Applied Acoustics 99 (2015) 60–67

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

Perceptual study of the evolution of gear defects

Younes Ramdane a,⇑, Hamzaoui Nacer b, Ouelaa Nouredine a, Djebala Abderrazek a
a
Mechanics and Structures Laboratory, University of Guelma, B.P. 401, Guelma, Algeria
b
Laboratory of Vibration-Acoustics, INSA of Lyon, Bâtiment A. St. Exupéry, 25 bis Avenue Jean Capelle, 69621 Villeurbanne Cedex, France

a r t i c l e i n f o a b s t r a c t

Article history: The objective of this work is to apply the sound perception approach to study and diagnose gear defects.
Received 19 May 2014 Simple and multiple defects of different levels of severity are artificially simulated on the gear teeth. The
Received in revised form 5 May 2015 corresponding sounds are then acquired to perform a sound base representative of the diversity of gear
Accepted 26 May 2015
defects. Acoustic sounds are generated using the processing software DynamX V.7. These sounds are ana-
lyzed with the paired comparison method to find a correlation between the sound perception and the
scalar indicators. The results show that perception tests allow classifying gear defect sounds by order
Keywords:
of degradation. The relation between the vibratory indicators and sound perception enabled us to obtain
Gear defect
Sound perception
applicable mathematical models for the other sounds not included in the listening tests. These models
Paired comparison method can be used to monitor the evolution of gear degradation without repeating perceptions tests.
Fault detection Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction As vibration and acoustics have the same generation mecha-

Gears are used to transmit and change the rate of machinery
shaft rotation. They can also change the direction of the rotation
axis and rotary motion to linear motion, and they have existed
since the invention of rotating machinery. Because of their
force-multiplying properties, early engineers used them for
hoisting heavy loads such as building materials. However, when
dedicated to high-speed machinery such as automobile transmis-
sions, gears are the optimal technological solution for low energy
loss, high accuracy and low play. There are several categories of
gears and they can be combined in a multitude of ways.
Vibration based methods for gear fault analyses have been estab-
lished for a long time; they have been widely used for monitoring
rotating machine conditions and for fault diagnostics. Regarding
gearbox fault diagnosis, many signal processing tools have been
used recently, such as the time–frequency method [1], spectral anal-
ysis [2], cepstral analysis [3], empirical mode decomposition and the
Hilbert spectrum [4], and wavelet multi-resolution analysis [5].
In addition, noise signals, which are closely correlated to vibra-
tions, have also been actively investigated for gearbox condition
monitoring and fault diagnosis over the last two decades [6,7].
Nevertheless, both vibration and acoustic signals can be contami-
nated by different background noises. Therefore specific analysis
with more advanced tools should be carried out to obtain reliable
measures for fault diagnosis.
⇑ Corresponding author.
E-mail address: ramdane_ys@yahoo.com (R. Younes).
nism, acoustic noise can also be used for machinery condition
monitoring when combined with effective signal processing meth-
ods, such as the sound perception method. Perceptual tests can be
used to analyze various sound characteristics, and to discover the
perceptive dimensions used by listeners to differentiate these
audio stimuli. These measures of relative differences allow creating
a similarity matrix perceived between stimuli. The dimensions are
obtained from the similarity matrix using the Multi-Dimensional
Scaling (MDS) method [8], which allows representing the dissimi-
larity perceived by listeners between stimuli in a multidimensional
space. The paired comparison method is mainly recommended as
the standard method for making judgments of dissimilarity
directly between audio stimuli [9,10]. Auditors have to assess the
overall dissimilarity of pairs, for example, by recording their esti-
mate on a linear scale ranging from ‘‘very similar’’ to ‘‘very dissim-
ilar’’. Individual dissimilarity matrices are filled with auditor’s
estimates and lead to an average matrix adapted to the MDS.
The perceptive approach to sounds has been used in several
domains, for example in automobiles for the study of the sounds
produced by air conditioning systems, and in railways for the per-
ception of interior noises in a high-speed train to improve comfort
and satisfy consumers’ expectations [11,12]. Many researchers
have investigated the comparison of listening tests and explained
the benefits of the paired comparison method [13]. Other research-
ers have studied a serious problem in the paired comparison
method: that of the considerable duration of listening tests in
the case of a high number of stimuli, as it leads to auditive fatigue.
The search for a test allowing the evaluation of a high number of

