Comparative analysis of the properties of the nodular cast iron with carbides and the austempered ductile iron with use of the machine learning and the support vector machine

ORIGINAL ARTICLE

Comparative analysis of the properties of the nodular cast

iron with carbides and the austempered ductile iron

with use of the machine learning and the support vector machine

Dorota Wilk-Kolodziejczyk

1,2

& Krzysztof Regulski

& Grzegorz Gumienny

Received: 16 September 2015 /Accepted: 12 February 2016 /Published online: 9 March 2016

The Author(s) 2016. This article is published with open access at Springerlink.com

Abstract The use of modern casting materials allows the

achievement of higher product quality indices. The conducted

experimental studies of new materials allow obtaining alloys

with high performance properties while maintaining low pro-

duction costs. Studies have shown that in certain areas of

applications, the expensive to manufacture austempered duc-

tile iron (ADI) can be replaced with ausferritic ductile iron or

bainitic nodular cast iron with carbides, obtained without the

heat treatment of castings. The dissemination of experimental

results is possible through the use of information technologies

and building applications that automatically compare the

properties of materials, as the machine learning tools in com-

parative analysis of the properties of materials, in particular

ADI and nodular cast iron with carbides.

Keywords Austemperedductileiron(ADI)

Nodularcastiron

with carbides (NCIC)

Cast iron

Data mining

Machine

learning

Support vector machine

1 Introduction

Ductile iron continues being one of the most dynamically

developing cast materials in the world. It is used by engineers

as a preferred material for parts of the machinery and equip-

ment. This forces the need to improve its functional properties

through the development of new types of ductile iron, satisfy-

ing the high demands of customers.

Studies to improve the properties of ductile iron have led to

the development of materials such as austempered ductile iron

(ADI) or ausferritic nodular cast iron with carbides (NCIC).

The microstructure of these materials ensures high properties

both mechanical and functional. An advanced variation of

ADI is carbidic austempered ductile iron (CADI), which of-

fers an increased resistance to wear. The required properties of

this material are obtained, like in the case of ADI, by heat

treatment of castings [

1–5].

ADI is obtained by a heat treatment, which consists in

quenching and isothermal holding in a temperature range of

250–400 °C. A unique combination of these cast iron proper-

ties, including the strength, ductility, and fatigue behavior,

makes this material a successful substitute for steel compo-

nents or even aluminum alloys. ADI has a high fatigu e

strength higher than aluminum and is resistant to abrasive

and adhesive wear [

5, 6], but above all, its use can significant-

ly reduce production costs by approx. 20 % compared with

wrought steel and 50 % compared with aluminum. Another

important feature is the cost of heat treatment lower than, for

example, the cost of carburizing. Besides, the mere process of

making iron castings is cheaper than the cost of making steel

castings (lower energy consumption), not to mention even the

forged steel, in the case of which the cost of production is

additionally increased by the operation of plastic forming.

ADI without heat treatment has better machinability, which

increases tool life. On the other hand, finished items have the

mechanical properties similar to numerous steel grades, while

being approx. 10 % lighter [

6, 7].

The technologically demanding process to obtain ADI is an

example of how important is accurate information about the

individual process parameters. Parameters such as the temper-

ature and time in different phases of the process, and also the

* Dorota Wilk-Kolodziejczyk

dwilk@agh.edu.pl

AGH University of Science and Technology, Krakow, Poland

Foundry Research Institute, Krakow, Poland

Lodz University of Technology, Lodz, Poland

Int J Adv Manuf Technol (2016) 87:1077–1093

DOI 10.1007/s00170-016-8510-y

chemical composition, strongly influence the final properties

of the material.

An alternative way of obtaining the bainitic or ausferritic

microstructure is by introducing elements that affect the aus-

tenite stability. In this way, an ausferritic or bainitic NCIC is

obtained. The change in the microstructure caused by the use

of alloying elements allows eliminating the costly heat treat-

ment. Isothermal transformation is necessary in the

manufacturing process of ADI significantly increasing the to-

tal cost of production. The results reviewed in this article show

beneficial effect of alloying elements on the microstructure of

ductile iron, which means the possibility of obtaining either

bainitic or ausferritic matrix with carbides increasing the wear

resistance [

8–14].

The studies of wear resistance indicate that the developed

ductile iron with carbides can successfully be used for ma-

chine parts, which are required to offer high resistance to wear,

replacing in this way, e.g., castings made of ADI [

5, 6].

Wear resistance as an important factor in the choice of con-

struction material is but only one in the vast range of other

material properties. In product design, the technologist must take

into account a number of other criteria comprised in the technical

requirements. Some materials, although significantly different

from each other in terms of, e.g., density, may have similar

characteristics as regards their mechanical properties and as such

can be interchangeably taken into account in the design of prod-

ucts. Each product must meet some specific requirements—

sometimes the critical factor will be appropriate damping capac-

ity, and in other cases, it may be the wear resistance.

Promoting new materials requires adequate dissemination of

information, best achieved with the use of modern computer

technology. Providing precise information can support the more

frequent choice of particular material. A tool for the efficient and

automatic classification of materials will serve as an aid

supporting decisions regarding the selection of construction ma-

terials. In this way, the technologist will be able to make the best

choice from among the wide range of various, but often similar ,

materials that meet the most demanding technological criteria,

such as the tensile strength (R

), the force needed to break the

material sample, and yield strength (R

p0.2

), the stress a material

can withstand without permanent deformation, elongation (A),

and hardness (HB). At this point, the problem is how to deter -

mine which materials can be classified as “si milar.” This is a

simple task in the case of two or three parameters, but with an

increasing number of attributes and the number of materials

taken into account, it becomes a multicriteria analysis. W ithout

calculations, the issue becomes quite difficult to comprehend,

even for the technologist, let alone for stakeholders such as cus-

tomers or persons preparing the offer .

Here come to the rescue the data mining tools and machine

learning techniques drawing inspiration from the studies of

artificial intelligence. The paper will present the use of a num-

ber of algorithms for classification, taking into account their

effectiveness in solving the stated problem characterized by a

strong non-linearity.

In the former authors’ research, the problems of

multicriteria decision support based on discrete and continu-

ous signals were investigated with a possibility of use of the

artificial intelligence methods in the domain of foundry engi-

neering and metal processing. Those solutions were widely

discussed in [

15, 16] where an application of decision trees

was presented in the conditions of metal processing and pro-

cess optimization and in [17–19] where casting defects were

indentified and classified with data mining techniques. Also a

problem of knowledge representation was studied within

those works in the context of data acquisition and integration

in order to further process information [

20–22]. The reasoning

models based on one of logic knowledge representations—

logic of plausible reasoning—were widely described in [23,

24]. An artificial intelligence was successfully applied in the

systems of pattern recognition and diagnostics—an important

domain of industry processes [

25–27].

The problem in the analysis of experimental data is often

the acquisition of the data needed for research. An invaluable

source of data that until recently was not taken into account in

research activities is the Internet. Searching for information to

select process parameters and support the decision-making

procedure is becoming nowadays more and more difficult

due to the dispersion of research centers, a variety of sources

(researchers , technologists, magazines, books, research re-

ports, results of experiments), even in a niche so small as the

production of ADI. Increasingly, all these sources of informa-

tion can be found on the World Wide Web, but to accomplish

this, it is often not enough to use Google.com, and even if it is,

the task may turn out to be quite difficult [

28, 29].

