This action might not be possible to undo. Are you sure you want to continue?

**Contents lists available at SciVerse ScienceDirect
**

Journal of Materials Processing Technology

j our nal homepage: www. el sevi er . com/ l ocat e/ j mat pr ot ec

Identiﬁcation, for control, of the process parameters inﬂuencing tertiary scale

formation at the hot strip mill using a binary choice model

J. Kennedy

a,1

, M. Evans

b,∗

, F. Robinson

c

a

The Engineering Doctorate Centre in Steel Technology, Materials Research Centre, College of Engineering, Swansea University, Singleton Park, Swansea SA2 8PP, UK

b

College of Engineering, Singleton Park, Swansea, SA2 8PP, UK

c

Tata Steel Strip Products UK, Port Talbot Works, Port Talbot SA13 2NG, UK

a r t i c l e i n f o

Article history:

Received 16 May 2011

Received in revised form27 February 2012

Accepted 2 March 2012

Available online 10 March 2012

Keywords:

Tertiary scale

Principal component analysis

Binary choice

Probit and Logit models

a b s t r a c t

Scale is highly detrimental to the surface quality of tinplate products and this problem is created during

the hot rolling process. In this paper, a statistical analysis is carried out to both quantify the inﬂuence

of hot mill process parameters on tertiary scale formation and to use this information to identify how

to optimise the running conditions at the hot mill. That is, to identify what base chemistries to use and

what temperatures to run the hot mill at so that the percentage of coils produced with signiﬁcant scale

formation is kept as low as is required for meeting customer requirements. Principal component analysis

is used to reduce the dimensionality of the dataset whilst retaining the majority of the variability in the

process variables. It was found that a Logit model containing these components was consistent with the

process data and it was further determined from this model that the most signiﬁcant variables were

the temperature at which the steel strip was entering the ﬁnishing mill, together with the percentage of

phosphorus, copper and nickel present within the strip. More importantly, for process control it was found

that at low levels of phosphorus (<0.011%) there is a less than 10% chance of signiﬁcant scale formation

when the other hot mill variables are kept at their average values. It was also found that to keep the rate

of defective coils below 10%, the maximum rougher mill temperature should be kept at 1040

◦

C or less.

The Logit model was capable of identifying many other optimal running conditions for the hot mill.

© 2012 Elsevier B.V. All rights reserved.

1. Introduction

For high-end ﬂat steel products scale is clearly detrimental

to surface quality. Scale is formed at the hot mill but it has its

greatest impact on quality after further processing. Scale can have

a particularly negative impact on quality for tinplate products. This

occurs due to the scale interfering with the interface between the

steel substrate and the tin, resulting in a product surface that is not

suitable for the tight tolerances required by customers of tinplate

products. Whilst it is well known which process variables are the

major determinants of scale formation (for example phosphorous

and temperature), it is less well known at precisely which levels

to set these important variables so as to minimise the chances

of scale forming at the hot mill. Whilst this information is very

important for maximising the number of rolled coils that can be

used for high end applications, quantifying these values for the

key process variables is complicated by the fact that there is more

than one temperature at the hot mill that could inﬂuence scale,

∗

Corresponding author. Tel.: +44 1792 295748; fax: +44 1792 295514.

E-mail address: m.evans@swansea.ac.uk (M. Evans).

1

Tel.: +44 1792 295748; fax: +44 1792 295514.

and more importantly, these temperatures are highly correlated

with each other making it difﬁcult to disentangle their individual

contributions to scale formation.

This work will use principal component analysis together with

some binary choice models to identify not only what the impor-

tant chemistries andtemperatures are, but alsothenumerical effect

of changing these processing parameters on the quantity of scale

forming at the hot mill. With so many different temperature vari-

ables, the former technique is used to reduce data dimensionality

and hence the correlation between these process variables and

the latter technique is used because the predictions froma binary

choice model are in the formof the % of coils with serious surface

defect issues. The aim is not just to identify the important tem-

perature variables but more importantly to ﬁnd what temperature

values and chemical additions should be avoided when trying to

produce high end sheet steel.

Haapamaki et al. (2005) used data mining techniques to inves-

tigate scale defect predictions. Haapamaki et al. (2005) concluded

that predictions obtained fromreal world data were not as good as

expected. Haapamaki et al. (2005) account for this by saying that

the successful measurement of scale (using the Parsytec surface

inspectionsystem) does not account for the size or severity of scale.

This weakness of the Parsytec systemis well known, and it should

be noted that even if the size or severity of individual scale defects

0924-0136/$ – see front matter © 2012 Elsevier B.V. All rights reserved.

doi:10.1016/j.jmatprotec.2012.03.003

J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630 1623

Fig. 1. Port Talbot Hot Mill layout.

cannot beknown, thefrequencyof scalecounts onthestripis agood

indication of scale severity. Haapamaki et al. (2005) also note that

some of the data sources couldbe inaccurate due totheir positionin

the process, i.e. scale thickness effecting surface temperature at the

crop shear because it is positioned before the scale breaker. Inaccu-

racies caused by process setup are concerns, however the effect of

these types of inaccuracies can be greatly minimised by using only

one grade of steel that has a tightly controlled chemistry and set of

processing conditions.

This paper will use over 1500 records of the same steel grade

and same gauge (thickness) to minimise the effects of processing

conditions. Due to the inﬂuence of scale on further processing, the

best way of minimising the impact of scale is by understanding

the inﬂuence of process variables at the hot mill. This will then

enable control engineers to set the process variables at levels that

minimise the formation of scale.

2. Literature reviewon scale type and formation

2.1. Scale

Scale is an oxide layer that builds up on the surface of the steel

when it is exposed to environmental conditions. Oxide can format

any point during the life cycle of the product; however it is par-

ticularly likely to occur at the hot mill. This is due to the high

temperatures at the hot mill, which increases the speed of any

chemical reaction. Bolt (2003) states that the important properties

to consider for the oxide layer are thickness, composition, adhesion

and structure. Bolt (2003) also mentions that structure and cohe-

sion are of importance to scale formation, but that these properties

are of less importance.

