Vzkumn zmr .

MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

USAGE OF FRACTAL GEOMETRY AND STATISTICS FOR QUALITY

MONITORING

Vlastimil Hota, Frantiek Novotn

Department of Glass And Ceramic Producing Machines,
Technical University of Liberec, Czech Republic

Abstract
Fractal dimension with statistic tools is powerful tool for a complex data
quantification, for a poor quality source searching, a production optimalization and a non-
stability of a production process subsystems searching.
The data having complex and structured character can be also met during a glass
manufacturing. The data may have a form of digitalized pictures, time
series (signals) or topologically one-dimensional dividing lines (especially a surface
roughness). When analysing this data, it is suitable to use - in addition to classic
mathematical statistics - modern tools of the fractal geometry expressing the complexity
degree of structured data by means of a single number, the fractal dimension.
The article presents results of developing three off-line softwares that can be
converted into on-line control systems in the future: an off-line classification of surfaces and
defects pictures, a description of time series, which were obtained from outputs of a glass
production control system and an evaluation of metal surfaces (iron aluminides in
comparison with the chrome-nickel steel) in contact with the glass melt as well as changes of
their quality.

1 Introduction
Although continuously growing a competitive press to increasing quality of products
activates a requirement of an objective measurement and control methods for materials,
processes and productions, many structures (e.g. defects, surface, crack, time series from
dynamic processes) can be hardly described by conventional methods, because they are
complex and irregular. However, a new approach is the application of fractal geometry [1-3]
that is successfully used in science. Even though an application in industry is sporadic and
experimental only, the fractal geometry can be used as a useful tool for an explicit, objective
and automatic description of production process data (laboratory, off-line and potential on-
line).
Now, there are tools application possibilities for monitoring of three basic data format
types: digitalised photos [4, 5, 6], time series [4, 5, 6, 7, 8] and topological one dimension
dividing lines (especially surface roughness) [5].

2 Classification of Surfaces And Defects Photos

The explicit, objective and automatic description of images complexity can be made
by different methods both statistic and fractal dimension. Only some of the possibilities are
presented below. An evaluation of 2D pictures of surface defects is demonstrated on

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

structures of the hole cracks in costume jewellery. Software Matlab 6.5 and HarFa 4.0 [9]
were used for these experimental evaluations.
The process of description has five steps practically:
a. Preparing of samples - structure must be visible, the costume jewellery is cut,
fig. 1, A.
b. Taking digital photographs. Photos of the hole cracks in costume jewellery are
from an electronic microscope, fig. 1, B (it is possible obtain classic
photographs and they scan).
c. Software preparation of the digital photographs, fig. 1, C (cutting of the
photographs, because only some parts of the photos are important for analysis).
d. Analyses of the images.
e. Evaluation analyses results.

Fig. 1 Preparing of samples, taking photographs, software

preparation.

Digital image is a matrix (or matrixes) of pixels (rectangular array of points, fig. 1, D).
Pixels can reach different numbers, which depend on the used format of digital images. The
pixels have numbers between 0 (black) and 255 (white) for the grey 8-bit palette bitmap and
the bitmap has only one matrix. (Colour bitmap has 3 matrixes for RGB colour model one
matrix of red, green and blue colour.)
Fig. 1, C shows two typical poor quality surfaces of costume jewellery holes. The
cutting C-1 has deep cracks and C-2 has a thin structure.

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

2.1 Thresholding
The analyses are based on a technique called "thresholding", that transforms grey or
colour image object into black & white (binary) one. The binary image can be determined
from the grey 8-bit palette bitmap, where black are all pixels which fulfil condition e.g. 0 <=
black <= 100 and all the other pixels become white (100 < white <=255), fig. 2. It means, that
all pixels lesser than or equal to the threshold 100 are black and greater than 100 are white.
(More than one threshold can be used or the technique for matrixes of colour images can be
used too.)

Fig. 2 Thresholding of grey images

The procedure of thresholding can be used for all thresholds of the grey image, 256
binary images are obtained. An analysis is done for all binary images and as far the analysis
produces single number classifying a binary image, a spectrum of dependence between single
number and threshold is given (e.g. fig. 4).
Thresholds between 50 and 150 are suitable for the images of the hole cracks, because
binary images, obtained by these thresholds, show the best structure of the surface.
Thresholds between 10 and 50 show the best large cracks. Binary images produced by the
thresholding with thresholds over 150 contain shadow.