The proposed analysis is done by two routes, and one of

them is to examine the problem of classification based on the

raw data concerning various properties. This method can be

improved by a preliminary analysis of clusters based on prop-

erties, making classification more rough but at the same time

flawless (Fig. 1).

2 The results of experiments with different materials

2.1 NCIC

The base cast iron was melted in an electric medium frequency

induction furnace. The charge was composed of pig iron of a

special c hemical c omposition presented in Tabl e

1 and

S235JR steel scrap with chemical composition according to

the PN-EN 10025-1:2007 standard.

The content of silicon and manganese was adjusted with

theadditionofFeSi75andFeMn75,respectively.Thechem-

ical composition of cast iron was examined with a

SPECTROMAXx emission spectrometer provided with spark

1078 Int J Adv Manuf T echnol (2016) 87:1077–1093

excitation from SPECTRO Analytical Instruments GmbH.

The spheroidization of cast iron was made by Inmold process,

using Lamet® 5504 master alloy from the Norwegian Elkem

Company. The chemical composition of the master alloy is

shown in Table

Due to its strong inoculating effect, the master alloy used

for the Inmold spheroidizing treatment has provided a high

degree of the refinement of eutectic grains in the tested ductile

iron. This produced a large number of fine eutectic grains and

carbides. This is important for the castings whose functional

properties are to be achieved in as-cast condition, i.e., without

the use of heat treatment.

In the study, the data was collected from 150 melts. As

alloying additions, elements such as copper, nickel, molybde-

num, and chromium were used in concentrations strictly de-

pendent on the effect they exert on the microstructure and

properties of cast iron. The additions of Mo, Cr, Ni, and Cu

enable obtaining the ductile iron whose metallic matrix is

based on bainite, ausferrite, martensite, or a mixture of these

constituents, additionally containing carbides to increase the

wear resistance without subjecting the casting to a heat treat-

ment. The effect of various alloying elements on the micro-

structure and properties of cast iron was investigated. (1) Due

to its pearlitizing effect, the addition of copper greatly reduces

the temperature of the beginning of austenite transformation.

The addition of copper resulted in the disappearance of ferrite

and in the formation of a pearlitic microstructure of the metal

matrix in the entire examined range of the casting wall thick-

ness values. (2) The addition of nickel raised the temperature

of the eutectic transformation by approx. 6 °C (slightly smaller

increase than that observed in the cast iron containing Cu).

This is due to the graphitizing effect of an intensity similar to

Cu. Compared with unalloyed cast iron, nickel reduces the

tendency to the matrix ferritization. However, in the case un-

der discussion, this effect was weaker compared to the cast

iron containing copper, because in castings with the examined

range of wall thickness values, the graphite nodule s were

surrounded by the precipitates of ferrite. (3) Compared to un-

alloyed cast iron, the addition of 1.5 % molybdenum resulted

in a significant drop of the crystallization temperature of the

eutectic (by approx. 16 °C). (4) The addition of Cr reduced the

temperature of the eutectic crystallization by about 10 °C,

compared to the unalloyed cast iron, but the drop was smaller

than in the cast iron c ontaining similar amount of Mo.

Chromium intensively raises the temperature of the pearlitic

transformation in nodular cast iron. The addition of about

1.5 % Cr favors matrix pearlitization and promotes carbide

formation in the microstructure of the cast iron metal matrix.

Studies have shown that to produce upper bainite in the

matrix of nodular cast iron with carbides, it was necessary to

introduce a combined addition of molybdenum and nickel.

Molybdenum concentration w as changing in the range of

1.3 to 2.1 % with nickel content comprised in the range of

0.7–1.1 % (depending on the casting wall thickness and con-

centration of other elements). Depending on the concentration

of Mo, Ni, and Cr, the surface content of carbides in the cast

Fig. 1 Diagram of the process of

modelling the problem of the

selection of material based on

required properties

Tabl e 1 Chemical

composition of special

pig iron

Chemical composition, wt%

CSiMnP S

4.44 0.69 0.025 0.046 0.009

Tabl e 2 Chemical composition of the spheroidizing-inoculating master

alloy [

Chemical composition, wt%

Si Mg Ca La Al Fe

44–48 5–60.4–0.6 0.25–0.40 0.8–1.2 Rest

Int J Adv Manuf Technol (2016) 87:1077–1093 1079

iron was from 0 to about 12 %; the content higher than 5 %

resulted in a significant drop of ductility. The microstructure

of upper bainite without the inclusions of other phases was

obtained in the nodular cast iron with carbides for the follow-

ing range of the chemical composition (wt %): C = 3.27–

3.91 %, Si = 2.23–2.56 %, Mn = 0.05–0.09 %, Cr = 0.00–

0.28 %, Mo = 1.87–2.13 %, and Ni = 0.66–1.05 %. Due to

the increased content of carbides in thin-walled castings, Mo

concentration could be reduced to about 1.5 %. This reduced

the surface content of carbides to about 5 % and increased the

ductility.

To obtain in the cast iron metal matrix the structure of upper

bainite without carbide precipitates, molybdenum concentra-

tion was decreased to about 1.3 %. The maximum tensile

strength of the nodular cast iron with carbides and the upper

bainite was about 1000 MPa at an elongation of about 3–4%.

At a 6 % elongation, the value of R

was 750 MPa with the

yield strength of 500 MPa.

The cast iron with a metal matrix microstructure composed

of upper bainite, lower ba init e, and carbi de s had slightly

higher mechanical properties. This was due to the nickel con-

centration raised to about 1.5 % with molybdenum concentra-

tion similar as in the cast iron with upper bainite. In thin-

walled castings, the concentration of nickel should be lower,

considering the high cooling rate and the possibility of mar-

tensite formation. The range of the chemical composition for

the nodular cast iron with carbides and a mixture of upper and

lower bainite was as follows (wt %): C = 3.57–3.83 %,

Si = 2.36–2.87 %, Mn 0.05–0.09 %, Cr = 0.00–0.34 %,

Mo = 1.91–2.05 %, and Ni = 1.33–1.62 %. I t allowed

obtaining the cast iron with a tensile strength R

in the range

of 850–1050 MPa at an elongation A of 4–2%.

Introducing to the cast iron with molybdenum, small (about

0.5 %) quantities of chromium and copper in a concentration

of 1.3–2.0 % produced ausferrite and carbides in the metal

matrix without the need for the heat treatment of castings.

Depending on the cooling rate (casting wall thickness), the

ausferritic microstructure of nodular cast iron with carbides

was obtained for the following range of chemical composition

(wt %): C = 3.53–3.84 %, Si = 2.40–2.63 %, Mn 0.28–0.34 %,

Mo = 1.41–1.62 %, Cr = 0.45–0.55 %, and Cu = 1.03–2.08 %.

In thin-walled castings, ausferrite was obtained with copper

content in the lower limit of the specified range. In the castings

with a wall thickness of 25 mm, the minimum Cu content

necessary to obtain ausferrite free from the precipitates of

pearlite was 1.5 %. Increasing the concentration of Cu to

about 2.0 % did not cause the despheroidization of the nodular

graphite.