The thickness of the oxide layer will be determined by its

processing conditions. Bolt (2003) states that the most impor-

tant parameters for thickness are the ﬁnishing temperature and

gauge of the coil. Depending on the temperature of the strip, oxide

could be formed at the run-out table (ROT) and also during coil-

ing but that scale formation at these stages is rather modest. Bolt

(2003) also concluded that thickness is of particular importance

to pickling lines, because the thicker the oxide is the easier it is

to promote cracking. This is supported in the work by Yang et al.

(2008), where they discuss the inﬂuence of silicon on thickness.

Yang et al. (2008) investigated steels with silicon contents ranging

from0.01 to 1.91wt.%. Yang et al. (2008) concluded that oxidation

rates and scale thickness decreased with increasing silicon. They

also determined that different oxides are formed at different per-

centages of silicon. This work is useful because it demonstrates the

importance of limiting the chemistry range to avoid the formation

of scales that have different properties. Taniguchi et al. (2001) fur-

ther demonstrates the importance of limiting the chemistry range.

Taniguchi et al. (2001) found that when silicon is at a high percent-

age (1.14wt.%), the penetration depth increases with increasing

silicon. This further adds to the understanding that for a successful

analysis to be carried out the chemistry range should be promot-

ing the same type of scale. The work of Munther and Lenard (1999)

concluded that thick scale provides a degree of lubrication to the

work rolls, whilst a thin scale has a higher coefﬁcient of friction and

is more adherent to the metal substrate, which makes it harder to

remove.

The composition of these formed oxides inﬂuences the prop-

erties of the scale. The classical three-layer scale model reveals

that when pure iron is oxidised under normal conditions wustite

(FeO) is formed closest to the steel substrate followed by magnetite

(Fe

3

O

4

) and hematite (Fe

2

O

3

). Each of these oxides has their own

individual properties. Bolt (2003) states that at hot rolling tem-

peratures, wustite has a high plasticity, but is extremely brittle at

room temperature, which means that a large amount of wustite

will result in poor scale cohesion at roomtemperatures. Hematite

is extremely hard and brittle and is highly detrimental during all

stages of processing due to the increased work roll wear its causes

and the poor strip quality it creates. Magnetite is the phase that

is least brittle at room temperature so promotes scale integrity.

This phase can be tolerated in small quantities, because it is not as

hard as hematite and shows some plasticity at hot rolling temper-

atures. Thinner scales are more adherent than thicker scales and

the reason for their increased adherence is the higher amounts of

hematite and magnetite. The adherence of the scale is proportional

to the amount of magnetite and hematite present in the scale. This

is because magnetite and hematite are a lot harder than wustite

and so is harder to remove through further processing of the coil.

The composition of the scale that develops is also highly tem-

perature dependent. At temperatures greater than 570

◦

C the main

composition of scale is wustite, which is generally the thickest

layer. This is then followed by magnetite and hematite. The lit-

erature on what composition of scale is observed at different

temperatures is highly variable, mostly due to different hot mill

layouts, different types of steel, different experimental methods,

and different temperatures. The literature review of Bolt (2000)

concludes that the mechanisms that will occur are heavily related

to the steel composition and the temperature the steel strip is at.

However, generally speaking, as the temperature is increased the

more likely it is that the scale formed will be composed of a higher

percentage of magnetite and hematite at the expense of wustite.

The increase inmagnetite andhematite will increase the adherence

of the scale and make it harder to remove.

2.2. Temperature

The thickness of the scale is proportional to temperature and

time. Sun et al. (2004) observed that as temperature and oxidation

time increase the scale thickness also increased. It was observed

that for the ﬁrst 20s the scale increased in thickness according to

linear growth, and then by parabolic growth after that. Sun et al.

(2004) explain this observation by the transport mechanismof oxi-

dation. Initially the scale is very thin which allows rapid diffusion

of 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.

2.3. Phosphorus

If the diffusion of iron into the scale is suppressed then mag-

netite and hematite will form in preference to wustite, and this

1624 J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630

Table 1

Process variables and their sample means and standard deviations.

Process variable x

j

Mean Standard deviation

% carbon x

1

0.069 0.0042

% manganese x

2

0.485 0.0226

% phosphorus x

3

0.014 0.0031

% sulphur x

4

0.014 0.0030

% silicon x5 0.005 0.0032

% copper x

6

0.016 0.0059

% nickel x7 0.010 0.0030

% chrome x

8

0.018 0.0040

% aluminium x

9

0.034 0.0047

% tin x

10

0.003 0.0018

% nitrogen x

11

0.012 0.0009

% solAl x

12

0.031 0.0043

MaximumRMtemperature (

◦

C) x

13

1121.75 17.87

Average RMtemperature (

◦

C) x

14

918.34 11.97

MinimumRMtemperature (

◦

C) x

15

1099.43 16.18

MaximumCS temperature (

◦

C) x

16

1099.88 19.89

Average CS temperature (

◦

C) x

17

951.69 48.59

MinimumCS temperature (

◦

C) x

18

1069.55 20.80

MaximumFMtemperature (

◦

C) x

19

907.16 7.18

Average FMtemperature (

◦

C) x

20

829.87 28.96

MinimumFMtemperature (

◦

C) x

21

881.05 5.75

ROT temperature (

◦

C) x

22

731.35 25.13

Scale count indicator y 0.327 0.469

RM=rougher mill, CS =crop sheer, FM=ﬁnishing mill and ROT=run out table. y =1

when the scale count on the bottomof the sample exceeds 200 and zero otherwise.

makes it harder to remove the scale. The composition of the scale

canbe inﬂuencedbythe steel composition. For tertiaryscale forma-

tion, Bolt (2000) states that phosphorus is of particular importance.

This is because phosphorus can suppress diffusion of iron into the

scale. This suppression can, locally, completely cut off the diffusion,

which means as the steel is oxidised magnetite and hematite will

formin preference to wustite. This results in the scale being harder

and more difﬁcult to remove.

3. The hot mill process

There is no single variable that can be attributed to scale forma-

tion, so to be able to determine variables that are signiﬁcant inscale

formation at the hot mill, it is important to identify all the process

variables through a detailed description of the hot mill process.

A schematic view of this process is shown in Fig. 1. The temper-

ature is the most important variable for scale formation as scale

increases exponentially with time at high temperatures. The most

critical area of the hot mill for scale formation is between the crop

shears and the ﬁnishing mill. To be able to understand the impact

of this critical area on scale formation, multiple temperature read-

ings are made for each coil including an average value, a maximum

and minimumtemperatures across the coil and the temperature at

the front and back of the coil. The chemistry of the steel substrate

also has a strong inﬂuence on the type of scale that will formin the

hot mill. The chemistry of the coil is measured on a cast basis at the

Basic Oxygen Steel making plant.

