73 oo O
Cm 00
O IS
O X X Sh
no o
4> > O Id
o CD CS <0 • »»M o
S
2 X «+H B
h

X O X X c3
H So
h
So o n0
X
c W a
B
CD
&
a> o o
<3
o a)
73
a3
x
o
O h
<L>
H
cd
C /3 O z d O

c
73 ♦
o <L>
Sh 73

x
a)
&
o3
I
<d a>
> O h X
-o ICD CD
• rH
73
O O
Sh
03 4~> X
o
C /3
<D
C /3
a 00
X £3 CD & X
n0
a>
C /3
0
<d B
in X pH X
£
fS <
d
110

C/3 > 00
CD
p«l O h •
CD -t-» fi X1
X! O
Sh X O c O 70
A u
o
cd X X X X X
H
Chapter 10. Recommendations

There are several recommendations that can be implemented from this thesis. The first is that
to limit scale formation the temperature needs to be kept low. Values greater than 1070 °C for
the average Roughing Mill and above 1050 °C for the average Crop Shear temperature are
considered high, with values greater than this increasing the chance o f scale formation. As the
temperature increases more scale suppression measures are required to limit scale formation,
with high temperatures more likely to generate a greater amount o f scale even with fully
functional scale suppression systems in place.

Chemistry is also a significant factor in scale formation, with Phosphorus being the most
significant o f the chemistry variables. It is recommended that the chemistry specification for
Phosphorus be limited to a maximum value o f 0.015 % rather than 0.020 % to limit scale
formation. Slabs with higher values should be treated with particular care when being
processed through the Hot Mill to limit scale formation.

It is recommended that the monitor that has been deployed continue to be utilised and
updated, to aid in the removal o f scale as a significant defect and to improve communication
between business units.