736 DOI 10.1002/mawe.201500347 Mat.-wiss. u. Werkstofftech. 2015, 46, No.

Cluster analysis of acoustic emission signals in pitting

corrosion of low carbon steel
Clusteranalyse des Schallemissionssignals bei Lochkorrosion
von niedrig legiertem Stahl

H. Bi1, Z. Li2, D. Hu1, I. Toku-Gyamerah2, Y. Cheng1

The pitting corrosion characteristics of low carbon steel specimens are studied by
acoustic emission (AE) and electrochemical techniques, in a 3.0 wt.% NaCl solution
acidified to pH 2.0. The acoustic emission signals generated by pitting corrosion are
classified based on multiple acoustic emission parameters using K-means clustering
algorithm, then each classified signals are analyzed by acoustic emission param-
eters correlation plot and distribution with time. Furthermore, each acoustic source
characteristics is extracted using Gabor wavelet transform (WT) in the time and fre-
quency domain. An error back propagation (BP) artificial neural network (ANN) is
trained according to the classified signals, so as to successfully identify the acoustic
emission signals from parallel experiments. Experimental results show that the hy-
drogen bubble activation, oxidized film rupture and pit growth are typical acoustic
emission sources in pitting corrosion process, which can be effectively classified by
cluster analysis and recognized by back propagation neural network. The data gath-
ered from laboratory tests combined with the real data from acoustic emission on-
line storage tank floor inspection can help to evaluate the bottom corrosion severity
and interpreter the corrosion source, further to make the on-site testing more reliable
and reduce the risk.

Keywords: Low carbon steel / pitting corrosion / acoustic emission / cluster analysis /
artificial neural network

Schlüsselwörter: Niedrig legierter Stahl / Lochkorrosion / Schallemission /

Clusteranalyse / künstliches neuronales Netzwerk

1 Introduction rochemical products. However, over 80% of the

tanks shutdown, bottom perforation and leakage ac-
Low carbon steel is widely used as the main material cidents are caused by tank bottom corrosion, it can
for aboveground metal storage tank floor in the pet- cause very serious consequences on the environment
rochemical industry. The atmospheric pressure metal and health and safety, producing a very wide range
storage tanks play an irreplaceable role in storage of hazards and disasters. Therefore, the tank bottom
and transportation of crude oil, oil products and pet- corrosion has attracted more and more attentions all
over the world in recent decades.

1 College of Electromechanical Engineering, Qingdao

University of Science & Technology, Shandong, Corresponding author: Haisheng Bi, College of Electro-
266042, China mechanical Engineering, Qingdao University of
2 College of Pipeline and Civil Engineering, China Uni- Science & Technology, Qingdao, Shandong, 266042,
versity of Petroleum, Qingdao, Shandong, 266580, China
China E-Mail: [email protected]

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7 Cluster analysis of acoustic emission signals in pitting corrosion of low carbon steel 737