Among the investigated types of nodular cast iron with

carbides, the cast iron with an ausferritic microstructure of

the metal matrix had the highest strength properties. Its tensile

strength (R

) varied between 750 and 1 100 MPa at an elon-

gation (A) of 6 to 3 %, respectively. The hardness in Brinell

scale (HB) of ausferritic cast iron was from 280 to 345 HB

units and was by up to 40 HB units lower than the hardness of

the cast iron with upper bainite.

The presence of carbides in the ductile iron metal matrix is

expected to increase the resistance to abrasive and adhesive

wear compared with other, currently used, types of cast iron.

The abrasive and adhesive wear of the newly developed cast

iron grades was compared with the grades used so far, ADI

included. The effect of the surface content of carbides on the

cast iron wear resistance was examined [

11, 14]. In this group

of materials, ADI had the lowest hardness among all the ex-

amined cast iron grades. The cast iron with upper bainite ma-

trix exhibited a slightly higher hardness of 272 HB units .

Carbides in the content of 5 and 12 % increased the hardness

of this cast iron to 293 and 411 HB units, respectively. The

cast iron with the matrix composed of a mixture of upper

bainite, lower bainite, and carbides was characterized by the

highest hardness of 477 HB units.

The mechanical properties of nodular cast iron with car-

bides depend on the type of metal matrix and on the carbide

content. The conducted studies have shown that surface con-

tent of carbides higher than 5 % causes a significant decrease

in plastic properties regardless of the type of metal matrix. It

has been found that the small fraction of carbides in the ma-

trix, reducing only slightly the cast iron ductility, significantly

raises the resistance to both abrasive and adhesive wear. The

examined cast iron was divided into five groups (from I to V)

depending on the matrix microstructure. The surface content

of carbides in each examined cast iron group was up to 5 %.

Table

3 shows the basic mechanical properties of different

types of the nodular cast iron with carbides.

Table 3 Mechanical properties

of nodular cast iron with carbides

Group no. Matrix type Mechanical properties

,MPa R

p0,2

,MPa A,% HB

IPearlite 450–700 330–450 2–5270–375

II Upper bainite 750–1000 500–750 3–6285–385

III Upper bainite with lower bainite 850–1050 650–800 2–4380–490

IV Martensite 780–950 500–700 <2 470–570

VAusferrite 750–1100 500–700 3–6280–345

1080 Int J Adv Manuf T echnol (2016) 87:1077–1093

The cast iron with a pearlitic matrix and carbides (group I)

showed the lowest strength properties, while the highest ten-

sile strength was obtained in the ausferritic cast iron with

carbides (group V). The highest elongation was obtained in

the nodular cast iron with carbides in which the matrix was

composed of upper bainite and ausferrite (groups II and V,

respectively). The ductile iron free from the carbide precipi-

tates showed elongation higher by maximum 2 %. The rela-

tively large discrepancy in the minimum and maximum prop-

erties within a given group was mainly caused by the surface

content of carbides varying between 2 and 5 %.

Through the introduction of alloying elements, such as Mo,

Cr, Cu, and Ni, it is possible to obtain bainite or ausferrite in

the cast iron without any heat treatment. This is due to the

impact of these elements on the time–temperature–transfor-

mation (TTT) curves. Each of the abovementioned elements

makes the curve of the start of austenite decomposition shift

towards longer times, the influence most intensive being that

of nickel. Copper impact on the stability of austenite is less

intense. Chromium and molybdenum are carbide-forming el-

ements, but to some extent, they can dissolve in austenite, thus

affecting its stability. A very characteristic effect is that of

molybdenum, which enhances the stability of austenite with

respect to its transformation to pearlite. Nevertheless, molyb-

denum does not affect in a more significant manner the stabil-

ity in a bainitic range. Figure

2 shows the CCT curves of cast

iron containing about 0.5 % Cr, 1.4 % Mo, and 1 % Cu. The

graph was plotted using the following cooling rates of sam-

ples: 6, 8, 10, 15, 20, 30, 40, 50, 60, 75, 90, 105, 120, 150,

180, 210, 240, 270, 300, 330, 360, 402, 450, 510, 600, 900,

and 1620 °C/min. The specified chemical composition en-

abled obtaining in the base cast iron for dilatometric tests a

metal matrix composed of ausferrite and carbides in castings

with the wall thickness of 3 to 25 mm. The TTT graphs of

selected cast iron grades were plotted using a RITA L78 High

Speed Quenching Dilatometer made by LINSEIS Company.

The HV30 hardness measurements were taken on samples

used for plotting of the TTT curves with a Dia Tester 2R made

by Wolp ert-Werke (Germany) under a load of 294. 2 N

(HV30). The measurement was performed in accordance with

the PN-EN ISO 6507-1 standard.

From the graphs, it follows that in the area of pearlitic trans-

formation (upper marked region), the stability of austenite is

similar to the stability in bainitic area (lower marked region).

This is due to a reduced concentration of molybdenum and the

addition of Cu, which is less effective than nickel in enhancing

the stability of austenite in the range of pearlitic transformation.

There is, however, a range of cooling speeds, which promotes

the occurrence of bainitic transformation without the need of

crossing the pearli tic area. The impact of copper on the begin-

ning of austenite decomposition in a bainitic range is much more

intense than that of nickel. In the examined cast iron, the mar-

tensitic microstructure was obtained in the range of cooling

speeds of 360–1620 °C/min. Pearlite precipitates occurred in

the metal matrix of cast iron cooled at a rate of 210 °C/min

and lower. Ferrite envelopes around the graphite nodules were

observed to occur in the cast iron cooled at a rate equal to or

lowerthan20°C/min.

Fig. 2 The CCT curves of nodular cast iron with carbides of the following chemical composition: 3.75 % C, 2.40 % Si, 0.33 % Mn, 0.51 % Cr, 1.41 %

Mo, and 1.03 % Cu. F ferrite, P pearlite, B bainite, A austenite, M martensite, C carbides

Int J Adv Manuf Technol (2016) 87:1077–1093 1081

2.2 ADI

One of the stages in the production of ADI is making the cast

iron with nodular graphite and the addition of elements such

as Mn, Ni, Cu, Mo, Cr, Sn, or other elements allowing the

formation of a pearlitic or pearlitic-ferritic structure with im-

proved hardenability. Here some attention deserves the fact

that not every grade of ductile iron is suitable for the produc-

tion of ADI. Some grades of low-alloy and high-alloy cast

iron (e.g., high-nickel cast iron) are either not used to produce

ADI or their use is not economically viable. The spheroidizing

treatment is performed with magnesium or its respective com-

pounds. In some cases, the rare earth elements such as, for

example, Ce or Re are also used. The spheroidizing treatment

of cast iron consists in the introduction of appropriate amounts

of magnesium into the bath of cast iron, resulting in the pre-

cipitation of graphite with nodular morphology. The refine-

ment and proper shape of graphite depend on the content of

sulfur and magnesium that are present in the alloy in a residual

form as its constituents.

Differences in mechanical properties of the specific ADI

grades are associated with differences in the structure. This, in

turn, is due to different variants of the applied heat treatment.

Through control of the parameters such as the temperature of

austenitizing or austempering, the properties of ADI can be

modelled. This is best seen in the tensile strength (R

)-to-

elongation (A) relationship. In the case of austempering, a very

important parameter is the temperature of the process. If the

cast iron is austempered at a temperature above 400 °C, due to

rapid transformation, the presence of carbides can be expected

in the structure with the resulting consequence of the cast iron

having inadequate ductility. The content of austenite relative

to the content of ferrite is highest in the range of 350 to 370 °C,

and then the cast iron has the highest elongation and the lowest

strength.