2.2 Percentage of Deep Cracks

The method is suitable for detection of
relatively large and single cracks and defects. The
method computes percentage of pixels with
neighbouring pixels of the same value. The
analysis searches black pixels (value 0) in a binary
image, which have five ore more neighbouring
black pixels. The black pixels represent defect,
structure, cracks, etc. Especially large cracks and
defects contain black pixels with five or more Fig. 3 Pixels on boundary crack
neighbouring black pixels. Fig. 3 shows part of
boundary crack. Black pixel in fig. 3, A has five neighbours and in fig. 3, B has 8.
Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

Fig. 4, A shows spectrum of dependence between percentage of black pixels with five
or more neighbouring black pixels of binary images and thresholds. For detection of large
hole cracks in costume jewellery thresholds from 10 to 50 are the most suitable, fig 4, B. For
the threshold 50, the cutting C-1 has more single cracks and defects, numerically: T50_C-
1=3,17% than the cutting C-2, numerically: T50_C-2=0,8%.

Fig. 4 Percentage of black pixels with five or more neighbouring

black pixels spectrum.

2.3 Box Dimension

The box counting method is shown in [2, 9, 10] and based on fractal geometry. The
analysis describes structure by single number: the box dimension DB. The box counting
method works by laying meshes of different sizes rB and then counting numbers of boxes N
needed to cover a binary image (fig. 5, A) completely (fig. 5, B, C). The number N(rB) of
boxes needed to cover the structure is given by a power law:
D
N (rB ) = const. rB B (1)
DB is the box dimension. Logarithmic dependence between log2N(r) and log2rB is called the
Richardson-Mandelbrot plot (fig. 5, D). The box dimension (that estimate fractal dimension)
can be determined by a slope s of the regression line in fig. 5, D:
log N (rB )
s=D =
B log rB . (2)
The software HarFa 4.0 [9] is used for the analysis and software tools developed in
Matlab 6.5 make data evaluation. The box dimension is multiplied by 1000 for a better
confrontation.
The fractal spectrum of the cuttings C-1 and C-2 are shown in fig. 6. The box
dimensions over threshold 150 are similar, because over the value an influence of shadow is
significant. Results of analysis for threshold 120 are: DB_C-1=1429,6 (C-1) a DB_C-2=1562,4
(C-2), where the higher value represents a greater complexity of the structure in the image.
The cutting C-2 is more structured than the C-1 and the box dimension quantifies the
structures.

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

Fig. 5 Box counting method

Fig. 6 Fractal spectrum

3 Description of Time Series By Fractal Dimension

A lot of production processes have dynamic subsystems, which have an influence to
the production. The production can get a dynamic behaviour from the subsystems and
measured data (time series) from production sensors contain this dynamic influence. The time
series are structured and common statistic data evaluations are not often sufficient [5, 6, 7].
A good quality of products is given by accurate settings of a production process and
sensors, which often produce time series in time monitor. The research target is to develop
software tools describing fast and exactly time series from production processes sensors using
the fractal geometry and statistic analyses. The fractal geometry uses the fractal dimension
that describes the character (complexity) of time series by one number. To analyse time series
statistical methods, power spectral analysis and the fractal dimension estimation are used. The
estimation of the fractal dimension is made by the rescaled range method, the aggregated
variance method [7, 10] and by the box counting method from an "iso-set" [6, 7]. The article
shows briefly a background of the rescaled range method and the box counting method, only.

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

3.1 Rescaled Range Analysis

The rescaled range analysis (R/S) represents method for estimating the fractal
dimension of self-affine fractals and uses statistical tools. The Hurst exponent H is computed
in the analysis, and the fractal dimension of the time series can be calculated from the
relationship between the Hurst exponent H and the fractal dimension:
DRS = 2 H , (3)
where R/S dimension DRS denotes the fractal dimension estimated from the Rescaled Range
analysis. The R/S dimension has value from 1 to 2, and Hurst exponent has value from 0 to 1.
More about the method can be found in [7, 10].

3.2 Box Dimension from Iso-set

An "iso-set" is constructed from the time series as in fig. 7 that contain zeros and ones
and the fractal dimension is estimated from this set. The "iso-set" (as time series) can be
generated by two basic ways, where the ones represent crossings of a pre-selected threshold.
The "iso-set" is generated by setting of suitable thresholds and marking the time at which the
time series cross these thresholds (fig. 7). The threshold values can be perceptually dependent
on the time series average (a floated threshold value) or can be pre-selected fixed values.
Time series I I
k * V max Thresholds V max Vmean

1
Iso-set
0

ts
Box without crossing
rB= b*ts Box containing
crossing

Fig. 7 Construction of iso-set by marking each time at

which trace crossed the chosen threshold values and a
principle of Box Dimension computing

The fractal dimension of the "iso-set" is estimated by using box counting method, that
is described in chapter 2.3. The principle of the box dimension method used for iso-set is
given in fig. 7. (The box size rB=ts.b) Starting from box size rB=ts (ts is sampling time
interval), the number of boxes that contain a crossing is recorded. The box size is then
increased by an increasing factor b and the procedure continues until the entire iso-set is
contained in one single box. This is illustrated for the factor b=2 in fig. 7. The box dimension
DB is determined from the central slope of the regression line of the Richardson-Mandelbrot
Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

plot (logarithmic dependence between log2N(r) and log2rB). For more information about the
method please see [6, 7, 8].