The tank bottom corrosion damage is a very com- 6, 7]. The hydrogen bubbles activities resulting from
plex process. Several ions such as Na+, Ca2+, SO42−, proton reduction and corrosion products movements
Cl−, S2−, and dissolved O2, H2S, CO2 in the sedimen- have been suggested as the main acoustic emission
tary water on tank bottom can induce the corrosion, sources in both crevice and pitting corrosion [1, 5, 8,
especially the chloride ions can invade the local de- 9]. Acoustic signals released from passive film rup-
terioration anticorrosion coating and result in pitting ture can also be detected in stainless steel, aluminum
corrosion under the coating, even perforation and and aluminum alloy pitting corrosion [9, 10]. Acous-
leaking [1]. Hence, periodic inspections are neces- tic emission sources of corroded metal and products
sary to prevent destruction from bottom corrosion. deposit can be observed in uniform corrosion [2].
Several conventional testing techniques such as Moreover, good correlations have been found be-
magnetic flux leakage (MFL) testing have been ap- tween corrosion rate and acoustic emission activity
plied to locate and describe the defects. However, (hits, events and amplitude), some empirical formu-
these techniques require a great deal of process inter- las to estimate corrosion severity are presented based
ruption and preparation, for example, emptying or on mass loss, polarization resistance measurements,
cleaning the tank bottom. Therefore, cost-effective electrochemical impedance spectroscopy (EIS) and
and on-line diagnosis techniques are increasingly hydrogen evolution rate in previous papers [6, 7,
needed for detecting and evaluating corrosion da- 11–13]. As well, extracting and distinguishing the
mage of storage tanks bottom. The acoustic emission features of different acoustic emission sources can
method is considered to meet the requirements of an help to provide a reliable basis for acoustic emission
on-line inspection technique and has gained popular- on-line tank floor inspection and evaluation [14–16].
ity. Nevertheless, very few studies have been carried out
Acoustic emission is a non-destructive technique on the cluster analysis of acoustic emission signals
(NDT), defined by American Society for Testing of for the pitting corrosion of low carbon steel. In that
Materials (ASTM) as the phenomena whereby tran- context, the aim of the present study is to classify
sient elastic waves are generated by the rapid release the acoustic emission sources based on multiple
of energy from localized sources within a material, acoustic emission parameters using the K-means
also known as the stress wave emission or micro- clustering algorithm and back propagation artificial
seismic activity. It is widely used to detect process neural network.
of plastic yield deformation, fatigue fracture, corro-
sion damage and surface friction. Extracting and
analyzing the effective signal to evaluate the features 2 Experimental procedure
of acoustic emission sources is the bottleneck to
acoustic emission on-site inspection. Various re- 2.1 Material
searchers have shown that the acoustic emission sig- Specimens were cut into small squares using wire-
nals are closely related to corrosion factors and elec- electrode cutting method and its size was 2.9 × 2.9 ×
trochemical mechanisms, other researchers also have 0.2 cm3. Composition of steel is given in Table 1.
attempted to identify and characterize acoustic emis- The exposed surface was polished with a series of
sion sources in corrosion processes [2–5]. Different silicon carbide sheets of up to 1200 grit. The speci-
acoustic emission signals are generated by different mens were rinsed with de-ionized (DI) water then
corrosion sources. According to the source of the acetone and alcohol, dried in a desiccator, and last
signals, all the acoustic emission sources can be di- weighted, labeled and stored to be used.
vided into two groups: direct acoustic emission
sources and indirect acoustic emission sources (sec- 2.2 Experimental device
ondary acoustic emission sources). The acoustic The experiments were conducted in 3.0% NaCl solu-
emission sources released from metal deformation tion with pH adjusted to 2.0 with the addition of hy-
and cracking itself are considered as direct acoustic
emission sources, whereas the hydrogen bubbles ac- Table 1. Composition of the studied materials
tivation, passive film rupture, pits propagation and Element Fe C Si Mn P S
corrosion products movement in the corrosion pro- wt.% Balance 0.17 0.3 0.61 0.05 0.045
cess belong to indirect acoustic emission sources [2,

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

738 H. Bi et al. Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7

Figure 1. Schematic diagram of experiment system

drochloric acid (HCl) at room temperature. The spe-

cimen was sealed by epoxy resin and inserted in the
center hole of a glass cell bottom, the edges of the
hole was filled with silicone sealant to insure no
leakage, Figure 1. An acoustic emission sensor was
mounted on the back of specimen for real-time mon-
itoring corrosion signals. The electrochemical meas-
urements were conducted simultaneously to confirm
the result of acoustic emission detection.

2.3 Electrochemical devices

Electrochemical measurements were carried out with Figure 2. Frequency response curve of the acoustic emission
a Princeton Applied Research PARSTAT 2273. The sensor
minimum current resolution of system is 1.2 fA and
the minimum potential step is 2.5 μV. The size of extract acoustic emission parameters. The frequency
the typical three-electrode electrochemical cell was response curve of the sensor was shown in Fig-
25 × 25 × 25 cm3. The sample used as the working ure 2. Acoustic emission acquisition threshold was
electrode (WE), its electrochemical potential was fixed at 30.2 dB according to background noise lev-
measured with a saturated calomel (SCE) reference el and filtered by band pass between 25 and
electrode (RE) and a platinum plate as the counter 850 kHz.
electrode (CE).