Table

3 shows a number of ADI variations designated with

symbols from ADI-1 to ADI-6. They repres ent different

chemical compositions of the cast iron subjected later to the

heat treatment of austenitizing and austempering. The data for

tests was collected from the 23 variants of the chemical com-

position. The data on the ADI mechanical properties obtained

by the application of different heat treatments covered 172

samples [

30]. Table 4 shows only a few selected variants of

this material.

A general scheme of the process for making ADI is shown

in Fig.

Characteristics of the set of samples are presented in

Table

From the results of the analysis, it follows that ADI ductil-

ity increases with the increasing content of austenite in the

structure, while the y ield strength (R

p0.2

) and the tensile

strength (R

) decrease. The content of austenite in the struc-

ture increases with the increasing temperature of the

ausferritizing treatment; during this time the content of ferrite

decreases. With the increasing temperature, the ferrite coarse-

ness increases. The results show that at lower temperatures of

the ausferritizing treatment (260 °C), the content of ferrite in

the cast iron increases, thus leading to an increase in the yield

strength (R

p0.2

) and tensile strength (R

) for both of the ex-

amined cast iron chemical compositions (Fig.

4). Fracture

toughness (the stress intensity factor) reaches its peak for the

hardness value of approx. 40 HRC, that is, when the cast iron

contains approx. 60 % ferrite and 25 % austenite. This means

that the optimum values are achieved for the ausferritizing

treatment carried out at a temperature of approx. 280 °C for

a time of approx. 3.5 h. The analysis of microstructure shows

that the optimum fracture toughness can be achieved with a

microstructure consisting of lower bainite and fine strips of

ferrite and with the retained austenite in an amount of approx.

30 % saturated with carbon to a level above 1.8 %. The opti-

mum fracture toughness is obtained for an isothermal anneal-

ing in the range of 300–320 °C [

31].

2.3 Preliminary data analysis—a comparison of materials

A collective comparison of the impact of individual

alloying elements on the properties of ADI and NCIC is

presented on scatterplots in Fig. 5. Well visible is t he drop

of plastic properties caused by the majority of alloying

elements. Only nickel increases the elongation of the ex-

amined materials and their strength.

Table 4 Chemical composition of different ADI variations

C Si Mn Mg Cu Ni Mo S P V Cr Ti Sn Al

ADI-1 3.65 2.76 0.36 0.045 0.6 0.9 – 0.014 0.016 –––––

ADI-2 3.51 2.45 0.3 0.036 –––0.01 0.017 –––––

ADI-3 3.5 2.65 0.4 0.035 0.055 1.6 0.3 0.01 0.021 0.05 – 0.02 ––

ADI-4 3.5 2.47 0.38 0.035 0.35 1.57 0.3 0.01 0.021 0.05 0.5 0.02 ––

ADI-5 3.54 2.81 0.43 0.05 0.56 1.52 0.3 0.009 0.031 – 0.3 –––

ADI-6 3.4 2.69 0.19 0.044 0.87 0.73 0.23 0.01 0.02 – 0.04 0.004 0.007 0.015

1082 Int J Adv Manuf T echnol (2016) 87:1077–1093

The conclusions from the comparison of properties are

as follows: (1) The ductile iron with upper baini te shows

nearly two times higher resistance to abrasive wear than

its pearlitic counterpart; in ADI, this increase is nearly

three times higher. (2) The highest abrasive wear resis-

tance was found in the ductile iron with a mixture of

upper a nd lower bainite. I ts loss of weight amounted to

approx. 64 % of the weight loss suffered by ADI. ( 3)

Compared to all other types of bainitic nodular cast iron,

ADI shows the highest wear rate in parts m ate with hard-

ened steel. (4) The conducted studies of the wear resis-

tance have proved that the develo ped nodu lar cas t iro n

with carbides can successfully be used for machine parts

which should offer high resistance to wear, thus replacing,

e.g., castings made of ADI. (5) The NCIC has lower

strength than ADI but its elongation is comprised i n the

same range of values as ADI (Fig.

6a). (6) Compared with

other materials (cast steel, some grades of steel, grey cast

iron, and ductile iron), ADI and NCIC have a lower elon-

gation but higher strength and hardness. Clear division

into subgroups based on the hardness values is also ob-

served in the ADI itself (Fig.

7). One of the subgroups has

a hardness at the level of NCIC, and the other shows

definitely higher values. (7) It can be stated that NCIC

has properties similar to ADI, adopting the values inter-

mediate between the ADI and other construction mate-

rials, mainly as regards the strength, hardness, and elon-

gation (Fig.

6b). This is confirmed by the canonical

Fig. 3 A general scheme of the

process for making ADI

Table 5 Characteristics of the set

of ADI samples

Chemical composition

CSi MnMg Cu Ni

3.56–3.62 % 2.45–2.63 % 0.12–0.32 % 0.037–0.065 % 0.44–0.93 % 0.43–1.9 %

Heat treatment

Austenitizing Ausferritizing

Temperature Time Temperature Time

900 °C 90–120 min 300–360 °C 120–180 min

Mechanical properties

Tensile strength Yield strength Elongation Hardness

862–1602 MPa 455–1418 MPa 0.3–20 % 233–629

Int J Adv Manuf Technol (2016) 87:1077–1093 1083

analysis of the impact of mecha nical properties on the

division of materials into groups (ADI, NCIC, and others)

and is the basis for d rawing so me conclusio ns, where

component 1 comprises mainly the strength and compo-

nent 2 the elongation and hardness.

3 Methods of data mining–machine learning

Machine learning methods, derived from research on

artificial intelligence, allow generating knowledge in

the form of models or rules from a set of training data

for the fu tur e us e o f t his knowledge in th e a nal y si s of

new facilities. Artificial intelligence methods have al-

ready been used in various problems of the engineering

and industrial nature, as evidenced by the rich literature

in this area. The collection of these machine learning

methods includes the most popular tools for classifica-

tion, such as minimum distance classifiers (k-nearest

neighbors (kNN)), induction of decision trees, artificial

neural networks, and also the most modern tools, like

support vector machine (SVM).

Fig. 4 Example of relationship between the tensile strength and temperature and time of the isothermal annealing

Fig. 5 Analysis of the impact of

various alloying elements on the

properties of ADI and NCIC

1084 Int J Adv Manuf T echnol (2016) 87:1077–1093

3.1 Machine learning using kNN, CART, CHAID,

and ANN classifiers

In all these algorithms, learning proceeds according to a cer-

tain pattern: on the basis of training data (approx. 70 % of

cases), a classification model is constructed. This model can

take a variety of forms: a set of rules, the distance matrix,

weights on connections, or vectors and matrices. The model

is tested during learning process with approx. 15 % of cases,

and its final form is validated (subject to verification) using the

last discriminated group of data—approx. 15 %. The evalua-

tion of the model can be performed with the use of a few

parameters. Three of them, i.e., the quality of learning, testing,

and validation, indicate the percentage of cases that are cor-

rectly classified during each phase of the model building. The

final performance of the model can be tested for the compat-

ibility of results with the dependent variable input data. In this

study, parameters such as the value of test χ

(chi-squared) or

(G test) could be po inted out as compliance tests (the

goodness-of-fit). Those tests are statistical methods of valida-

tion of goodness of fit to a distribution and for independence

in contingency tables and also used to compare the goodness

of fit of two models. Another criterion could be the percentage

incompatibility that determines model misclassification for

the same set of input data and the dependent variable.