http://dx.doi.org/10.1016/j.apacoust.2015.05.010
0003-682X/Ó 2015 Elsevier Ltd. All rights reserved.
 R. Younes et al. / Applied Acoustics 99 (2015) 60–67
Two gears
transmission
Variable Motor
speed drive
P4
P3
P2
P1
Brake
Accelerometer
Ball bearing
Fig. 1. Experimental setup.
stimuli led to the adaptation of a new method based on visual per-
ception, called comparison with reference method. The task con-
sists in assessing the similarity of a group of stimuli compared to
a Ref. [14].
Kanzari [15] carried out a vibro-acoustic analysis of sounds
resulting from gear teeth defects. From the results obtained, she
established a correlation with scalar indicators. These results
showed that scalar indicators such as vibratory speed, the spectral
center of gravity, and the peak to peak value clearly explain the
dissimilarity judgments for gear sounds in rotating machines.
Although the main advantage of using acoustic condition mon-
itoring is that acoustic signals can be acquired remotely, it is nec-
essary to avoid safety hazards and eliminate the need for high
temperature vibration sensors with their associated mounting
problems. However, acoustic condition monitoring used for
machinery has received very little attention. This is because the
acoustics from nearby rotating machines tend to contaminate the
airborne sound generated by the machine being monitored. The
influence of the acoustic environment definitely plays an impor-
tant role in machine condition monitoring, and these noise sources
must be filtered out by advanced signal processing techniques.
Also, regarding surface vibration measurement, sensors are sensi-
tive to the direction of the vibration source, whereas for acoustic
measurement the requirement for a microphone is that it should
be pointed toward the machine.
In the case of closed transmissions solid-borne noise is the
major component [16]. The vibration spectra produced by gear-
boxes appear to be complicated, but they can generally be broken
down into: tooth meshing harmonics, ghost components, side-
bands, low harmonics of shaft speed and intermodulation compo-
nents. For a well-meshed set of gears, only the fundamental gear
meshing frequencies are likely to be measured.
According to the literature study presented below, we found
very few works on fault detection by sound perception. The
Fig. 2. Photos of various defects.
61
objective of this work is to apply the sound perception approach
to the study and diagnosis of gear defects. The paired comparison
method is used for assessing various samples of gear defect sounds
measured in different cases to find a relationship between scalar
indicators and gear sounds.
By listening to vibratory signals of different defect configura-
tion, we hope to construct a linear correlation between the defect
aggravation and a combination of 2 scalar indicators used in mon-
itoring. The advantage will therefore be to construct a model of
defect detection with scalar indicators, from its birth to the aggra-
vation stage. To conceive this type of model in a new real case, it
will be necessary to constitute a signals’ base of faulty and healthy
configuration. This approach is rather similar to a method of artifi-
cial intelligence as neural networks, but does not require training
stage for regulating optimization parameters. The originality of this
article locates at the level of the contribution of subjective
approach by human audition as a supplement to the actual physi-
cal methods.
The paper is structured as follows. Section 2 is devoted to the
theoretical study of the MDS method. Section 3 contains the pre-
sentation of the experimental setup and the measurement chain.
Sound perception tests are presented in Section 4. Finally, the
results obtained from perceptive tests with the approach proposed
are presented and discussed in Section 5.
2. Multidimensional scaling method
The multidimensional scaling method allows the representation
of objects in a space starting from the proximity relationship exist-
ing between each couple of objects. Each object can then be repre-
sented in this space. A large number of algorithms can be used to
determine the coordinates of objects in space, starting from the
distances between them. Some of these algorithms allow taking
into account the specificities of the various subjects or groups of
subjects of the various stimuli, and even both of them
simultaneously.
Multidimensional analysis is an iterative reduction process in
which objects are placed in a space of several dimensions adjusted
to similarity or dissimilarity measures between objects. In this
space, the similar objects are placed close to each other, and the
non-similar items are removed. Multidimensional analysis is used
to find dimensional information common to a set of variables. The
dissimilarity scores provided by the subjects are translated into
distances in a comparison test of objects presented in pairs.
2.1. INDSCAL algorithm
The Multidimensional Scaling algorithms used in the literature
are: EXSCAL [17]. CLASCAL [18], expanded CLASCAL expanded [19],
62 R. Younes et al. / Applied Acoustics 99 (2015) 60–67

emergent properties of the objects that will later be useful for per-
ception. On the other hand, taking the weighting into account
allows the rapid determination of what the properties were of
the different objects used by each subject in the comparison step.