2.4 Acoustic emission monitoring

3 Results and discussion
Acoustic emission instrumentation consisted of a
piezoelectric sensor (VS150-RIC type from Vallen, The corrosive solution was filled into a glass cell at
integral preamplifier: 34 dB gain) and an acquisi- 80% of the container’s height, standing for about 2 h
tion device AMSY-5 from Germany Vallen-Sys- until the solution become stable. Then the open cir-
teme GmbH, the AMSY-5 was a fully digital multi- cuit potential (OCP) was measured and the curve
channel acoustic emission measurement system to was shown in Figure 3. The change of open circuit

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7 Cluster analysis of acoustic emission signals in pitting corrosion of low carbon steel 739

rupture of oxidized film invaded by chloride ion,

the deposits and friction of corrosion products
could also generate obvious acoustic emission sig-
nals.

3.1 Acoustic emission parameters extraction

and correlation analysis
Using waveform or frequency spectrum directly for
cluster analysis is a huge computation, and the sin-
gle frequency spectrum is difficult to describe the
differences between the signals. To avoid the man-
agement of big amounts of data, acoustic emission
signals (“hits”) are usually characterized by wave-
form parameters, the typical characteristic param-
Figure 3. Open circuit potential decay curves for low carbon eters are represented in Figure 4. These parameters
steel in 3% NaCl solution at pH 2.0
can be related to the corrosion types and severity
that generated the signals. From this diagram, five
potential was less than 1 mV over 10 minutes, which typical characteristic parameters are defined as fol-
indicated that the corrosion system had reached a low:
steady state. Then potentiostatic test was kept at a 1. Amplitude: maximum amplitude of a given
level of 0.80 V (vs. SCE) to control the pitting corro- acoustic emission signal in dB (logarithmic scale
sion. The 0.80 V level was the pitting potential when with a reference voltage of 1 μV at the sensor out-
anode oxidized film breakage observed in potentio- put).
dynamic tests with a scan rate of 0.5 mV s–1. The 2. Rise time: time between the first threshold cross-
result of an electrochemical corrosion process of the ing and the peak amplitude.
carbon steel was shown below. The corrosion pro- 3. Energy: the integral of the squared amplitude
cess of metal comprises an oxidation and reduction over time of signal duration and is expressed in
reaction. term of energy units (eu), 1 eu = 10−14 V2 s.
Anodic reaction: 4. Duration: time between the first and the last
threshold crossing.
M ! M2þ þ 2e ð1Þ 5. Counts: number of times that the threshold is
overshot for a given acoustic emission signal.
Cathodic reaction:

2Hþ þ 2e ! H2 ð2Þ

O2 þ 4Hþ þ 4e ! 2H2 O ð3Þ

Overall reaction:

ðMnþ þ nCl Þ þ nH2 O ! MðOHÞn þ nHþ þ nCl

ð4Þ

Acoustic emission monitoring was conducted si-

multaneously to acquire and store data. In the ac-
celerated corrosion process, the oscillation, move-
ment, and breakage of hydrogen bubbles gener- Figure 4. Acoustic emission signal typical waveform and
ated acoustic stress waves, the exfoliation and parameters

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

740 H. Bi et al. Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7

Besides above five characteristic parameters, three tween the acoustic emission characteristic param-
other derived parameters were also introduced in this eters. And the relationship can be established using
paper: the Pearson correlation coefficient. The Pearson cor-
(I) RA value: a calculated feature derived from relation coefficient can be defined as
“Rise time” divided by “Amplitude”, it is also P P
known as “Signal Slope”, in ms V–1. zx zy ðx 
xÞ ðy 
yÞ
r¼ i
¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
i
P ffi ð8Þ
Rise time ðN  1Þ ðx 
xÞ2 ðy 
yÞ2
RA ¼ ð5Þ
Peak Amplitude i i