The following chapters describe research involving the

construction of 18 models using various methods of machine

learning and comparing the results to choose the best classifi-

cation tool for the disclosed problem of casting material selec-

tion according to preset mechanical properties.

For this purpose, the kNN classifier was used. This method

uses a matrix of distances between each of the elements in the

training set. Classification is based on a comparison of select-

ed (nearest) neighbors of the new object. Neighborhood is

determined based on the distance, which is the Euclidean dis-

tance, or other metrics. K parameter determines the amount of

Fig. 6 a Tensile strength and elongation compared for NCIC and ADI. b Canonical analysis of the impact of R

, R

p0.2

,HB,andA

on the division into

three groups of materials

Fig. 7 Properties in groups of materials

Int J Adv Manuf Technol (2016) 87:1077–1093 1085

neighbors. It bases on a set of training cases and does not

perform the training process, but when the classification of a

new case is called, it looks in the matrix for distances k of the

nearest neighbors. Most of the values of the dependent vari-

ables among the k-neighbors are the values of the classifier

decisions. In other words, the largest class among the k-neigh-

bors becomes the classifier response. This method has been

known for years and its greatest advantage is the simplicity,

intuitiveness, and no need to build a “fixed” model—the al-

gorithm can be based on dynamically changing data, and thus

the process of “training” is most flexible.

The second applied algorithm is called artificial neural net-

works (ANN). It is a very popular method in the current re-

search on machine learning, successfully used in the control of

devices and optimization of processes in the industry. These

models allow with the aid of training data the modification of

weights on the connections in a network of neurons, which are

the computing cells. Here, the training process consists in

optimizing the matrix weights, which results in the best rep-

resentation of the training signal sequence. This method is

efficient and effective for most problems but has one major

drawback—the resulting model cannot be easily understood

by a human. The model does not generate rules or equations

that the user could interpret or verify based on his own domain

knowledge [

32, 33].

Induction algorithms of classification and regression trees

(CART) and chi-square automatic interaction detection

(CHAID) were used, where the idea was to proceed with the

division of a data set into subsets until each subset contains

objects belonging to only one class of the dependent variable.

The division was made basing on various tests—Gini index

for the CART algorithm or χ

test for the CHAID algorithm.

These methods allow the construction of decision rules based

on the explanatory (independent) variables. Decision trees are

a graphical representation of the rules disclosed in a user-

friendly form. These algorithms make it possible not only to

create rules but also to quantify the importance of each vari-

able in the model, which is sometimes as important as the

model itself. The variable is defined as important in the clas-

sification process, i.e., providing information on classes, when

it shows appropriate degree of readiness to participate in the

division of the dependent variable that is measured during the

construction of the tree. The determined validity enables cre-

ating a ranking of independent variables in terms of their im-

pact on the dependent variable. The validity is the degree of

covariance with the dependent variable. Some obvious advan-

tages of classifiers based on trees are as follows: (1) their

graphical representation, which is clear, easily interpreted,

and verifiable on the basis of domain knowledge; (2) the abil-

ity to determine the validity of the predictors; (3) insensitivity

to noise and outliers; (4) the possibility of action based on

variables both quantitative and qualitative; and (5) the out-

come in the form of a set of rules that can be used in other

applications. These methods are, however, not as effective as

neural networks and support vector machine. They do not

achieve so good results mainly due to the discretization of

quantitative variables, which demand forced generalization

[

34, 35].

The last in a group of the applied techniques is the already

mentioned SVM. It is a method that determines to which of

the two classes should belong a given set of input data. The

task is accomplished through the designation of the widest

possible boundary between classes basing on training data.

This is a linear, binary, and non-probabilistic classifier.

Linear classifier means that the boundary of classification is

a linear function. SVM is non-probabilistic because the clas-

sification is strict, determined by the vectors, not based on the

statistical analysis. Binary means that the possible result of

classification is 0 or 1. Of course, this does not mean that it

is not suitable for predicting the dependent variables, which

can assume more than two values, or for solving the non-

linear problems. In situations when there are more than two

classes, the multiclass support vector machine may use two

solutions: (1) one-versus-all, when it is building a cascade

structure of binary classifiers, and (2) one-versus-one when

it is forming a matrix of binary classifiers for all possible pairs

of the values of the dependent variable. The result of classifi-

cation in one-versus-one solution is the label of this class,

which has been selected most frequently by the binary classi-

fiers. Support vector method, although relatively young, has

already found a number of practical applications, like [

36–41].

In this article, special attention was paid to the algorithm

which gave the best results, and it was the aforementioned

support vector machine. It also deserves attention for the mere

fact that being the “youngest” of all the presented techniques,

it has already been included into a still relatively little known

branch of machine learning.

3.2 Support vector machine algorithm

As has already been mentioned, the SVM classifier is a linear

classifier, which does not mean that it is not capable of “han-

dling” the problems of a non-linear nature. In this case, linear

means that in every point of classification boundary, the clas-

sification is based on linear fun ction (vectors), but as the

whole problem, the boundary does not have to be a straight

line. Although conventional tools, such as a linear discrimi-

nant analysis (LDA), are seeking linear discriminant functions

in a space of the characteristics of the training variables, it is

the SVM model which is the normal vector and a linear com-

bination of support vectors, i.e., the nearest ones to the divi-

sion border.

A solution to the problem of non-linearity (when the vec-

tors in the training set are not linearly separable) is the so-

called kernel trick-mapping the training vectors to a space of

1086 Int J Adv Manuf T echnol (2016) 87:1077–1093

larger dimension, where their linear separability can be ex-

pected. The calculations are carried out using kernel functions.

The method consists in finding a border (hyperplane) that

will separate the classes in a possibly acceptable manner. The

system is additionally enriched with the introduced slack var-

iable (ξ

), which is a penalty for points located on the wrong

side of the border. The model is also defined by a parameter C

deciding about the size of the margin [

32].

The hyperplane can be represented as w ⋅ x − b = 0, where

w is the vector normal to the hyperplane,

parameter deter-

mines the offset of the hyperplane from the center to the co-

ordinate system, and ‖ ⋅ ‖ is the Euclidean norm in the space

ℜ

. The search for SVM model is reduced to the optimization

problem, the essence of which consists in minimizing the

Lagrange functional, ‖w‖, b,andξ

, and maximizing the mul-

tipliers α

and μ

L≡

wkk

þ C

i¼1

−

i¼1

⋅w þ bðÞ−1 þ ξ

½−

i¼1

ð1Þ

4 The results of machine learning

The machine learning process has been divided into two scenar-

ios. (1) In the first scenario, raw material data should help to

build a classifier most competen t in identifying the material of

the required mechanical properties. (2) The second variant of the

analysis includes a pretreatment of data done with the method of

clustering to form groups (clusters) of materials with similar

properties, building next a classifier that will be capable of iden-

tifying clusters of materials with the selected features.