As illustrated in Fig. 1, steel with a speciﬁc alloying composi-

tion is reheated in furnaces and passed through scale breakers. Let

variables x

1

–x

12

be the amounts of various alloying elements, by

wt.%, contained within this steel. At the edging and rougher mills

the gauge of the incoming slab is reduced from250mmto 35mm

(and the processed steel at this stage is often termed the trans-

fer bar). The hot mill at the Tata works in Port Talbot (UK) has a

reversing rougher hence several passes are necessary toachieve the

required transfer bar thickness. The transfer bar is then held at the

coil boxtoincrease the heat of the bar prior toentering the ﬁnishing

mill. The length of time the transfer bar can be held at the coil box

depends on the temperature that the transfer bar is required to be

at entry to the ﬁnishing mill. The temperature of the transfer bar is

dependent on the dimensions of the bar, the processing history

(after the reheat furnace), and the time the transfer bar takes in

completing the process. Let variables x

13

–x

15

be various tempera-

tures (as described above) recorded at the rougher mill.

The crop shear temperature is the temperature just before the

transfer bar enters the ﬁnishing mill. The crop shear temperature

is likely to be of vital importance for predicting scale formation

because tertiary scale is formed during the early stages of the ﬁn-

ishing mill. Let variables x

16

–x

18

be various temperatures recorded

at the crop shear stage of the process. On entry into the ﬁnishing

mill the slab encounters the descalers and the purpose of these is

to remove scale on the surface of the coil during the hot rolling pro-

cess. Silk (2001) investigates descaler effectiveness and determines

the main components of successful descaling, which include the

impact pressures used and the angle of the descaler conditioning

sprays. Understanding the impact of the descaler is an important

section of understanding the inﬂuence of scale removal, but will

not be consideredinthis paper. Due tothe statistical methods being

used, theinﬂuenceof thedescaler will bedifﬁcult toextract without

prior knowledge of the inﬂuence of temperature and chemistry.

The purpose of the ﬁnishing mill is to further reduce the gauge

to between 1.4 and 18.0mm. The ﬁnishing mill also controls the

shape and temperature of the strip that in turn controls the met-

allurgical properties achieved by the strip. Let variables x

19

–x

21

be

the various temperatures recorded at the ﬁnishing mill. The phase

transformation occurs on the run-out table (ROT) and the cooling is

controlled with a water cooling system. The temperature control of

the runout table is vital incontrolling the austenitic phase transfor-

mation. The temperature, cooling rate and cooling path will affect

the microstructure present, which in turn will control the mechan-

ical properties of the steel. Let variables x

22

be the temperature

recorded at the ROT. However, the ROT temperature control sprays

are not effective at removing scale, because the scale is formed at

the entrance to the ﬁnishing mill.

4. Data collection

TheParsytec inspectionsystemis acameraoperatedsystemthat

detects defects on the surface of the coil. The detection of defects is

achieved by the identiﬁcation of all non-homogeneous sections of

the strip. When the affected sections are detected then the Parsytec

software classiﬁes the detected sections into features. These fea-

tures are compared to the online defect catalogue and then the

best result is what the detected section is labelled as the appropri-

ate defect. It is possible for defects to be misclassiﬁed (MC) or not

classiﬁed (NC). This can occur because the environment around the

cameras has a lot of potential contamination including:

1. Water fromthe descaler and cooling sprays at the ﬁnishing mill.

2. Overlapping defects.

3. Substrate appearance. Every steel compositionwill have a differ-

ent surface appearance under the Parsytec cameras that makes

detection rates variable.

The Parsytec system monitors the top and the bottom of the

strip, but they are located in different locations on the mill. The top

surface monitor is located at the start of the ROT, whilst the bottom

surface monitor is located at the end of the ROT just before coil-

ing. Due to the fact that under high coiling temperatures scale can

continue to growon the ROT and that due to water and other con-

tamination after leaving the ﬁnishing mill it was determined that

the bottomsurface monitor’s scale count would give more reliable

results.

The dataset is constructed using data associated with coils from

Tata Steel Europe’s Port Talbot Hot Mill, collectedusingthe Parsytec

J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630 1625

systemdescribed above. Each of the coils in the dataset is intended

for a high surface quality tinplate application, so scale levels must

be low. To minimise the effect of different heating/cooling cycles,

only a single gauge of 2.1mm is investigated in this paper. Coil

data was collected between September 2009 and March 2010 for

all the process variables described above and Table 1 gives some

descriptive statistics for this sample of data, together with the full

name of eachprocess variable and its short hand descriptor—x

j

. The

sample size was n=1577.

The dependent variable, y, is a binary classiﬁcation based on the

scale count fromthe Parsytec system. The dependent variable takes

on a value of unity when the scale count on the bottomof the table

exceeds 200 and zero otherwise. The cut of 200 was chosen for two

main reasons. The ﬁrst reason was that lower scale counts than this

are unlikely to be a problemfor 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 howadherent the scale is. The second reason is that

a split of 200 makes the population of bottomsurface affected coils

represent more than 30% of the dataset which makes it easier to

be conﬁdent about the output of the model, i.e. it creates a more

balanced data set from which to apply the binary choice model

too.

5. Statistical methodology

5.1. Principal component analysis

Principal component analysis (PCA), see for example Jolliffe

(2002) for a good review of this technique, is an extensively used

technique of multivariate linear data analysis. The purpose of PCAis

to reduce the dimensionality of the data but still maintain its vari-

ability and to remove the correlationamongst the process variables

used to control the hot mill process.

The principal components (PC) are calculated as follows. The

p process variables (x

1

–x

p

) are standardised to have a zero mean

and unit variance to avoid the problems associated with units of

measurement. Then p linear combinations of these variables can

be formed as follows

PC

1

= a

11

z

1

+a

12

z

2

+· · · +a

1p

z

p

PC

2

= a

21

z

1

+a

22

z

2

+· · · +a

2p

z

p

.

.

.