3. Experimental setup

3.1. Test rig

The test rig contains two spur gear transmissions lubricated

normally. The gears are made of stainless steel with a module
M = 2 mm and a face width F = 20 mm. The first transmission con-
sists of a pinion with 42 teeth (P1) mounted in the input shaft and
a wheel with 50 teeth (P2). The second transmission consists of a
wheel with 65 teeth (P3) and a pinion with 45 teeth (P4) mounted
in the output shaft. The test rig also contains a brake to simulate a
variable load, as shown in Fig. 1.

3.2. The experiments conducted

Fig. 3. Training phase interface. Simple and multiple defects of different severities were simu-
lated on the gear teeth using a mini grinder. The corresponding
sounds were then acquired to obtain a sound base representative
MDSCAL [20], and INDSCAL. In this article the INDSCAL algorithm is of gear defect diversity.
considered for the results processing. Regarding the sound measurements resulting from our test
The INDSCAL Algorithm (INdividual Differences SCALing) was device, the recording base was built using the vibration signals
developed by Caroll and Chang [21] and, unlike MDSCAL, it allots from the accelerometer which was saved in ‘‘⁄.wav’’ format with
a different scale for each subject. But like MDSCAL, it assumes that the acquisition and vibratory analysis software DynamX V.7. This
all the listeners use the same dimensions but not with the same is possible, since when transmissions are enclosed in a casing, as
weighting. These weightings appear in Eq. (1) in the form of a in the case of a gearbox, the noise radiated is mainly due to the
Wkr factor representing the weight that listener k allots to percep- vibrations of the latter [22].
tive dimension R of perceptive space.
" #12 4. Sound perception tests
X
R
dijk ¼ W kr ðX ir  X jr Þ2 ð1Þ
4.1. Stimuli
r¼1

where dijk is the distance between the objects i and j according to
subject k. X ir and X jr are the coordinates of these objects on dimen-
sion R of the perceptual space.
The INDSCAL measure generates varying axes between subjects.
On the one hand it simplifies the identification step of the
In this part, the objective is to study the ability of listeners to
identify different types of defect, ranging from a small single defect
to a combination of defects. To do this, we simulated several faults
with different degrees of degradation as follow:
 Healthy Gears HG.
 Small Defect on Pinion1 SDP1.
 Average Defect on Pinion 1 ADP1.

Measure of stress factor

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8
Dimensions number of the MDS

Fig. 4. Paired comparison interface test. Fig. 5. Measure of stress factor.

 R. Younes et al. / Applied Acoustics 99 (2015) 60–67 63

Nb val sing : 2 R : 0.93582 0.4

1.2
S1
0.3
1

0.2
0.8
similaritées reconstruites

Dimension 2
0.1 S6
S7
0.6
S8
0 S4
0.4

0.2 -0.1 S5

S3
0 -0.2

S2
-0.2 -0.3
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
Dimension 1
-0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fig. 7. Perceptual space.
similaritées mesurées

Fig. 6. Diagram of dispersion between measured and reconstructed similarities. 4.3. Tests interface and subjects

The test interface was programed in MATLAB language and con-

 Critical Defect on Pinion 1 CDP1.
tained two phases: the first was called training phase in which the
 Critical Defect on Pinion 1&Small Defect on Pinion 2 CDP1SDP2.
auditors were asked to listen to the sounds. The listeners could
 Critical Defect on Pinion 1&Average Defect on Pinion 2
become familiar with the sounds of the test in operation. The sec-
CDP1ADP2.
ond phase was a dissimilarity pair comparison phase. 30 listeners
 Critical Defect on Pinion 1& Critical Defect on Pinion 2
(07 women and 23 men), aged from 22 to 50 years old (average
CDP1CDP2.
age = 31, SD = 9.5), were asked to perform these two tests. Fig. 3
 Critical Defect on Pinion 1, Critical Defect on Pinion 2 & Critical
represents the training phase.
Defect on Pinion 3 CDP1CDP2CDP3.