(II) AF value: a calculated feature obtained from ðx 
xÞ ðy  y
Þ

“Counts” divided by “Duration”, which determines Where zx ¼ and zy ¼ ,

x,

y, σx
σx σy
an average frequency over one acoustic emission hit,
and σy are mean and standard deviation of x and y,
in kHz
respectively. If r ¼ 1, the two variables would be
Counts perfectly positively correlated; If r ¼ 1, the two
AF ¼ ð6Þ would be perfectly negatively correlated; If r ¼ 0,
Duration
the two would be no relationship.
(III) RD value: the ratio of “Rise time” and “Dura- 16257 groups of acoustic emission hits were se-
tion”, it can describe the steepness of one acoustic lected from the experimental acquisition data, corre-
emission waveform lation among eight characteristic parameters, A
(Amplitude), R (Rise time), E (Energy), D (Dura-
Rise time tion), C (Counts), RA value, AF value and RD value
RD ¼ ð7Þ were analyzed using Pearson correlation. The corre-
Duration
lation matrix was shown in Table 2.
After extracted the characteristic parameters, to As shown in the correlation coefficient matrix in
avoid unequal usage of the inputs that have different Table 2, the Pearson correlation coefficient between
physical meaning and dimension, eliminate indica- R and RA value is almost linearly correlated. Using
tors dimension effects and solve their comparability, both features does not provide much more informa-
the original data were standardized using “Z-score” tion than using one. It contributes little to the cluster
statistical method. but enlarging the work remarkably. In addition, it
During the pattern recognition analysis of neural would have a negative impact on neural network
network, each two variables are required to be less analysis later. Since parameter R is greatly dependent
relevant as much as possible for avoiding “dimen- on the voltage threshold, and it contains less infor-
sionality curse”, a linear increase in the number of mation on acoustic emission signals than RA value.
features requires an exponential increase in the size Therefore, the seven acoustic emission parameters,
of data set. Therefore, before cluster and neural net- A, E, D, C, RA value, AF value and RD value were
work analysis, correlation analysis is required be- selected for cluster and neural network analysis.

Tabelle 2. Pearson correlation coefficient matrix

Pearson Correlation A R E D C RA value AF value RD value

A 1.000 0.050 0.506 0.271 0.400 –0.151 –0.087 –0.303
R 0.050 1.000 0.137 0.354 0.353 0.904 –0.207 0.215
E 0.506 0.137 1.000 0.329 0.505 –0.019* –0.118 –0.150
D 0.271 0.354 0.329 1.000 0.858 0.237 –0.366 –0.207
C 0.400 0.353 0.505 0.858 1.000 0.144 –0.221 –0.191
RA value –0.151 0.904 –0.019* 0.237 0.144 1.000 –0.228 0.346
AF value –0.087 –0.207 –0.118 –0.366 –0.221 –0.228 1.000 0.335
RD value –0.303 0.215 –0.150 –0.207 –0.191 0.346 0.335 1.000
* Correlation is significant at the 0.05 level, others are at the 0.01 level (two tailed).

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7 Cluster analysis of acoustic emission signals in pitting corrosion of low carbon steel 741

3.2 K-means clustering

K-means algorithm, also known as fast clustering
algorithm, was first proposed by James MacQueen
in 1967. The data was recorded as points on K-di-
mensional space, and K clusters were determined
based on minimizing the summation of square dis-
tance of each data point in a subset to the center of
the subset in which it is partitioned. In the cluster-
ing process, the category of each sample was ad-
justed continuously to minimize the sum of squared
errors of each category. Suppose given a data set
X ¼ fx1 ; x2 ; . . .; xN g, xn 2 Rd is available, the al-
gorithm aims at partitioning this data sets into K
disjoint subsets,C1 ; C2 ; . . . Ck through minimizing
the clustering error as follow Figure 6. Acoustic emission hits’ counts-duration distribution