PC

p

= a

p1

z

1

+a

p2

z

2

+· · · +a

pp

z

p

(1a)

where the z variables are the standardised values of the x process

variables. The so called loadings, a

11

to a

1p

, are then chosen so as to

maximise the variance of PC

1

subject to the normalising condition

a

2

11

+a

2

12

+· · · +a

2

1p

= 1 (1b)

PC

1

is then said to be the ﬁrst principal component and is a

linear function of the zs (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 zs is picked up by the p principal

components. That is

variance(PC

1

) +variance(PC

2

) +· · · +variance(PC

p

) = variance(z

1

)

+variance(z

2

) +· · · +variance(z

p

) = p (1b)

with

variance(PC

1

) >variance(PC

2

) >· · · >variance(PC

p

)

Various different algorithms can be used to ﬁnd values for these

loadings, ranging fromthe 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 of

the correlation matrix amongst all the x variables (see for example

the text by Mardia et al., 1979). The Eigen values fromthis decom-

positionmeasurethevariationinall theprocess variables explained

by each principal component. Thus the ﬁrst Eigen value 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 prin-

cipal 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 anEigenvalue greater than1 shouldbe consid-

ered as essential and therefore retained in the analysis. This simple

rule allows for a substantial reduction in the number of variables

included in the ﬁnal statistical model.

5.2. Binary choice models

5.2.1. A generalised framework

A binary choice model views the outcome of a discrete incident

(such as the occurrence of a large amount of scale) as a reﬂection of

an underlying regression. The cost associated with any additional

processing or downgrading of a heavily scaled coil (deﬁned above

as 200+ observations of scale) is modelled using an unobserved

variable y

∗

i

such that

y

∗

i

= ˇ

0

+

p

j=1

ˇ

j

x

ij

+ε = x+ε

i

(2a)

where x

ij

is the ith measurement on process variable j associ-

ated with the hot mill and the ˇ

j

are the parameters that are to

be estimated from the data set.  is a vector containing the p+1

parameters of this model and x is the data matrix with the columns

of this matrix containing each of the process variables (including

an initial column of 1s) and each rowcontains the ith measurement

on these variables. ε

i

is assumed to follow some standardised dis-

tribution with known variance—to be discussed further below. The

cost of having large amounts of scale is not directly observed, just

whether the coil is heavily scaled or not. Therefore observations at

the hot mill take the form

y

i

= 1 if y

∗

I

>0

y

i

= 0 if y

∗

I

≤ 0

(2b)

Provided that a constant (ˇ

0

) is included in Eq. (2a) the assump-

tion of a zero threshold for y

∗

i

is an innocent one. The process

variables are then taken to determine the probability of y

∗

i

>0

Prob(y

∗

I

>0|x) = Prob(y

i

= 1|x) = F(x, )

Prob(y

∗

I

≤ 0|x) = Prob(y

i

= 0|x) = 1 −F(x, )

(2c)

where F(x, ) is a function, dependent on x and , that determines

the probability that y

i

=1 given the values in x, with the natural

requirement that F(x, ) should tend to 1 as x

 tends to inﬁnity

(and F(x, ) tends to zero as x

** tends to minus inﬁnity). There are
**

many cumulative density functions that ﬁt these criteria, but most

applications of this type of model are conﬁned to the use of the

normal distribution (giving rise to the Probit model—see Maddala,

1991 for a good introduction to this model) and the logistic dis-

tribution (giving rise to the Logit model—see Maddala, 1991 for a

good introduction to Logit model).

The following model canbe seenas a generalisationof these two

binary choice models that enables discrimination between these

and other functions through use of standard likelihood ratio tests.

1626 J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630

FromEqs. (2a)–(2c) it follows that

Prob(y

i

= 1) = P

i

= Prob

_

_

_

ε

i

>−

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

_

_

_

= 1 −F

_

_

_

−

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

_

_

_

(3a)

Eq. (3a) makes it clear that F is the cumulative distribution func-

tion for ε. One very general speciﬁcation can be derived from the

generalised F distribution. Prentice (1975) has suggested a non

degenerate and standardised version of this distribution where the

probability density function is given by

f (w) =

(m

1

/m

2

)

m

1

{(m

1

)(m

2

)}/(m

1

+m

2

)

exp(m

1

w)

(1 +{m

1

/m

2

}e

w

)

(m

1

+m

2

)

(3b)

where

w

i

= −

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

(3c)

and where f( ) is the derivative of F( ) with respect to w and ( ) is

the gamma function. In Eqs. (3a)–(3c), the parameters m

1

and m

2

determine the cumulative distribution function for w

i

and thus ε

i

.

That is, m

1

and m

2

determine the formof the binary choice model

suchas the Logit andProbit model. Unfortunately, there is noclosed

formexpression for F( ) within this setting except for some special

cases. To take some well known examples, ﬁrst consider the case

where m

1

=m

2

=1. Then

1 −F[w

i

] =

exp(w

i

)

1 +exp(w

i

)

(4a)

which is of course the cumulative (standardised) logistic function

giving rise to the Logit model which can also be expressed as the

log odds model

ln

_

P

i

1 −P

i

_

= −

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

(4b)

For ﬁnite m

1

and m

2

but with m

1

=m

2

this logistic function

becomes the generalised logistic. Then when m

1

=1 and m

2

=0

1 −F[w

i

] = 1 −exp[−exp(w

i

)] (5a)

which is of course the cumulative (standardised) extreme value

function giving rise to the proportional hazards model which can

also be expressed as

ln[−ln(1 −P

i

)] = −

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

(5b)

Other special cases include the Probit model when m

1

=∞and

m

2

=∞and the generalised gamma cumulative function when just

m

2

=∞(see Lawless, 2003 for a descriptions of this function).

Whilst there is no closed formexpression for the integral of Eq.