4.1 Comparison of classification algorithms

In the case under discussion, the input vectors, based on which

the classification of the material is done, are composed of four

dimensions: tensile strength (R

), the force needed to break the

material sample; yield strength (R

p0.2

), the stress a material can

withstand without permanent deformation; hardness in Brinell

scale (HB), and elongation (A

). Based on data that includes 264

samples of materials with different properties—26 4 vectors cor-

responding to the samples are available—most data relate to test

materials. Based on material standards, the grades (symbols) of

those materials were defined for the specified chemical compo-

sitions. Some of the samples may represent the same material,

although their properties will vary within a wide range of values.

These differences result from the different scenarios of the heat

treatment or inoculation and are the largest obstacle to a correct

and unambiguous classification. There are many vectors which

have very similar properties, although they have different class

labels; in other words, different materials can have very similar

properties. In contrast, there are materials called by the same

name, e.g., ADI-23, but after different variants of the treatment

acquire different properties.

The following algorithms were used: ANN, kNN, the trees

(CART:Gini, CART:G-square, CHAID) (presented in

Section

3.1), and SVM. Some of them were used with various

configurations of the parameters. Altogether, 18 models were

made, of which SVM proved to be the best fit. The results are

shown in Table

6. Neural networks (ANN) were considered in

several versions. The best structures are given in the table. The

best has proved to be a multi-layer perceptron (MLP) with

activation function: linear and tanh. Radial networks (RBF)

did not yield so good results.

The kNN algorithm used v-cross-validation to determine the

best value of k. K value determines a number of neighbors that

are taken under consideration during classification process.

Neighbors are the nearest object in the training set. Each neigh-

bor is classified—it takes a label of known classes. The result of

the classification takes a class that dominates the neighborhood.

Classification trees were built basing on CART algorithm

using in the first case the Gini index as a criterion for division,

and G

(G-squared test) in the second case. The CHAID tree has

retained the default parameters. Induction of the classification

trees (as it was mentioned earlier) bases on the division of the

training data. Based on the sequence of test on the input vari-

ables (R

, R

p0.2

,HB,A

), the algorithm seeks to create

Table 6 Comparison of classification results using ANN, kNN, CART, CHAID, and SVM

Symbols ANN kNN CART CHAID SVM

Fitness MLP MLP k-4 Gini G2 Chi2 Line RBF RBF RBF RBF RBF RBF

4–4–40 4–8–40 C =7 C =5 C =8 C =9 C =10 C =12 C =15

γ =3 γ =5 γ =0.25 γ =3 γ =5 γ =2

17.8 9.2 22.3 0.5 3.7 0.1 0.3 8.2 9.9 0.1 8.1 9.2 7.1

43.8 35.3 32.5 6.5 17.0 2.1 4.3 24.0 29.4 2.1 23.9 26.8 21.1

Incompatibility [%] 13.8 16.6 48.0 44.2 50.5 73.7 56.0 46.0 40.0 56.0 38.0 42.0 34.0

Learning [%] 86.2 83.4 – 34.5 31.6 46.4 56.1 76.7 87.7 53.5 86.4 90.9 85.1

Testing [%] 63.3 66.7 –––– 44 54 60 44.0 62 58 66.0

Validation [%] 60.0 63.3 41.2 50.7 52.1 52.1 53.1 71.2 80.9 51.0 80.4 82.9 80.4

Int J Adv Manuf Technol (2016) 87:1077–1093 1087

homogeneous classes of dependent variable. Inpurity—measure

specifying the number of cases outside the class —can be mea-

sured with various metrics: Gini index, G

,orχ

test. None of

these algorithms have yielded sufficiently interesting results.

Artificial neural networks are the structures of simple mathe-

matical elements designed for the signal processing. ANN—au-

tomatic artificial neural network—is the algorithm of developing

neural networks with various parameters as number of neurons,

number of layers, and activation functions. The algorithm creates

number of structures which are learned, tested, and validated and

in turn returns the best architectures of networks—architectures

that return the result the most similar to that expected.

The first and second rows of the Table

6 determine the

parameters of the models. Subsequent rows of the table in-

clude various parameters for an assessment of the models

described in Section

3.1. It may be noted that the algorithms

of kNN and of the decision trees induction (CART and

CHAID) definitely fail to yield good results in this particular

case—their quality in validation group is only 50 %.

The best results were achieved with the neural networks

and SVM. Neural networks show particularly good perfor-

mance in fitting the test results to the dependent variable in

the input data, which means that the algorithm has a high

capability of the reference sequence replication. Poor results

of validation, i.e., verification based on the data which do not

participate in the network training process, point to the net-

work tendency to “overtraining” and a weakness in the gen-

eralization of models.

In this aspect, the best performance offered the SVM algo-

rithm. In spite of a high percent non-compliance rate for most

of the cases, in validation test, it a chiev ed the maximum

values of up to 80 %. This means the greatest ability to gen-

eralize patterns and gives the best chance for correct operation

in the case of new samples not included in the training data.

However, none of the algorithms have proved to be fault-

less, and as regards some of the assessments, it could even be

said that the results were unsatisfactory, to mention as an ex-

ample the quality of testing. This means that the classification

gives only approximate results, and thus the decision which

Table 7 The table of

cardinality values for

clusters obtained with

EM algorithm

Cluster No Percent

17227.27

3 108 40.91

24617.42

43814.39

Total 264 100

Fig. 8 Assignment of materials to clusters based on the properties such as R

, R

p0.2

,HB,andA

using a cluster analysis algorithm EM

1088 Int J Adv Manuf T echnol (2016) 87:1077–1093

material has the properties specified by the user can not be

fully automatic. Consequently, the choice of construction ma-

terial still requires some amount of domain knowledge. This

formulation of preliminary conclusions allowed introducing

some modifications to the currently adopted problem

embodiment.

4.2 Supporting classification analysis: clustering

As stated in the previous section, the classification according

to material symbols is not fully effective and cannot be the

sole criterion for the choice of material. In fact, industrial users

may be satisfied with the support at the level of material group

selection—this meaning an answer to the question whether to

choose cast iron, or cast steel, or steel, etc. Basing on these

assumptions, an additional, intermediate process of the data

analysis has been proposed. It is called the cluster analysis.

The raw data in the form such as described in Section

4.1

was subjected to a pretreatment in the form of cluster analysis

done by the expectation maximization (EM) method. The al-

gorithm described in detail in [33] allows grouping of objects

into clusters based on the values of the characteristics

describing these objects (variables) using two-criteria optimi-

zation: minimizing the distance within the cluster and maxi-

mizing the inter-cluster distances (Fig.

8). It is a probabilistic

algorithm—instead of assigning definitely an example to a

group, it estimates the probability of such affiliations. The

calculation of the affiliation takes into account the distribution

characteristics. The method comprises two steps carried out

alternately until between successive runs there is no noticeable

improvement: (1) estimation (expectation). For the currently

estimated distribution parameters of the examples, it allocates

to the examples the probability of belonging to a group. (2)

Maximization: Changing the current distribution parameters

for such that will lead to a model more in line with the data

(the distribution of examples). For this purpose, it uses the

probability of membership in various groups determined in

step 1.

Analyzing the resulting clustering, it can be concluded that

clusters 1 and 3 are mainly materials from the family of ADI.