Fig. 2 presents the photos of the various defects created on the 4.4. Dissimilarities evaluation
gears.
After each combination of simulated defects, the corresponding The paired comparison test is often used to evaluate the dissim-
signal is generated. Finally, a database is obtained that can be used ilarities between the sound samples [23–27]. This subjective test
to possible to perform several analyses. The recording database consists in playing all the pairs of possible sound samples to the
comprises 22 sounds for a given rotational speed. A preliminary lis- listeners. In general, to limit the number of pairs to be presented
tening was used to select the most different sounds between them,
after which a second selection was carried out to obtain 8 sounds
Measures
on which the paired comparison method was performed. 0.3 2
Dim1 = 0.088565 x PP+ -0.64616 with R = 0.93669

0.2
4.2. Sounds restitution
DIM1 obtained by listening test

The evaluation dissimilarity tests were carried out in an ordi- 0.1

nary room of the mechanics and structures laboratory (LMS) of
the University of Guelma, Algeria. The sounds were reproduced 0
through stereo headphones. The reproduction chain was composed
of a PC with I5 Dell processor, a test interface in the Matlab envi-
-0.1
ronment, and Sennheiser HD201 headphones.

-0.2
Table 1
Sounds coordinates in two dimensional perceptual space.
-0.3
Sounds Dim1 Dim2 Type of defect
Sound1 0.4714 0.3165 Without defect
-0.4
Sound 2 0.3281 0.2775 Simple defect
Sound 3 0.1394 0.1631
Sound 4 0.1073 0.0068 -0.5
Sound 5 0.1323 0.1010 Double defect -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
Sound 6 0.2158 0.1105
DIM1 calculated by Mathematecal model
Sound 7 0.1887 0.0956
Sound 8 0.2949 0.0259 Triple defect
Fig. 8. Diagram of dispersion between the first dimension and the scalar indicators.
64 R. Younes et al. / Applied Acoustics 99 (2015) 60–67

to the listeners, a random selection of the stimuli presentation With dij,k being the Euclidean distances and dij,k the dissimilarity
order is performed within each pair to cancel the effect of the stim- measures between conditions i and j for subject k. The multidimen-
uli presentation order according to the Ross series [28]. Moreover, sional scaling methods presented in this section are used when the
only a few pairs of identical stimuli are included to ensure the con- experimenter has no idea of the number of dimensions to be
sistency of listeners’ responses. The total number of pairs of stimuli considered in order to represent the perceptual space. In this case
is n (n  1)/2 pairs of sounds, where n is the number of conditions the number of dimensions is chosen by analyzing the error caused
to be tested. In practice, each auditor listens to two sounds of each by the given stress value. In practice, the optimum number of
pair as many times as desired. Then, he (she) is asked to move a dimensions is determined when the adding a dimension gives little
cursor between ‘‘identical’’ to the left and ‘‘very different’’ to the additional information for the reconstruction of the distance
right, depending on the degree of dissimilarity between the sounds matrix. The stress curve in Fig. 5 obtained during our tests
perceived. The cursor position for a pair of sounds is translated into shows the emergence of a two-dimensional elbow. Thus the
a dissimilarity value from 0 to 1. This step takes about 10–15 min. two-dimensional solution was chosen.
Fig. 4 shows the comparison test interface as it appears to the
listener.
5.3. Calculation of dissimilarities between subjects: correlation
5. Results and discussion coefficient