(3b) whenm

1

andm

2

are not restrictedto the special cases outlined

above, values for this integral at a givenvalue for w

i

canbe obtained

numerically. If m

1

and m

2

are both ﬁnite, values for this integral at

a given value for w

i

are found by calculating areas under the Beta

distribution with m

1

and m

2

degrees of freedom. More precisely,

P

i

= 1 −F

_

_

_

−

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

_

_

_

= B(s

i

, m

1

, m

2

) (6a)

where B(s

i

, m

1

, m

2

) is the area under the Beta distribution with m

1

and m

2

degrees of freedomto the left of s

i

, where

s

i

=

_

1 +

m

1

exp[w

i

]

m

2

_

−1

(6b)

For ﬁnite m

1

and m

2

=∞, values for this integral at a given value

for w

i

are foundby calculating areas under the Chi distributionwith

2m

1

degrees of freedom. More precisely,

P

i

= 1 −F

_

_

_

−

_

_

ˇ

0

+

p

j=1

ˇ

j

x

ij

_

_

_

_

_

= (2v

i

, 2m

1

) (6c)

where (2v

i

, 2m

1

) is the area under the Chi distribution with (2m

1

)

degrees of freedomto the left of 2v

i

, where

v

i

= m

1

exp

_

w

i

√

m

1

_

(6d)

The integrals calculated using Eqs. (6c) and (6d) tend to those

obtained fromthe standard normal distribution (and hence yields

the Probit model) as m

1

tends to ∞. For ﬁnite m

2

and m

1

=∞, the

procedure is to replace v

i

with −v

i

and to replace m

1

with m

2

in

Eqs. (6c) and (6d). Most statistical packages have routines capable

of calculating such areas and as demonstrated by Prentice (1975)

this is best achieved by reparametrising m

1

and m

2

as

p = (m

1

+m

2

)

−1

and q = (m

−1

1

+m

−1

2

)(m

−1

1

+m

−1

2

)

−1

(6e)

5.2.2. Parameter estimation

Since the y

i

are just realisations of a binomial process with

probabilities given by Eqs. (6a) or (6c) we can write the likelihood

function as

L =

y

i

=1

P

i

y

i

=0

(1 −P

i

) (7a)

so that the log likelihood is given by

log[L] =

n

i=1

y

i

log(P

i

) +

n

i=1

(1 −y

i

)log(1 −P

i

) (7b)

where n is the sample size of the data set and P

i

is calculated from

Eqs. (6a) or (6c) for given values of m

1

and m

2

. Given these values

for m

1

and m

2

, the ˇ parameters in Eqs. (6a) and (6c) are chosen so

as to maximise the log likelihood in Eq. (7b).

5.2.3. Model identiﬁcation

Taking a general to speciﬁc approach to model identiﬁcation,

all the process variables shown in Table 1 are initially included

in the model. Then a grid search is carried out at over different

values for m

1

and m

2

using this general model. The log likelihood

deﬁned by Eq. (7b) at each (m

1

, m

2

) combination is stored and the

largest of them recorded. Call this statistic log[L]

max

. Clearly the

(m

1

, m

2

) value combination associated with this value determines

the distributionmost supportedby the data. However, any (m

1

, m

2

)

combination that produces

r

= −2[log(L)

max

−log(L)

r

] (8a)

with a value less than 3.84 is also supported by the data at the 5%

signiﬁcance level. log[L]

r

is the log likelihood associated with the

rth (m

1

, m

2

) combination. This follows fromthe fact that

r

, under

the null hypothesis that the rth (m

1

, m

2

) combination is the true

combination, is asymptoticallyChi squaredistributedwith1degree

of freedom.

Once the correct values for m

1

andm

2

are determinedinthe way

describedabove, adatabasedsimpliﬁcationof themodel thentakes

place. Themodel associatedwithlog(L)

max

is calledtheunrestricted

model. If the jth explanatory variable is then removed from this

J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630 1627

Table 2

Correlation matrix for the temperature process variables.

x

13

x

14

x

15

x

16

x

17

x

18

x

19

x

20

x

21

x

22

x

13

1

x

14

0.92 1

x

15

0.04 0.13 1

x

16

0.76 0.73 −0.01 1

x

17

0.53 0.57 0.01 0.69 1

x

18

0.30 0.26 0.04 0.38 0.3 1

x

19

0.46 0.42 −0.03 0.61 0.46 0.23 1

x

20

0.48 0.44 −0.04 0.63 0.51 0.26 0.70 1

x

21

0.16 0.12 0.03 0.18 0.12 0.13 0.16 0.25 1

x

22

0.06 −0.03 0.02 0.02 0.06 −0.01 −0.01 0.08 0.02 1

model, a restricted model is obtained with a log likelihood given by

log(L)

R

. This is achieved by removing that x

j

with the smallest stu-

dent t statistic. The ﬁnal simpliﬁed model is obtained by removing

all those variables for which

R

= −2[log(L)

max

−log(L)

R

] (8b)

has a value in belowof 3.84. Finally, the adequacy of this parsimo-

nious model canbeassessedinoneof twoways. First theproportion

of correct predictions can be calculated, where a prediction is con-

sidered correct if Eqs. (6a) or (6c) yields a value above 0.5 when

y

i

=1 or alternatively when Eqs. (6a) or (6c) yields a value equal

to or below 0.5 when y

i

=0. The count R

2

deﬁned in Eq. (8c) then

lies between 0 and 1 with better performing models having a value

closer to 1

count R

2

=

number of correct prediction

total number of observations

(8c)

6. Results and evaluation

6.1. Principal component testing procedure

Using principal component analysis the correlations existing

between all the process variables were investigated. The analysis

was performed to reduce both the number of process variables

and to remove the correlations amongst the process variables

used in the model to predict scale count and from this to gain an

understanding of 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 fromforming the principal com-

ponents of these chemistries. However, Table 2 shows that all the

temperature variables are highly correlated—especially for exam-

ple between the maximumand average rougher mill temperatures

(betweenx

13

andx

14

). Ingeneral, it canbeseenfromTable2that the

most highly correlated relationships occurred between the maxi-

mum and average temperatures at the rougher mill and at the crop

shear.

Given these high levels of dependency between the tempera-

ture process variables, principal components of these temperatures

need to be constructed to remove these correlations. Table 3 shows

howmuch of the variability in all the temperature variables shown

in Table 1 are accounted for by all the different principal compo-

nents.

The ﬁrst three principal components accounted for nearly 64%

of the variability present in all ten temperature process variables,

whilst the ﬁrst four principal components accounted for nearly

three quarters of the variability present in all ten temperature pro-

cess variables. Using Cattell’s rule (1952) of only needing to use

those principal components having an Eigen value of around unity

or more, the ten temperature process variables can be replaced by

Table 3

Principal component results for the temperature process variables.