N P
P k acoustic emission signals intensively during this
Eðm1 ; m2 ; . . . mk Þ ¼ Iðxi 2 Ci Þ kxi  mj k2 short time. The acoustic emission hits’ amplitude
i¼1 j¼1
distribution over time was shown in Figure 5. It in-
ð9Þ dicated that there existed two clusters among the
 acoustic emission signals. Therefore, the K-means
1 ðxi 2 Ci Þ
IðXÞ ¼ : ð10Þ number was set to 2 preliminarily in this stage, the
0 ðxi 2
= Ci Þ cluster analysis was shown in Figures 6 and 7.
The acoustic emission signals were divided into
When satisfied the predetermined error, stop the two clusters accurately with K-means algorithm,
iterative procedure. Where mj is the center of cluster Figure 6 and 7. Amplitude distribution was within
Cj [18–21]. 32–45 dB in Cluster 1, and the amplitude was a little
It was observed that current density was relatively low, most of the ring-down counts number in acous-
large before 250 s of potentiostatic polarization in tic emission hits was no more than 25, the signal ab-
the initial stage, a large amount of hydrogen bubbles solute energy value was less than 100 eu but concen-
were produced by the proton reduction. Meanwhile trated, and the duration was mostly below 300 μs.
in about 100 ~ 230 s, the oxidized films exfoliation, While in Cluster 2, the amplitude was higher than
shrink and rupture were observed, which generated 45 dB, the counts and duration were larger than that

Figure 5. Acoustic emission hits’ amplitude distribution with Figure 7. Acoustic emission hits’ energy-duration distribution
time (1 eu = 10−14 V2 s)

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

742 H. Bi et al. Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7

emission signals clustering was shown in Figures 8

and 9.
The correlation plot, between acoustic emission
hits’ duration and energy was shown in Figure 8. It
was shown that the three classes of signals were dis-
tinguished by clustering algorithms. Comparing with
the previous acoustic emission signals within 250 s
at the initial stage, a large number of Cluster 3 sig-
nals newly appeared and its duration was mainly be-
low 100 μs, while the Cluster 1 signals reduced shar-
ply. From the amplitude distribution, Cluster 2 and
Cluster 3 were mainly acoustic emission sources in
the whole sampling time and Cluster 3 signals domi-
nated overwhelmingly, consistent with our observa-
Figure 8. Acoustic emission hits’ energy-duration distribution tions, Figure 9.
Combined with electrochemical test, after 1000 s
potentiostatic test, the hydrogen bubbles activity re-
of in Cluster 1, more abrupt energy released and its duced remarkably, the corrosion current density be-
value up to 1500 eu. come constant, the stable pits growth and the surface
The pitting went into induction period when local corrosion products began to deposits. Based on
oxidized film breakage, the pit nucleation leaded to
the formation of a small area of bare, un-filmed sur-
face of metal. After a short of induction of about
1000 s, the pitting process began to transition from
metastable pit growth to stable pit growth, the
growth of a stable pit leaded to metal damage and
pit propagation. The acoustic emission activity and
intensity were still at a high level in stable pitting
stage, the pits propagation, hydrogen bubbles and
movement of corrosion products were considered as
the main acoustic emission sources. So the cluster
number K was set to 3 in this stage, the acoustic

Figure 10. Acoustic emission waveform and WT coefficients

diagram of hydrogen bubble activity. (a) waveform and (b) 2D
Figure 9. Acoustic emission hits’ amplitude distribution with WT diagram (Frequency resolution = 10 kHz, Wavelet size =
time 600)

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7 Cluster analysis of acoustic emission signals in pitting corrosion of low carbon steel 743

Figure 11. Acoustic emission waveform and WT coefficients Figure 12. Acoustic emission waveform and WT coefficients
diagram of oxide and corrosion products film breakage. (a) diagram of pit growth. (a) waveform and (b) 2D WT diagram
waveform and (b) 2D WT diagram (Frequency resolution = (Frequency resolution = 10 kHz, Wavelet size = 600)
10 kHz, Wavelet size = 600)