3.3 Example of results

Application possibilities of the describing fast and exactly time series from production
process sensors using the fractal geometry and statistic analyses are large. The fig. 8 shows
results of a tank furnace siege time series analysis, where the standard deviation, the R/S
dimension, the aggregated variance dimension and the box dimension were used. The average
temperature indicates a temperature profile of the siege and implicitly a temperature profile of
glass melt on the siege. The fractal dimensions result is a relatively large decrease of the
dimension in positions 5 and 6 and a subsequent ascension between positions 6 and 8. The
large decrease indicates simple time series and higher dimension represents complex time
series. The decrease is in position, where a change of longitudinal currents is expected. The
complex time series represent changes in temperature and the movement of melt gives the
changes. It is appear that the fractal analysis can be used for detection of glass melt currents.
The fractal measures, obtained from time series derived from a production process, can be
also used for monitoring and control of the process [6, 7].

Fig. 8 Results of a tank furnace siege time series analysis

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

4 Classification of Dividing Lines Profile Evaluation

The fractal analysis and statistic tools are tested for a quantification of metal surfaces
changes of relatively new materials: iron aluminides in comparison with currently used
chrome-nickel steels in contact with the glass melt.
Analyses were performed on samples of the iron aluminide Fe28Al4Cr0,1Ce and the
chrome-nickel steel X15CrNiSi25 21 - EN 10095 (AISI 310) that were exposed to static and
dynamic glass melt effects in different temperatures.
The methodology of surface profile evaluation is shown in fig. 9. Firstly, classic or
better a digital camera takes a photo of a surface layer profile from a microscoped
metallographic sample, fig 9, A. The classic photo of the layer profile has to be scanned,
that extend time needed to analyses.

Photo of surface layer

(chrome-nickel steel
after dynamic test)

(Conversion of photo to
digital form and)
dividing line generation
(line between metal alloy
and surrounding - profile)

r1=250 N1=18,5 L1 = 4625 Computing of compass

r2=500 N2=4,5 L2 = 2250 dimension: measurement
of profile length by
different ruler rRi:
Li ( rRi ) = N i ( rRi ) rRi

Computing of compass
dimension: generation of
Richardson Mandedlbrot plot,
choosing central slope and a
compass dimension compute DR
Central slope from central slope by relation:
log L(rOBV )
DR = 1
log rOBV

Fig. 9 Analysis of surface layer, dividing line generation from

photography, evaluation by statistic and compass dimension
Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

Secondly, a dividing line is generated from the digital photography, fig. 9, B, by the
software tool that exactly defined the curve between material alloys and a surrounding - a
dividing line as a curve is obtained. The generated dividing line is a binary image, where the
line is from the black pixels (value 0) and the surrounding from white pixels (value 1). The
width of images is 2272 that matches 57,7 m. Statistic tools and (or) the fractal dimension
can describe the curve fig. 9, C, D.

4.1 Evaluation by statistic tools and fractal geometry

The average surface roughness (Ra) and the maximum roughness (Rm or Rmax) were
used for the evaluation of basic statistic roughness parameters. The maximum roughness Rm is
a distance between a deepest valley and a highest peak. The average surface roughness (Ra) is
an area between the roughness profile and its mean line, or an integral of the absolute value of
the roughness profile height over the evaluation length. Examples of results are shown in fig.
10.
As mentioned above, the fractal dimension describes complexity by a single number.
The fractal dimension can be estimated by many different methods. A compass method is one
of them and the method is based on the measurement of the dividing line (roughness profile)
by different size of a ruler (fig. 9, C) via the equation:
Li ( ri ) = N i ( ri ) ri (4)
Li is a length in i-step of the measurement, Ni is a number of steps needed for the
measurement and ri is a ruler size. If the line is fractal and hence the fractal dimension is
larger than the topological dimension, the measured length increases as the ruler size is
reduced (fig. 9, C). Using equations (1) and (4):
D 1 D
Li (ri ) = N i (ri ) ri = const.ri R ri = const. ri R
(5)
DR is the compass dimension.
The compass dimension is determined from slope s of the regression line from the
dependence between log2 r a log2 L (fig. 9, D):
log 2 L( r )
DR = 1 s = 1
log 2 r . (6)
Although the typical dependence consists of three-parts slope, only central part (the
central slope) is important for the compass dimension computing. The compass dimension DR
is multiplied by 1000 for better confrontation, DR 1000.