Cluster 2 is mainly NCIC and some individual cases of ADI,

as well as certain types of nodular cast iron with high R

Cluster 4 groups include other materials—nodular cast iron

with lower R

, ferritic ductile iron, and steel. These results

Fig. 9 The distributions of variables in individual clusters

Int J Adv Manuf Technol (2016) 87:1077–1093 1089

confirm previous analyses described in Section 2.3. The car-

dinality values for clusters are presented in Table 7.

Individual clusters have different distributions of the values

of the input variables, which in this case are the mechanical

properties. It can be seen (Fig.

9) that the cluster grouping

NCIC and the most similar ADI, that is cluster 2, have lower

strength than cluster 1 (ADI), lower ductility, and hardness but

higher elongation. At the same time, in many areas, it coin-

cides with a group of materials from the ADI family (cluster

3). What is important is the fact that nearly no area of the

properties coincides with cluster 4, that is, with other mate-

rials. So, it can be concluded that some of the materials from

the NCIC family are very similar to the ADI subset and as

such are substitutable (Fig.

10).

4.3 Classification using cluster analysis results

The results of cluster analysis allow re-considering the prob-

lem of classification. Each material has been assigned to one

of four clusters created according to the division by mechan-

ical properties. As a result, the problem of classification will

consist in indicating a group of similar materials and not an

individual m aterial as has been done previously (see

Section 4.1.). In this way, an error is avoided which results

from the overlapping properties of materials and a bias asso-

ciated with variations in the properties of different samples of

the same class of materials. As the analysis using a CART

algorithm shows, for thus processed data, the results are very

promising (Fig.

11a).

Even the CART algorithm, which in previous studies did

not give good results, now allows for very efficient classifica-

tion and drawing of further conclusions. An important role in

this classification will play the variables such as the tensile

strength (R

) and the yield strength (R

p0.2

), but rough division

can be done using hardness (HB) only (Fig. 11b).

As has been indicated previously, the induction of classifi-

cation trees allows discovering the relationships which enable

the division of partitions (SVM allows insight into the support

Fig. 10 Two-dimensional comparison of the values of features for clustering

1090 Int J Adv Manuf T echnol (2016) 87:1077–1093

vectors that are not easily understood by the user).

Visualizations shown in Fig. 10a allow for the model evalua-

tion by a domain expert. The successive bars in each of the

tree nodes reflect the percent content of objects from each

cluster in a partition represented by the node. If the leaf bears

only one bar, it means a homogeneous class and inerrancy of

the rule (represented by the entire branch). It can therefore be

read from a tree that, e.g., the low hardness and low yield

strength of the material mean the material from group 4, that

is, cast steel, ductile iron, and some grades of steel, and this

can be said with the confidence equal to 93 %. If hardness

exceeds 363 HB and the yield strength is higher than

797 MPa, it means cluster 1, that is, ADI (98 % confidence).

The most significant errors of the trees are also visible. The

classification matrix indicates where there is a risk of making a

mistake—cluster 2 is the one most commonly confused with

materials from clusters 1 and 3, which means that there are

ADI types with very similar properties to the NCIC cast iron.

General quality parameters of CART trees are shown in

Table

8 and are compared with other algorithms.

It turns out that the initial clustering allows in further step

an error-free classification of the materials. The best results of

validation were obtained for the neural networks, but all the

used algorithms, even decision trees, have gained an over

95 % quality, which represents a huge improvement, with

the CART algorithm having a particularly high sequence fit

in the similarity test G

. After the introduction of cluster anal-

ysis done with EM algorithm, the classification quality has

approached 100 %, which means that clustering enables an

error-free classification only in the situation where the user

does not want to find a particular material but one of the four

material groups.

5 Summary

The article presents the results of experimental studies of ma-

terials and is enriched with the data analysis aimed to classify

these materials according to the specified mechanical proper-

ties. The advantages of modern casting materials, such as ADI

and NCIC, were presented.

To present these materials t o the potential users and

make them more popular, models based on machine learn-

ing techniques were built. Their task is to support the se-

lection of construction materials at the stage of product

design. The models are based on the training data sets

Table 8 The results of

classification based on cluster

analysis—comparison of

algorithms

Clusters CART SVM ANN

Fitness Gini Index Line RBF RBF MLP MLP

C =8 C =10 C =10 4–3–44–8–4

γ =0.25 γ =3

3.1 0.57 0.57 0.21 0.06 0.12

42.9 6.53 6.53 4.20 4.06 6.12

Incompatibility [%] 9.8 6.00 6.00 1.38 1.38 2.07

Learning [%] 97.9 % 97.4 97.4 98.7 98.6 97.9

Testing [%] – 94 94 100 100.0 100.0

Validation [%] 97.6 % 96.5 96.5 99.02 100.0 100.0

Fig. 11 a The CART tree for data preceded by cluster analysis. b The importance of predictors

Int J Adv Manuf Technol (2016) 87:1077–1093 1091

obtained from experimental studies carried out on the sam-

ples of materials and from the norms and standards

concerning other materials, such as cast steel, grey cast

iron, ductile iron, etc.

In the train ing process, 18 models with different parameters

using different algorithms were built but all failed to give the

satisfactory results of classification; only the cluster analysis

allowed re-training of models and achieving nearly faultless

classification of the groups of materials with similar properties.

Thus, constructed methodology of comparative analysis

also provides a convenient exploratory path for analyses ex-

panded with the new content of materials either taken from the

groups listed here or entirely new like bronzes or aluminum

alloys, whose structural characteristics can be combined and

used as a range of options by the technologist or designer.

The proposed methodology indicates the similarities

between materials and allows combining various materials

in clusters. Errors in the algorithm classification also em-

phasize the subtlety of the differences in material proper-

ties in particular groups, to mention as an example the

austenitic ductile iron and ADI-6 with the chemical com-

position as given in Table

The developed method of comparative analysis cannot be

the sole criterion for the selection of material, due to the mere

fact that it takes into account only the characteristics selected

for analysis, disregarding other material properties, e.g., wear

resistance or density. However, this is due to a small number

of the experimental data and not to the limitations imposed by

the algorithms. The methods shown in the article can be suc-

cessfully applied to more complex input vectors and thus can

be used in applications capable of supporting not only the

rough decisions or offers made to customers in production

plants but also the detailed process of product design done

by the technologist.

Acknowledgments Financial support of The National Centre for

Research and Development LIDER/028/593/L-4/12/NCBR/2013 is

gratefully acknowledged.

Open Access This article is distributed under the terms of the Creative

Commons Attribution 4.0 International License (http://

creativecommons.org/licenses/by/4.0/), which permits unrestricted use,

distribution, and reproduction in any medium, provided you give appro-

priate credit to the original author(s) and the source, provide a link to the

Creative Commons license, and indicate if changes were made.