above analysis, it was inferred that Cluster 1 was tribution of WT coefficients, the energy was very
mainly produced by the hydrogen bubbles, Cluster 2 concentrated, the frequency bands corresponding to
was mainly generated by the oxidized film rupture at center energy was in the range of 130–180 kHz, and
the initial stage and corrosion products cracking at the peak frequency was about 165 kHz. The acoustic
stable pit growth stage, while Cluster 3 was only resonance frequency of a single bubble in water ne-
generated by pit propagation at stable stage. glecting the effects of surface tension and viscous at-
To further confirm the inference above, the Gabor tenuation given by Minnaert is
wavelet transform (WT) was presented to analyze sffiffiffiffiffiffiffiffiffi
the three cluster acoustic emission signals character- 1 3γp0
f0 ¼ ð11Þ
istics of time-frequency localization. WT coeffi- 2πr0 ρ0
cients as magnitudes over time and frequency were
shown in Figures 10–12, respectively. The color Where r0 is the bubble radius, P0 the ambient pres-
coding in the wavelet transform encodes the magni- sure, γ the ratio of specific heats for the gas, and ρ0
tudes of the WT, the colors usually are autoscaled: is the liquid density. However, A. Prateepasen and
the peak value of the transform corresponds to red his colleagues considered the maximum diameter of
and the minimum value corresponds to pink. The en- a stable bubble, Dmax and the collapse instantaneous
ergy features of signal in the time-frequency domain pressure, Pb they revised the Minnaert equation as
were extracted [22–25]. [3, 21, 26–31]
The bubble acoustic emission signal was similar
sffiffiffiffiffiffiffiffiffi
to the resonant signal, Figure 10. It was strongly re- 1 3γpb
lated to the hydrogen bubble oscillation, movement fb ¼ ð12Þ
and cavitation in solution. In the time-frequency dis- πDmax ρ0

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

744 H. Bi et al. Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7

The suggested value of Pb was 3.11 × 107 Pa.

The largest stable bubble diameter was about
0.9 mm captured by a high-speed camera. The bub-
bles were assumed to be filled with an ideal gas
with γ = 1.4 J kg–1 K–1), and liquid density was
ρ0 = 1.0 × 103 kg m–3. Therefore, the break-up fre-
quency of the maximum diameter of the stable
bubble was, fb ≥ 127.8 kHz, computed using
Eq. (12) that was well consistent with the fre-
quency range shown in Figure 10.
When compared with the bubble signals, the en-
ergy released from the oxide film rupture was higher
but more dispersive, and with frequency mainly be-
tween 90 and 200 kHz, Figure 1b. The film rupture
was a complex microscopic process, so acoustic Figure 13. Acoustic emission hits’ energy-duration distribu-
emission signal had rich frequency components, the tion
maximum energy corresponding time was at 65 μs,
which was relatively lagging behind. While pitting
growth signal energy distributions was little dis-
persed and the frequency was in the range of 95–
130 kHz the maximum energy appeared more ahead
at about 15 μs, Figure 12. The frequency characteris-
tics of hydrogen bubbles and pit growth signals were
consistent to the conclusion obtained by A. Pratee-
pasen in studying stainless steel pitting, which
further to confirmed the above inference of the three
acoustic emission sources [2, 6].

3.3 Back propagation artificial neural network

Some parallel experiments had been done to further
verify the accuracy of the three clusters signals cor-
responding to three different acoustic emission
sources in pitting corrosion. The acoustic emission
parameters A, E, D, C, RA value, AF value and RD
value were selected as seven inputs and three acous-
tic emission sources as outputs of back propagation
artificial neural network, and a 7-6-3 with a single
hidden layer back propagation network was estab-
lished [17, 32]. The data from classified acoustic
emission sources were used as training samples to
train the network. Then the acoustic emission data
collected from parallel experiments were classified
and recognized with trained neural network. The sa-
tisfactory results showed that the back propagation
artificial neural network can recognize different
acoustic emission sources successfully. The acoustic
emission energy-duration distribution of the identi-
fied signals was shown in Figure 13. Figure 14. Surface morphology of specimen after 3 hours po-
Figure 13 showed acoustic emission data from tentiostatic test in 3% NaCl solution at pH 2.0. (a) Surface mor-
parallel experiments were properly identified corre- phology (×100) and (b) 3D Surface topography of pits (×200)