5.1. Multidimensional analysis The differences between subjects are calculated from their
responses for different pairs of stimuli. The dissimilarity measure-
Multidimensional scaling analysis is a technique used to repre- ment adopted depends on the type of response.
sent an ensemble of similarities between items as points in a The dissimilarity between the two subjects’ k and l is defined as
Euclidean space. In our case the items are sounds, and the proxim- follows [30]:
ity between the sounds is evaluated using dissimilarity judgments.
The principle of this analysis is to match the measured dissimilar- d1 ðk; lÞ ¼ 1  rðk; lÞ ð3Þ
ities between the distances of stimuli with points representing
these stimuli in a Euclidean space. The dimensions of this space where r(k, l) is the Bravais–Pearson linear correlation coefficient
are continuous, i.e. all stimuli may be classified according to these [31] between the dissimilarity scores given by subject k and those
dimensions. pronounced by subject l.
The similarity score given by the Bravais–Pearson coefficient is
5.2. Determination of the number of dimensions R = 0.91 (Fig. 6) confirms good dissimilarity restitution.
In Table 1 we present the coordinates of the simulated sounds
The optimal number of dimensions for representing the percep- in two dimensional space (DIM1 and DIM2), where DIM1 repre-
tual space is determined from the error value in the reconstruction sents the evolution of the degradation, and DIM2 represents the
of the dissimilarities in Euclidean distances. This error is deter- difference in severity of each sound compared to other sounds
mined by the stress value which is generally represented by the according to the perceptive approach test (see Fig. 7). It is notewor-
number of dimensions, as shown in Eq. (2) [29]. thy that the coordinate results are in good agreement with the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P degrees of gear deterioration, namely without defect S1, single
2
ði;jÞ2X ðdij;k  dij;k Þ (S2, S3), double (S5, S6, S7) and triple defects (S8). Nevertheless,
Stress ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ð2Þ it should be noted that the critical defect (S4) was perceived as
2
ði;jÞ2X dij;k being closer the double defects than the simple defects.

0.4 0.3
Measures Measures
Dim1 = 1.5981 x OL +0.013647 x K -0.7405 with R2 = 0.97767 Dim2 = 1.4118 x OL + 0.0026603 x SCG -1.8297 with R 2 = 0.82752
0.3
0.2
DIM2 obtained by listening test

0.2
DIM1 obtained by listening test

0.1 0.1

0
0
-0.1

-0.1
-0.2

-0.3
-0.2

-0.4

-0.3
-0.5 -0.2 -0.1 0 0.1 0.2 0.3
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 DIM2 calculated by Mathematecal model
DIM1 calculated by Mathematecal model
Fig. 10. Diagram of dispersion between the second dimension and the scalar
Fig. 9. Diagram of dispersion between the first dimension and the scalar indicators. indicators.
 R. Younes et al. / Applied Acoustics 99 (2015) 60–67 65

Table 2
The coefficients R2 and their mathematical model.

DIM1 DIM2
2
R 0.98 Dim1 = 1.5981 ⁄ OL + 0.0136 ⁄ K-0.7405 R2 0.83 Dim2 = 1.4118 ⁄ OL + 0.0026 ⁄ SCG-1.82
0.96 Dim1 = 0.05988 ⁄ PP + 0.70222 ⁄ OL-0.64 0.73 Dim2 = 0.18536 ⁄ KF + 0.0023 ⁄ SCG-1.49
0.95 Dim1 = 2.0511 ⁄ OL-6.7359 ⁄ 105 ⁄ RMS-0.10 0.59 Dim2 = 0.0501 ⁄ PP + 0.0022 ⁄ SCG-1.57
0.95 Dim1 = 2.0038 ⁄ OL + 0.0326 ⁄ CF-1.0297 0.42 Dim2 = 0.0010 ⁄ SCG + 5.61  105 ⁄ RMS-0.97

0.4 which presents the best similarity score, thus a better alignment of
cloud points on the regression line.
S1
The scalar indicators are calculated from the vibratory acceler-
0.3 M1 ation signals, thus their unit is m/s2, except the kurtosis and the
crest factor which are dimensionless, and the CGS in Hz. The
K-factor is defined as the product between the peak value and
0.2
the RMS. The overall level of vibration of a machine is a measure
of the total vibration amplitude over a wide range of frequencies.
Dimension 2