Components Eigen values Variation (%) Cumulative variation (%)

PC

1

4.201 42.01 42.01

PC

2

1.106 11.06 53.07

PC

3

1.043 10.43 63.51

PC

4

0.9741 9.74 73.25

PC5 0.8418 8.42 81.67

PC

6

0.7769 7.77 89.44

PC7 0.4999 5.00 94.44

PC

8

0.2805 2.81 97.24

PC

9

0.2040 2.04 99.28

PC

10

0.0718 0.72 100.00

just the ﬁrst three or may be the ﬁrst four principal components.

These principal components had the form

PC

1

= −0.0233(x

13

−1121.8) −0.0251 (x

14

−1099.4)

−0.0016(x

15

−918.3) −0.0223(x

16

−1099.9)

−0.0179(x

17

−1069.6) −0.0045(x

18

−951.7)

−0.0493(x

19

−907.2) −0.0644(x

20

−881.1)

−0.0045(x

21

−829.9) −0.0003(x

22

−731.4) (9a)

PC

2

= 0.0156(x

13

−1121.8) +0.0214(x

14

−1099.4)

+0.0431(x

15

−918.3) +0.0008(x

16

−1099.9)

−0.0007(x

17

−1069.6) +0.0008(x

18

−951.7)

−0.0384(x

19

−907.2) −0.0622(x

20

−881.1)

−0.0116(x

21

−829.9) −0.0185(x

22

−731.4) (9b)

PC

3

= −0.0037(x

13

−1121.8) +0.0005(x

14

−1099.4)

+0.0597(x

15

−918.3) −0.0027(x

16

−1099.9)

+0.0003(x

17

−1069.6) +0.0022(x

18

−951.7)

−0.0149(x

19

−907.2) −0.0007(x

20

−881.1)

+0.0116(x

21

−829.9) +0.0235(x

22

−731.4) (9c)

PC

4

= 0.0043(x

13

−1121.8) +0.0080(x

14

−1099.4)

+0.0085(x

15

−918.3) +0.0040(x

16

−1099.9)

+0.0078(x

17

−1069.6) −0.0054(x

18

−951.7)

−0.0047(x

19

−907.2) −0.0086(x

20

−881.1)

−0.0249(x

21

−829.9) +0.0233(x

22

−731.4) (9d)

When the loadings in front of some variables are similar, then

instead of 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 prin-

cipal component, the loadings in front of the maximum rougher

mill (x

13

), the average rougher mill (x

14

) and the maximum crop

shear (x

16

) 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 of the minimum rougher mill (x

15

) and the average crop

shear (x

17

) temperatures are very similar, suggesting that an

appropriate dimensionality reduction would be to look at the dif-

ferences between these two temperatures. The loadings in front

1628 J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630

-570

-568

-566

-564

-562

-560

-558

1 0.75 0.5 0.25 0

M

a

x

i

m

i

s

e

d

l

o

g

l

i

k

e

l

i

h

o

o

d

p (as defined in Eq. (6e))

q = 1

q= 0.75

q= 0.5

q = 0.25

q = 0

Logit model

Probit model

5% significance level for values of p and q supported by the data

Binary choice model most spported by the data

Fig. 2. The maximised log likelihood obtained at different values for p and q.

of the minimum crop shear (x

18

) and the minimum ﬁnishing mill

(x

21

) temperatures are very similar, suggesting that an appropri-

ate dimensionality reduction would be to look at the differences

between these two temperatures. The loadings in front of the

maximum (x

19

) and average ﬁnishing mill (x

20

) temperatures are

also very similar, suggesting that an appropriate dimensionality

reduction would be to look at the differences between these two

temperatures.

6.2. Distribution analysis

The ˇ parameters in Eqs. (6a) or (6c) were estimated by max-

imisingthe loglikelihoodinEq. (7b) at givenvalues for pandq(with

p and q deﬁned by Eq. (6e)). All the chemistry variables in Table 1

were included in Eqs. (6a) or (6c), and all the temperature variables

were replaced by the ﬁrst three principal components speciﬁed in

the previous section. Parameter estimates are not shown for the

sake of briefness but these results are available from the authors

uponrequest. Fig. 2shows themaximisedloglikelihoods associated

with different values for p and q. It can be seen fromthis ﬁgure that

the largest maximised log likelihood occurs where p=1 and q=0.5.

The area of the graph below the dashed line in Fig. 2 shows which

maximised log likelihoods are signiﬁcantly different fromthis one

at the 5% signiﬁcance level using the test statistic in Eq. (8a). It is

clear from this ﬁgure 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 com-

putational aspects of the work simpler, all further results shown

beloware based on the Logit model.

6.3. Logit model results

Using the data based simpliﬁcation procedure described in Sec-

tion 5.2.3, a simpliﬁed Logit model was derived and this model is

summarised in Table 4. Only a few chemistries were statistically

Table 4

Logit model result.

Process variable Coefﬁcient Std. error t-Value

Constant −2.49 0.46 −5.38

Percentage of phosphorus (x

3

) 228.28 25.13 9.08

Percentage of silicon (x5) −88.87 22.25 −3.99

Percentage of copper (x

6

) 72.66 13.33 5.45

Percentage of nickel (x7) −184.66 28.79 −6.41

Percentage of chrome (x

8

) −64.12 19.74 −3.25

PC

1

−0.391 0.04 −10.3

PC

2

−0.232 0.07 −3.32

Table 5

Quasi-elasticities.

Process variable Quasi-elasticity

Percentage of phosphorus (x

3

) 0.62

Percentage of silicon (x5) −0.09

Percentage of copper (x

6

) 0.24

Percentage of nickel (x7) −0.39

Percentage of chrome (x

8

) −0.22

PC

1

−0.08

PC

2

−0.05

signiﬁcant at the 1% signiﬁcance level, with these being phospho-

rous, silicon, copper, nickel andchrome. Further, only 2of the above

derived principal components were statistically signiﬁcant at this

level. Further, of the n=1577 observation of y, 1140 of these were

predicted correctly by the model. This is an overall correct predic-

tion rate of 73%. However, the model was signiﬁcantly better at

predicting the hot mill conditions leading to no signiﬁcant scale

formation (i.e. y

i

=0), as the model predicts when y

i

=0, some 98%

of the time.