4.2 Examples of results

The examples are shown in fig. 10. Six digital photos of every metal sample profile in
different position were made and analysed. The presented results Ra, Rm and DR 1000 are an
average of six measurements on a tested sample and fig. 10 shows only one example of the
dividing lines.

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

Chrome-nickel steel EN 10095 (AISI 310) FA - iron aluminide on base Fe3Al

Dividing line - roughness profile (size in Rm [m] Rm [m] Dividing line - roughness profile (size
pixels) Ra [m] Ra [m] in pixels)
Do 1000 [-] Do 1000 [-]
3.07 1.56

0.79 0.29
Ground state

1029.2 1055.8

4.28 1.54

1.01 0.47
1100 C, 24 hour

1056.4 1012.0
O

28.11 5.52

8.59 1.39
1250 C, 96 hour

1121.6 1100.3
O

24.08 22.64

6.09 8.09
1350 C, 96 hour

1185.4 1124.0
O

Fig. 10 Examples of dividing lines chrome-nickel steel material and iron aluminide, after static
glass melt effects in different temperatures and results of analyses

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces
 Vzkumn zmr . MSM 242100001
Optimalizace vlastnost stroj a pracovnch proces

5 Conclusion
Many natural structures cannot be described by conventional methods, because they
are complex and irregular. A new approach is the application of fractal geometry that is
successfully used in science, but an application in industry is sporadic and experimental only.
The data for the fractal measurement may have a form of digitalized pictures, time
series (signals) or topologically one-dimensional dividing lines (especially a surface
roughness).
The results of our research show that the fractal dimension is potentially a powerful
tool for explicit, objective and automatic description and quantification of complex data. In a
near future the methodology can be used for an off-line and an on-line measurement. The
possibilities of successful applications in industry are believed to be large.
This work was done under of the research project No. MSM 242100001, which is
financed by the Ministry of Education of the Czech Republic.
References
[1] MANDELBROT. B. B. The fractal geometry of nature. New York: W. H. Freeman
and Co.,1982.
[2] PEITGEN, H.O., JUERGENS, H. and SAUPE, D. Chaos and Fractals: New Frontiers
of Science. New York; Berlin; Heidelberg: Springer-Verlag, 1992.
[3] BUNDE, A. and HAVLIN, S. Fractals in science. Berlin: Springer, 1994.
[4] HOTA, V. and NOVOTN, F. Possibilities of Fractal Geometry Application in
Glass Industry. In: Glass Odyssey. Montpellier (France) 2002. p. 120, 6 pages.
[5] HOTA, V. and NOVOTN, F., HOTA A. Interpretation of the production data by
means of the fractal geometry and statistical analyses. Skl a keramik 53 (2003), no. 7-8, p.
148-154
[6] HOTA, V., NOVOTN, F., CHRPOV, E., LANG, M.. Application of Fractal
Geometry In Glass Industry Overview. ISQVPFD 2002. Bled, Slovinsko, 2002. 9 pages.
ISBN 961-6238-58-2
[7] HOTA, V., CHRPOV, E., LANG, M. Application of Fractal Dimension In:
Carding. Paper And Others Production Processes - Final Report of Sub Team; CeVis,
Technical University of Liberec (2000), available in: www.noviscam.herts.ac.uk
[8] KERPE, A.R. Acquisition of Turbulent Data, Major Project Report, University of
Hertfordshire, 1998/99
[9] ZMESKA, O., NEZADAL, M., BUCHNICEK, M. AND SEDLAK, O. HarFA and
HARFA: E-JOURNAL. http://www.fch.vutbr.cz/lectures/imagesci/ Brno: stav fyzikln a
spotebn chemie, 2002
[10] EVERTSZ, C.J.G., PEITGEN, H.O., VOSS, R.F. Fractal Geometry and Analysis,
World Scientific Publishing Co.Pte. Ltd., Singapore, 1996
Vlastimil Hota, Department of Glass And Ceramic Producing Machines, Technical
University of Liberec, Halkova 6, 461 17 Liberec, Czech Republic
tel. +420 485 354 129, fax: +420 485 354 157, e-mail: vlastimi.hotar@vslib.cz
Frantiek Novotn, Department of Glass And Ceramic Producing Machines, Technical
University of Liberec, Halkova 6, 461 17 Liberec, Czech Republic
tel. +420 485 354 103, fax: +420 485 354 157, e-mail: frantisek.novotny@vslib.cz

Vzkumn zmr . MSM 242100001

Optimalizace vlastnost stroj a pracovnch proces