0.1 S7
S6 M8 In Fig. 8 this correlation is shown for Dimension1 (DIM1), with
M7 one physical parameter. In this case Dimension1 is best character-
M6
S8 ized by the Peak-to-Peak, with a coefficient of determination
0 S4 R2 = 0.93 (P < 0.001). The linear combination obtained is given by:
M5
DIM1 ¼ 0:088565  PP  0:64616 ð4Þ
-0.1 M4 S5 Fig. 9 shows the correlation obtained for DIM1 with two phys-
S3
ical parameters. The presence of OL and K in the mathematical
M2
-0.2
model (5) confirms the previous DIM1 analysis, since these two
M3
indicators are directly related to the defect degradation level. The
coefficient of determination is R2 = 0.97 (P < 0.001).
S2
-0.3
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 DIM1 ¼ 1:5981  OL þ 0:013647  K  0:7405 ð5Þ
Dimension 1 Finally, Fig. 10 presents the correlation obtained for DIM2 with
two physical parameters. The model of this dimension given by
Fig. 11. Sounds classed by listeners (Si), and those calculated by mathematical
models (Mi).
expression (6) is characterized by the OL and the SCG which is
related to the spectral balance. The coefficient of determination
is R2 = 0.82 (P < 0.002).
5.4. Correlations between scalar indicators and dimensions
DIM2 ¼ 1:4118  OL þ 0:0026603  SCG  1:8297 ð6Þ
2
Since these dimensions perfectly describe the sound dispersion We present in Table 2 the value of the coefficients R and their
in the proximity space, and in order to link the vibration indicators mathematical model; we chose the model that has the largest coef-
to acoustic perception, it is necessary to find a correlation between ficient value.
this dispersion and the scalar indicators calculated: Kurtosis (K),
Crest Factor (CF), Spectral Center of Gravity (SCG), Root mean 5.5. Application of the mathematical correlation models
square (RMS), Peak-to-Peak (PP), overall level (OL), the peak value
and the K-factor. Thus the dimensions of this space will be a linear Fig. 11 presents, in the sound proximity space, the classification
function of different indicators. coordinate values of the 8 sounds obtained by the perception tests
To this end, an ascending linear regression was performed using (Si) and the two models (Mi) obtained previously from Eqs. (5) and
vibratory indicators as inputs. The dimension selected will be that (6). This figure shows good agreement, therefore the models

Fig. 12. Kurtosis of the 22 sounds.

66 R. Younes et al. / Applied Acoustics 99 (2015) 60–67

0.7
Mo5
Zone1 Zone2 Zone3 Zone4
0.6
Mo4

0.5
Mo2
0.4
S1 Mo9
Mo1
Dimension 2 0.3 Mo3

0.2 S8
S7Mo15 Mo22
0.1
Mo21
Mo13 Mo17
Mo18
0 Mo8 Mo19 S6
S5
Mo14 Mo16
Mo20
-0.1 Mo12Mo11
S2 S4
Mo6 S3
-0.2 Mo7
Mo10

-0.3
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
Dimension 1

Fig. 13. Perceptual space for the 22 sounds.

Zone I: The sounds Mo1 and Mo2 represent the state without
defect, measured in two different time intervals.
Zone II: The sounds Mo3–Mo10 represent the simple defects,
with increasing degrees of severity: Mo3–Mo5 small defect,
Mo6–Mo9 average defect, and Mo10 critical defect.
Zone III: Includes the very critical simple defects Mo11–Mo13
(tooth removal) and the double defects Mo14–Mo21 with vari-
able severity levels. It is noteworthy that the sounds Mo11–
Mo13 were classed by the mathematical model in the double
defect zone, such as sound S4 which was classed by the listen-
ers in the same zone (see Fig. 6).
Zone IV: Sound Mo22 represents the triple defect.

Finally, we can conclude from this study that the mathematical

Fig. 14. DIM1 of the 22 sounds.
obtained are reliable with a coefficient R2 = 0.97 for DIM1 and
R2 = 0.82 for DIM2.
Based on previous studies of bearing and gear defects [32,33], it
has been shown that kurtosis is the indicator most sensitive to
shocks. Fig. 12 shows the kurtosis of the 22 sounds. It was found
that the kurtosis values increased as the defect worsened, reaching
its maximum value for sounds 11 and 12 corresponding to a large
defect (removal of approximately 3=4 of the width of the tooth). The
combination with other defects on the intermediate and the output
wheel led to decrease of kurtosis. It became less sensitive due to
the increase in the number of shocks in the signal which tended
to merge into each other.
correlation models obtained for dimensions 1 and 2 can be used as
a tool for monitoring the evolution of gear degradation without
repeating perceptions tests.
Fig. 14 shows the evolution of DIM1 for the 22 sounds calcu-
lated by expression (5) of the mathematical model. If there is no
fault (sounds 1 and 2), the DIM1 values are very low. They increase
with the progression of the fault on the wheel 1 (sounds 3–10). For
sounds 11 and 12 corresponding to a very critical defect, listeners
perceive them as a double fault (see Fig. 11), DIM1 turns positive.
For sound 13 corresponding to the total tearing away of the tooth,
thus no shock, DIM 1 significantly decreases and reaches almost
zero. The combination of two to three defects on different wheels
gives DIM1positive values.
6. Conclusion