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7 Cluster analysis of acoustic emission signals in pitting corrosion of low carbon steel 745

sponding to the three acoustic emission sources by dation of China (51301201) and the Key State
back propagation artificial neural network. The Science and Technology Projects of China
acoustic emission hits data was chosen from the last (2008ZX05017-004-1). We also wish to express our
one hour of potentiostatic polarization test stage, in gratitude to the Shengli Oilfield Technical Test Cen-
this stage acoustic emission signals were mainly ter, Sinopec for AMSY-5 device and the Germany
generated by pits propagation and corrosion pro- Vallen-Systeme GmbH for AGU-Vallen Wavelet
ducts cracking, while hydrogen bubble signals were software.
relatively less in this stage. The dominated pitting
signals were concentrated in low-duration, low-en-
ergy region, while the corrosion product cracking 5 Reference
signals energy was relatively dispersed, the duration
was higher than the former. The pitted surface was [1] M. Fregonese, L. Jaubert, Y. Cetre, Prog. Org.
observed by Zeiss Axio 3D Confocal Metallographic Coat. 2007, 59, 239.
Microscope, Figure 14. The 3D surface topography [2] A. Prateepasen, C. Jirarungsatian, Corrosion
characterized the morphology and the size of the 2011, 67, 225.
pits. A large number of honeycomb pits distributed [3] C. Jirarungsatian, A. Prateepasen, Corros. Sci.
on the pitted surface of specimen, and the pits depth 2010, 52, 187.
was on an average of 255 μm by the statistical analy- [4] C. Jomdecha, A. Prateepasen, P. Kaewtrakul-
sis, which further confirmed that the specimen sur- pong, NDT& E Int. 2007, 40, 584.
face was indeed undergoing severe pitting corrosion. [5] M. Fregonese, H. Idrissi, H. Mazille, L. Re-
naud, Y. Cetre, Corros. Sci. 2001, 43, 627.
[6] A. Prateepasen, C. Jirarungsatean, P. Tueng-
4 Conclusions sook, S. Lee, J. Lee, I. Park, S. Song, M. Choi,
Advanced Nondestructuve Evaluation I, Pts 1
This work aimed at classifying the acoustic emission and 2, Proceedings 2006, 321–323, 545.
signals generated by low carbon steel pitting corro-
[7] A. Prateepasen, P. Kaewtrakulpong, C. Jira-
sion, and the following conclusions can be obtained:
rungsatean, in Advanced Nondestructuve Eva-
During the pitting corrosion process of low carbon
luation I, Pts 1 and 2, Proceedings, Vol. 321–
steel in acidic NaCl solution, there are three main
323 (Eds.: S.S. Lee, J.H. Lee, I.K. Park, S.J.
kinds of acoustic emission sources: hydrogen bub-
Song, M.Y. Choi), 2006, pp. 549.
ble, oxidized film exfoliation and breakage, and pit
[8] M. Fregonese, H. Idrissi, H. Mazille, L. Re-
propagation. The K-means algorithm can be used to
naud, Y. Cetre, J. Mater. Sci. 2001, 36, 557.
distinguish the three typical acoustic emission
[9] F. Bellenger, H. Mazille, H. Idrissi, NDT& E
sources well.
Int. 2002, 35, 385.
Pit propagation, corrosion products exfoliation
[10] a) B. Assouli, A. Srhiri, H. Idrissi, NDT& E
and cracking are the main acoustic emission sources
Int. 2003, 36, 117; b) H. Chang, E. Han, J.
in the stable pit growth stage, which is consistent to
Wang, W. Ke, NDT& E Int. 2006, 39, 8; c)
the result of electrochemical polarization test.
V. Akop’yan, Y. Rozhkov, S. Shevtsov, Russ.
The three types of acoustic emission sources can
J. Nondestr. Test. 2007, 43, 390; d) S. Krako-
be recognized by back propagation artificial neural
wiak, K. Darowicki, J. Solid State Electr.
network, extracting the acoustic emission signal
2009, 13, 1653; e) M. Boinet, J. Bernard, M.
characteristics and further combined with the results
Chatenet, F. Dalard, S. Maximovitch, Electro-
of acoustic emission on-line storage tank floor in-
chim. Acta 2010, 55, 3454; f) G. Du, J. Li,
spection, it can help to interpret and evaluate the
W. Wang, C. Jiang, S. Song, Corros. Sci.
testing result and make the result more reliable.
2011, 53, 2918.
[11] A. Prateepasen, C. Jirarungsatean, P. Tueng-
Acknowledgement sook, in Advanced Nondestructive Evaluation
I, Pts 1 and 2, Proceedings, Vol. 321–323 (Ed.:
The authors greatly acknowledge the financial sup- S. S. L. J. H. P. I. K. S. S. J. C. M. Y. Lee),
port provided by the National Natural Science Foun- 2006, pp. 553.