FromTable 4 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 chrome. Phosphorous and copper on the other

hand have a positive impact on this ratio. For a feel for which of

these variables has the biggest effect on the probability of scale

forming at the hot mill it is better to convert the coefﬁcient val-

ues in Table 4 into quasi-elasticities. These elasticities, contained

in Table 5, show the percentage change in the probability of scale

forming following a percentage change in each of the process vari-

ables. These elasticities are calculated around the mean values –

shown in Table 1 – for each process variable as the elasticity will

vary with the magnitude of each process variable.

FromTable 5 it can be seen that when the phosphorous content

is at 0.014wt.% (i.e. at its mean value shown in Table 1) a further

1% increase in the phosphorous content will increase the proba-

bility of scale formation by just over 0.6 percentage points. It can

also be observed from this table that when the nickel content is

at 0.01wt.% (i.e. at its mean value shown in Table 1) a further 1%

increase in the nickel content will decrease the probability of scale

formation by just under 0.4 percentage points. Given the size of the

other elasticities in Table 5, these are therefore the most impor-

tant 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 of scale forming when the temperature

variables are at their mean values shown in Table 1.

As for the temperature variables, it canbe seenfromTable 5 that

when the ﬁrst principal component is at its mean value a further

1% increase in this component will decrease the probability of scale

formation by just under 0.1 percentage points. As temperatures

enter the ﬁrst principle component with negative loadings (see Eq.

(9a) above for PC

1

), this implies temperature has a positive impact

of the probability of scale forming. This relationship is looked at in

more detail below, but before then it is important to look next at

some probability plots as the above elasticities only apply at the

mean values for each process variable. To get the full picture, the

complete range of values must be looked at and this is shown in

the probability plots of Fig. 3.

In Fig. 3 the probabilities shown on the vertical axis are calcu-

lated fromEqs. (6a) or (6c) by varying one of the process variables

whilst holding the other process variables at their mean values.

These probabilities canalsobe interpretedas the proportionof coils

beenmanufacturedthat have a signiﬁcant number of ﬂaws—sopre-

venting them being used for the high value added markets. Plots

like those shown in Fig. 3 can therefore be used for optimising

the process, i.e. ﬁnding how to run the hot mill so that only small

proportions of manufacture coils have scale counts that prevent

J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630 1629

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 a b

c d

0.025 0.02 0.015 0.01 0.005 0

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Phosphorous content, %wt

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0.02 0.018 0.016 0.014 0.012 0.01 0.008 0.006 0.004 0.002 0

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Silicon content, %wt

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0.04 0.035 0.03 0.025 0.02 0.015 0.01 0.005 0

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Copper content, %wt

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0.025 0.02 0.015 0.01 0.005 0

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Nickel content, %wt

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 e f

g h

0.035 0.03 0.025 0.02 0.015 0.01 0.005 0

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Chrome content, %wt

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

25 20 15 10 5 0 -5 -10

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

First principal component, PC

1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

6 5 4 3 2 1 0 -1 -2 -3 -4

P

r

o

b

a

b

i

l

i

t

y

o

f

s

i

g

n

i

f

i

c

a

n

t

s

c

a

l

e

f

o

r

m

i

n

g

Second principal component, PC2

400

500

600

700

800

900

1000

1100

x21 x20 x19 x18 x17 x16 x15 x14 x13

T

e

m

p

e

r

a

t

u

r

e

,

0

C

Temperature variables at the hot mill

Fig. 3. Probability of scale formation with variations in the signiﬁcant process variables.

1630 J. Kennedy et al. / Journal of Materials Processing Technology 212 (2012) 1622–1630

them being used in high end applications. For example, to keep

the percentage of defective coils below 10%, Fig. 3a reveals the

phosphorous content must be below0.011wt.%, when running the

hot mill at the mean values for all the other process variables (see

Table 1 for these mean values). Again, to keep the percentage of

defective coils below 10%, Fig. 3d reveals the nickel content must

be below0.013wt.%, when running the hot mill at the mean values

for all the other process variables. As a ﬁnal example, Fig. 3f shows

that to keep the percentage of defective coils below 10%, the ﬁrst

principal component must be below1.9 when running the hot mill

at the mean values for all the other process variables.

Using Eq. (9a) and PC

1

=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 calcu-

lated temperatures are shown in Fig. 3h. It can be seen that to keep

the percentage of defective coils below10%, the maximumrougher

mill temperature must be kept at 1040

◦

C or less. When looking at

the average rougher mill temperature the ﬁgure should be 1025

◦

C

or less. The corresponding ﬁgures for the maximum, average and

minimum crop shear temperatures are 1020

◦

C, 970

◦

C and 540

◦

C

respectively. The corresponding ﬁgures 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 1.

This type of calculationcanbe done for any PC

1

value, andthere-

fore for any frequency of signiﬁcant scale formation to ﬁnd what

temperatures to run the hot mill at to achieve this required fre-

quency of scale formation. The ﬁnal impression to be gleamed from

Fig. 3 is that the variables to keep under strict control, so as to

remove the possibility of very high proportions of coils being pro-

ducedwitha signiﬁcant number of defects (probabilities above 50%

for example), are phosphorous, copper and the various tempera-

tures as summarised in the ﬁrst principal component.

7. Conclusions

The principal component analysis above showed that the tem-

peratures at the hot mill were all highlycorrelated. However, nearly

75% of the variation in the ten hot mill temperatures could be

accounted for by just four principal components that were com-

pletely 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

withjust ﬁve chemistries—phosphorous, silicon, copper, nickel and

chrome. The Logit model also had a correct prediction rate of 73%.

It appears fromthis 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 of forming

scale at levels greater than a count of 200 on the bottom surface.

Therefore, reducing the average temperature entering the ﬁnishing

mill will reduce the amount of scale formed on the surface of the

coil.

Areas for future work could include using a more reliable

dependent variable, i.e. other than the counts obtained from the

hot mill Parsytec system. To this end data could be collected after

the pickling stage and be used in combination with the pickle line

processing. The effect of other variables such as different descal-

ing practices could also be investigated, together with the use of a

more general classiﬁcation of scale count into say low, moderate,

high and very high.

Acknowledgements

The authors are grateful for the support of Tata Strip UK, and

personnel within Port Talbot Hot Mill. The authors are also grateful

for thefundingreceivedfromtheEngineeringandPhysical Sciences

and Research Council (EPSRC) and Tata Steel Strip Products UK.