5.6. Application of correlation models for all sounds
To check the reliability of these models for the measured
sounds not subjected to the perception tests, we present in
Fig. 13 the coordinate values of all the 22 sounds calculated by
the expressions of DIM1 and DIM2 based on the most revealing
scalar indicators of the gear degradation state. In the calculation,
the scalar indicators values which best explain the dissimilarity
judgments (OL, K, and SCG) are used.
It should be noted that according to the proximity space, the 22
sounds are well classified into four zones in ascending order of
defect severity.
The aim of this work was to study the evolution of gear defects
using the sound perception approach. To do this, several defects
were simulated on gears assembled on a laboratory test rig. A per-
ceptual study based on the paired comparison and the MDS meth-
ods was carried out to study gear sounds for different degradation
levels. This study allowed identifying the vibration parameters for
establishing a vibro-acoustic relationship in rotating machinery,
namely gear transmission systems.
This work studied the capacity of the listeners to identify differ-
ent types of defect, ranging from a small simple defect to the com-
bination of several defects. To do this, several defects with various
levels of degradation were simulated.
 R. Younes et al. / Applied Acoustics 99 (2015) 60–67
According to the results obtained in this work we showed that:
 Sound coordinates in the proximity space are in good agree-
ment with the levels of gear degradation.
 The scalar indicators: Spectral Center of Gravity, Overall Level
and Kurtosis clearly explain the dissimilarity judgments for
the gear sounds in the case of a simulated defect.
 The correlations between the objective and subjective aspects
helped to highlight the strong relationship between the vibra-
tion indicators and the distances between gear sounds in prox-
imity space.
 The relation between the vibratory indicators and sound per-
ception enabled us to obtain applicable mathematical models
for the other sounds that were not part of the listening tests.
 The mathematical correlation models obtained for dimensions 1
and 2 can be used as a tool to monitor the evolution of gears
degradation without repeating perceptions tests.
References
[1] Oehlmann H, Brie D, Tomczak M, Richard A. A method for analysing gearbox
faults using time–frequency representations. Mech Syst Signal Process
1997;11(4):529–45.
[2] Ruiz BJA, Lopez Lopez JF, Quintero Riaza HF. Vibration analysis in gear damage:
mathematical modeling and experimental validation. In: 2013 XVIII
symposium of image, signal processing, and artificial vision (STSIVA); 2013.
p. 1–5.
[3] Cousinard O, Rousseau P, Bolaers F, Marconnet P. Paramétrage utilisation et
apport de l’analyse cepstrale en maintenance prévisionnelle. Méc Ind
2004;5:393–406.
[4] Liu B, Riemenschneider S, Xu Y. Gearbox fault diagnosis using empirical mode
decomposition and Hilbert spectrum. Mech Syst Signal Process
2006;20(3):718–34.
[5] Djebala A, Ouelaa N, Benchaabane C, Laefer DF. Application of the wavelet
multi-resolution analysis and Hilbert transform for the prediction of gear tooth
defects. Meccanica 2012. http://dx.doi.org/10.1007/s11012-012-9538-1.
[6] Derouiche A, Hamzaoui N, Boukharouba T. Reconstruction of sound sources of
gear transmission mechanism by planar near field acoustical holography. In:
Proceedings of the second international conference ‘‘condition monitoring of
machinery in non-stationary operations’’ CMMNO’ 2012. p. 247–55.
[7] Baydar N, Ball A. Detection of gear failures via vibration and acoustic signals
using wavelet transform. Mech Syst Signal Process 2003;17:787–804.
[8] Borg I, Groenen P. Modern multidimensional scaling, theory and