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw

746 H. Bi et al. Mat.-wiss. u. Werkstofftech. 2015, 46, No. 7

[12] S. Park, S. Kitsukawa, K. Katoh, S. Yuyama, [23] A. Gallego, J. Gil, E. Castro, R. Piotrkowski,
H. Maruyama, K. Sekine, Mater. Trans. 2006, Surf. Coat. Tech. 2007, 201, 4743.
47, 1240. [24] R. Piotrkowski, E. Castro, A. Gallego, Mech.
[13] S. Park, S. Kitsukawa, K. Katoh, S. Yuyama, Syst. Signal Pr. 2009, 23, 432.
H. Maruyama, K. Sekine, Mater. Trans. 2005, [25] R. Gonzalez-Carrato, F. Marquez, V. Dimlaye,
46, 2490. D. Ruiz-Hernandez, Mech. Syst. Signal Pr.
[14] G. Dai, W. Li, L. Zhang, F.F. Long, Mater. 2014, 48, 339.
Eval. 2002, 60, 976–978. [26] M. Ashokkumar, Ultrason. Sonochem. 2011,
[15] A.V. Sokolkin, I.Y. Ievlev, S.O. Cholakh, Russ. 18, 864.
J. Nondestr. Test. 2002, 38, 113. [27] W. Kracht, J.A. Finch, J. Colloid Interf. Sci.
[16] A.V. Sokolkin, I.Y. Ievlev, S.O. Cholakh, Russ. 2009, 332, 237.
J. Nondestr. Test. 2002, 38, 902. [28] M. Devaud, T. Hocquet, J.-C. Bacri, V. Leroy,
[17] M. Riahi, H. Shamekh, B. Khosrowzadeh, Eur. J. Phys. 2008, 29, 1263.
Russ. J. Nondestr. Test. 2008, 44, 436. [29] T.G. Leighton, Ultrasonics 2008, 48, 85.
[18] S. Didier-Laurent, H. Idrissi, L. Roue, J. Power [30] S.J. Spencer, R. Bruniges, G. Roberts, V.
Sources 2008, 179, 412. Sharp, A. Catanzano, W.J. Bruckard, K.J. Da-
[19] M. Riahi, H. Shamekh, Russ. J. Nondestr. Test. vey, W. Zhang, Miner. Eng. 2012, 36–38, 21.
2006, 42, 537. [31] J.W.R. Boyd, J. Varley, Chem. Eng. Sci. 2001,
[20] L. Calabrese, G. Campanella, E. Proverbio, 56, 1749.
Constr. Build. Mater. 2012, 34, 362. [32] a) D. Crivelli, M. Guagliano, A. Monici, Com-
[21] J. Li, G. Du, C. Jiang, S. Jin, Anti-Corros. posites Part B-Engineering 2014, 56, 948; b)
Method M 2012, 59, 76. P. Ramasamy, S. Sampathkumar, Composites
[22] Y. Ding, R.L. Reuben, J.A. Steel, NDT& E Int. Part B-Engineering 2014, 60, 457.
2004, 37, 279.
Received in final form: December 1st 2014 T 347

© 2015 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.wiley-vch.de/home/muw