References

Bolt, H., 2000. The Properties of Oxide Scales on Hot Rolled Steels: A Literature

Review. Corus Research, Development &Technology, Ijmuiden Technology Cen-

tre.

Bolt, H., 2003. Understanding the properties of oxide scale on hot rolled strip. In:

METEC Congress 03/3rd European Rolling Conference (ERC), Dusseldorf, June

16–18, 2003.

Burt, C.L., 1945. Howthe Mind Works. Allen & Unwin, London.

Cattell, R.B., 1952. Factor Analysis. Harper, NewYork.

Childs, D., 1970. The Essentials of Factor Analysis. Holt, Rinehart and Winston, Lon-

don.

Haapamaki, J., Tamminen, S., Roning, J., 2005. Data mining methods in hot steel

rolling for scale defect prediction. In: Proceedings of AIA, February, pp. 453–464.

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, NewJersey.

Maddala, G.S., 1991. Introduction to Econometrics, 3rd edition. John Wiley & Sons

Ltd., NewYork, p. 322.

Mardia, K.V., Kent, J.T., Bibby, M., 1979. Multivariate Analysis. Academic Press,

London.

Munther, P., Lenard, J., 1999. The effect of scaling on interfacial friction in hot rolling

of steels. Journal of Materials Processing Technology 88, 105–113.

Prentice, R.L., 1975. Discrimination among some parametric models. Biometrika 62,

607.

Silk, N., 2001. The practical aspects of hydraulic de-scaling. Steel Times International

October, 38–44.

Sun, W., Tieu, A.K., Jiang, Z., Zhu, H., Lu, C., 2004. Oxide scales growth of low-carbon

steel at hightemperatures. Journal of Materials Processing Technology 155–156,

1300–1306.

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

scale/substrate interface area of Si-containing low-carbon steels at high tem-

peratures. Materials Science and Engineering A 308, 250–257.

Yang, Y., Yang, C., Lin, S., Chen, C., Tsai, W., 2008. Effects of Si and its content on the

scale formation on hot-rolled steel strips. Materials Chemistry and Physics 112,

566–571.

- Kelompok Mata Kuliah Teknik Pengelasa
- Mechanical Metallurgy (Dieter 1988).pdf
- [A._K._Rakhit]_Heat_Treatment_of_Gears_A_Practica(BookFi.org).pdf
- 0521819423_Damage.pdf
- Penerapan-Lubang-Resapan-Biopori-Sebagai-Alternatif-Untuk-Meminimalisir-Banjir-Di-Kawasan-Perumahan-Ciledug-Indah-I-RIZKI-MUZAMMIL.pdf
- lecture 6
- Bab 6 Momentum Sudut Dan Rotasi Benda Tegar
- Proses bulk deformasi
- Mekanika Teknik
- Calcium Carbonate Scale Formation—Assessing the Initial
- 10Nergaard
- 09 Hardness Test-1
- 316 316l Data Sheet
- Prinsip e28093 Prinsip Dasar Pengendalian Korosi
- 628-1379-1-SM
- Effect of Magnetic Water Treatment on Calcium Carbonate Precipitation
- Analysis of CaSO4 Scale Formation Mechanism in Various
- Mechanism of Calcium Carbonate Scale Deposition Under
- pembentukan kerak
- Mechanism of Calcium Carbonate Scale Deposition Under
- Chapter 6 Entropy Ppt
- induktor
- ipi116935
- 140211131924_tews_244_chew_the_fat

Sign up to vote on this title

UsefulNot useful0.0 (0)

- Formula Mech Engg.by Navin Shukla
- revitranby Sebastian De Ossa Restrepo
- referat1by Leslie Nelson
- IJERD (www.ijerd.com) International Journal of Engineering Research and Development IJERD : hard copy of journal, Call for Papers 2012, publishing of journal, journal of science and technology, research paper publishing, where to publish research paper, journal publishing, how to publish research paper, Call For research paper, international journal, publishing a paper, how to get a research paper published, publishing a paper, publishing of research paper, reserach and review articles, IJERD Journal, How to publish your research paper, publish research paper, open access engineering journal, Engineering journal, Mathemetics journal, Physics journal, Chemistry journal, Computer Engineering, how to submit your paper, peer review journal, indexed journal, reserach and review articles, engineering journal, www.ijerd.com, research journals, yahoo journals, bing journals, International Journal of Engineering Research and Development, google journals, journal of engineering, online Submisby IJERD

- Analysis of the Effect of Alloying Elements on the Martensite-Start Temperature Ofby Guilherme Gomes
- TS_Refrigerated-Heated Bath Circulatorsby Gabriel Glotz
- Effects of Cryogenic Cooling by Liquid Nitrogen Jet on Forces, Temperature and Surface Residual Stresses in Grinding Steelsby Chirag Jain
- CS2by Sarbari Ganguly

- face recongination
- Anderson
- Thines Lab 2
- 16421283
- Forni
- Adjusting for Population Stratification in a Fine Scale With
- 17 2012 Magn Anomaly Detection Boris Ginzburg
- Structural Health Monitoring 2016 Comanducci 505 24
- 1
- MultiModal Identification System in Monozygotic Twins
- 03-559X-003
- dvd009-masterthesisreport
- Austenite Double Decomposition in High C Steels
- cover.pdf
- Effect of Mo in TRIP Steels (PR-311-075)
- UAS Prak. Statistik Rizki Auliya
- MET_45_2_093_096_Dorcak
- IOSRJEN(www.iosrjen.org)IOSR Journal of Engineering
- K10547_a001
- [IJCT-V2I2P15] Authors
- Formula Mech Engg.
- revitran
- referat1
- IJERD (www.ijerd.com) International Journal of Engineering Research and Development IJERD
- Analysis of the Effect of Alloying Elements on the Martensite-Start Temperature Of
- TS_Refrigerated-Heated Bath Circulators
- Effects of Cryogenic Cooling by Liquid Nitrogen Jet on Forces, Temperature and Surface Residual Stresses in Grinding Steels
- CS2
- Multistage cryo…application
- Responses to NIST's FAQ
- Identification, For Control, Of the Process Parameters Influencing Tertiary Scale