applications. Springer; 1997.
[9] International Electrotechnical Commission, Rapport technique IEC 60268 –
Part 13: Listening tests on loudspeakers; 1998.
[10] Lipshitz SP, Vanderkooy J. The great debate: subjective evaluation. J Audio Eng
Soc 1981;29(7–8):482–91.
67
[11] Vincent H. Etude de la qualité sonore d’appareils de soufflage et de
climatisation. M.Sc Thesis. Paris: Pierre & Marie Curie University; 2005.
[12] Parizet E, Hamzaoui N, Jacquemoud J. Noise assessment in a high-speed train.
Appl Acoust 2002;63:1109–24.
[13] Parizet E, Hamzaoui N, Sabatie G. Comparison of some listening test methods:
a case study. Acta Acust United Acust 2005;91:356–64.
[14] Michaud PY, Meunier S, Herzog P, Aubigny GD, Lavandier M. Méthode de test
adaptée à l’évaluation perceptive d’un grand nombre de stimuli audio
Application aux enceintes acoustiques. In: 10th French congress of acoustics,
CFA2010. Lyon, France.
[15] Kanzari M. Vibroacoustic diagnosis of gears defects: sound perception
approach analysis. M.Sc Thesis. France: INSA of Lyon; 2009.
[16] Reboul E. Vibroacoustique of high frequencies mechanisms: application to
gears transmissions. Ph.D. Thesis. France: Central School of Lyon; 2005.
[17] Winsberg S, Carroll JD. A quasi-nonmetric method to multidimensional scaling
via an extended euclidean model. Psychometrika 1989;54:217–29.
[18] Winsberg S, De Soete G. A latent class approach to fitting the weighted
euclidean model, CLASCAL. Psychometrika 1993;58(2):315–30.
[19] Koehl V. Influence of structure dispersions on sound perception. Ph.D. Thesis.
France: INSA of Lyon; 2005.
[20] Torgerson WS. Multidimensional scaling: I. theory and method. Psychometrika
1952;17(4):401–19.
[21] Carroll JD, Chang JJ. Analysis of individual differences in multidimensional
scaling via an n-way generalization of ‘‘eckart-young’’ decomposition.
Psychometrika 1970;35:283–319.
[22] Reboul E. Vibroacoustique of high frequencies mechanisms: application to
gears transmissions. Ph.D. Thesis. France: Central School of Lyon; 2005.
[23] Susini P, Adams M, Winsberg S. A multidimensional technique for sound
quality assessment. Acta Acust United Acust 1999;85:650–6.
[24] McDermott BJ. Multidimensional analyses of circuit quality judgments. J
Acoust Soc Am 1969;45(3):774–81.
[25] Suzuki T, Yasui S, Ojima Y. Evaluating adaptive paired comparison
experiments. In: Lenz H-J, Wilrich P-T, Schmid W, editors. Frontiers in
statistical quality control 9. HD: Physica-Verlag; 2010. p. 341–50.
[26] Mattila VV. Ideal point modeling of speech quality in mobile communications
based on multidimensional scaling (MDS). J Audio Eng Soc 2002;112:1–14.
[27] Wältermann M, Scholz K, Möller S, Huo L, Raake A, Heute U. An instrumental
measure for end-to-end speech transmission quality based on perceptual
dimensions: framework and realization. Interspeech 2008, Brisbane. Australia.
[28] David HA. The method of paired comparison. New-York: Oxford University
Press; 1988.
[29] Leman A. Diagnosis and automatic evaluation of vocal quality starting from
hybrid indicators (desqhi model). Ph.D. Thesis. France: INSA of Lyon; 2011.
[30] Saporta G. Probabilité, analyse des données et statistique. Paris,
France: Technip Edition; 1990.
[31] Stemler SE. A comparison of consensus, consistency, and measurement
approaches to estimating interrater reliability. Pract Assess Res Eval
2004;9(4):1–19.
[32] Djebala A. Application of the wavelet transform to the study and the vibratory
analysis of the mechanical systems. Ph.D. Thesis. Annaba, Algeria: Badji
Mokhtar University; 2008.
[33] Benchaabane C, Djebala A, Ouelaa N. Paramètres d’identification des défauts
d’engrenages par analyse vibratoire. In: Third international congress design
and modeling of mechanical systems, Tunisie; 16–18 Mars 2009.