References

1. Daber S, Prasad R (2008) Influence of austenitising temperature on

the formation of strain induced martensite in austempered ductile

iron. J Mater Sci 43:4929–4937

2. Olson BN, Moore DJ, Rundman KB (2002) Potential for practical

applications of ausforming austempered ductile iron. AFS

Transactions 111:965

3. Putatunda SK (2001) Development of austempered ductile cast iron

(ADI) with simultaneous high yield strength and fracture toughness

by a novel two-step austempering process. Mater Sci Eng A 315:

70–80

4. Bayati H, Elliot R (1999) The concept of an austempered heat

treatment processing window. Int J Cast Metals Res 11:413–417

5. Skoczylas P, Krzyńska A, Kaczorowski M (2011) The comparative

studies of ADI versus Hadfield cast steel wear resistance. Arch

Foundry Eng 11:123–126

6. Tybulczuk J, Kowalski A (2003) ADI. The properties and application

in industry . Castings atlas. Instytut Odlewnictwa, Kraków (in Polish)

7. Keough JR (2010) Austempered ductile iron (ADI)—agreenalter-

native. American Foundry Society, USA

8. Pietrowski S, Gumienny G (2010) Bainite obtaining in cast iron

with carbides castings. Arch Foundry Eng 10:109–114

9. Gumienny G (2010) Bainitic-martensitic nodular cast iron with

carbides. Arch Foundry Eng 10:63–68

10. Gumienny G (2010) Chromium and copper influence on the nodu-

lar cast iron with carbides microstructure. Arch Foundry Eng 10:

47–54

11. Gumienny G (2012) The effect of nodular cast iron metal matrix on

the wear resistance. Arch Foundry Eng 12:179–186

12. Pietrowski S, Gumienny G (2012) Microsegregation in nodular cast

iron with carbides. Arch Foundry Eng 12:127–134

13. Gumienny G (2013) Carbidic bainitic and ausferritic ductile cast

iron. Arch Metall Mater 58:1053–1058

14. Gumienny G, Klimek L (2014) Effect of temperature on ausferritic

nodular cast iron microstructure. Arch Foundry Eng 14:21–24

15. Regulski K, Szeliga D, Kusiak J (2014) Data exploration approach

versus sensitivity analysis for optimization of metal forming pro-

cesses. Key Eng Mater 611–612:1390–1395

16. Regulski K et al. (2012) Data exploration approach in control of

metal forming manufacturing chain: example of fasteners produc-

tion. Steel Research International spec. ed: 611-612:1319–1322

17. Kluska-Nawarecka S, Wilk-Kołodziejczyk D., Regulski K et al

(2011) Rough sets applied to the RoughCast system for steel cast-

ings. In: Nguyen NT et al. (ed) Intelligent information and database

systemsof the series Lecture Notes in Computer Science, Berlin;

Heidelberg: Springer-Verlag 6592:52–61

18. Nawarecki E, Kluska-Nawarecka S, Regulski K (2012) Multi-

aspect character of the man-computer relationship in a diagnostic-

advisory system. In: Hippe ZS et al (eds) Human-computer systems

interaction: backgrounds and applications Springer, Heidelberg

19. Kluska-Nawarecka S, Wilk-Kolodziejczyk D, Reguslki K (2011)

Practical aspects of knowledge integration using attribute tables

generated from relational databases, computational collective intel-

ligence—technologies and applications, Springer Lecture Notes in

Computer Science

20. Górny Z, Kluska-Nawarecka S, Wilk-Kolodziejczyk D et al (2015)

Methodology for the construction of a rule-based knowledge base

enabling the selection of appropriate bronze heat treatment param-

eters using rough sets. Arch Metall Mater 60:309–315. doi:

10.

1515/amm-2015-0050

21. Kluska-Nawarecka S, Wilk-Kolodziejczyk D, Dajda J et al (2014)

Computer-assisted integration of knowledge in the context of identifi-

cation of the causes of defects in castings. Arch Metall Mater 59:743–

746. doi:

10.2478/amm-2014-0124

22. Kluska-Nawarecka S, Regulski K, Krzyżak M et al (2013) System

of semantic integration of non-structuralized documents in natural

nguage in the domain of metallurgy. Arch Metall Mater 58:927–

930. doi:

10.2478/amm-2013-0103

23. Kluska-Nawarecka S, Sniezynski B, Parada W et (2014) The use of

LPR (logic of plausible reasoning) to obtain information on inno-

vative casting technologies. Arch Civ Mech Eng 14:25–31

1092 Int J Adv Manuf T echnol (2016) 87:1077–1093

24. Gorny Z Kluska-Nawarecka S, Wilk-Kolodziejczyk D (2013 )

Heuristic models of the toughening process to improve the proper-

ties of non-ferrous metal alloys. Arch Metall Mater 58:849–852

25. Glowacz A, Glowacz W, Glowacz Z et al (2015) Recognition of

armature current of DC generator depending on rotor speed using

FFT, MSAF-1 and LDA. Maintenance Reliability 17:64–69

26. David J, Švec P, Frischer R et al (2014) The computer support of

diagnostics of circle crystallizers. Metalurgija 53:193–196

27. David J, Jancíiková Z, Frischer R, et al (2013) Crystallizer’sdesks

surface diagnostics with usage of robotic system. Arch Metall

Mater 58:907–910

28. Wilaszek K, Wojcik T, Oplainski A et al (2012) Internet identity

analysis and similarities detection. Multimedia Communications,

Services and Security, Berlin, pp 369–379

29. TurekW, Opalinski A, Kisiel-Dorohnicki M et al (2011) Extensible

web crawler—towards multimedia material analysis. Multimedia

Communications, Services and Security, Berlin, pp 183–190

30. Ole jarczyk-Wożeńska I, Adrian H, Adrian A et al (2012)

Parametric representation of TTT diagrams of ADI cast iron.

Arch Metall Mater 57:981–986. doi:

10.2478/v10172-012-0065-9

31. Olejarczyk-Wo żeńska I, Adrian A, Mrzyglod B et al (2014)

Math ematical model of the process of pearlite a ustenitiz ation.

Arch Metall Mater 59:613–617

32. Kmiec M, Glowacz A, Dziech A et al (2012) Towards robust visual

knife detection in images: active appearance models initialised with

shape-specific interest points. Multimedia Communications,

Services and Security, Communications Computer Information

Science 287:148–158

33. Witten I, Frank E (2000) Data mining: practical machine learning

tools and techniques. Morgan Kaufmann, New York

34. Sztangret L, Szeliga D, Kusiak J et al (2012) Application of inverse

analysis with metamodelling for identification of metal flow stress.

Can Metall Q 51:440–446. doi:

10.1179/1879139512Y.

0000000035

35. Rauch L, Sztangret l, Pietrzyk M (2013) Computer system for identifi-

cation of material models on the basis of plastometric tests. Arch Metall

Mater 58:737–743

36. Tomar D, Agarwal S (2015) A comparison on multi-class classifi-

cation methods based on least squares twin support vector machine.

Knowl Based Syst 81:131–147

37. Li Yet, Yun Li, Yan Gao at al (2013) Fault diagnosis of metallur-

gical machinery based on spectral kurtosis and GA-SVM. Adv

Mater Res 634–638:3958–3961. doi:

10.4028/www.scientific.net/

AMR.634-638.3958

38. Peng S, Hua Q, Chenb Y et al (2015) Improved support vector

machine algorithm for heterogeneous data. Pattern Recogn 48:

2072–2083

39. Widodo A, Yang B (2007) Application of nonlinear feature extrac-

tion and support vector machines for fault diagnosis of induction

motors. Exp Sys App 33:241–250

40. Yuan S, Chu F (2007) Fault diagnosis based on support vector

machines with parameter optimisation by artificial immunisation

algorithm. Mech Syst Signal Process 21:1318–1330

41. Laurain Vet, Tóthc R, Pigad D al (2015) An instrumental least

squares support vector machine for nonlinear system identification.

Automatica 54:340–347

Int J Adv Manuf Technol (2016) 87:1077–1093 1093