Author’s Accepted Manuscript

Automatic condition monitoring system for crack

detection in rotating machinery

M.J. Gómez, C. Castejón, J.C. García-Prada

www.elsevier.com/locate/ress

PII: S0951-8320(16)30003-5
DOI: http://dx.doi.org/10.1016/j.ress.2016.03.013
Reference: RESS5522
To appear in: Reliability Engineering and System Safety
Received date: 1 January 2016
Revised date: 15 March 2016
Accepted date: 22 March 2016
Cite this article as: M.J. Gómez, C. Castejón and J.C. García-Prada, Automatic
condition monitoring system for crack detection in rotating machinery, Reliability
Engineering and System Safety, http://dx.doi.org/10.1016/j.ress.2016.03.013
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
 Automatic condition monitoring system for
crack detection in rotating machinery

M.J. Gómez ∗ , C. Castejón, J.C. Garcı́a-Prada

MAQLAB Group, Mechanical Dept. University Carlos III of Madrid, Av. de la
Universidad, 30, 28911, Leganés, Madrid, Spain

Abstract

Maintenance is essential to prevent catastrophic failures in rotating machinery. A

crack can cause a failure with costly processes of reparation, especially in a rotating
shaft.
In this study, the Wavelet Packets transform energy combined with Artificial
Neural Networks with Radial Basis Function architecture (RBF-ANN) are applied
to vibration signals to detect cracks in a rotating shaft. Data were obtained from a
rig where the shaft rotates under its own weight, at steady state at different crack
conditions. Nine defect conditions were induced in the shaft (with depths from 4%
to 50% of the shaft diameter). The parameters for Wavelet Packets transform and
RBF-ANN are selected to optimize its success rates results. Moreover, ‘Probability
of Detection’ curves were calculated showing probabilities of detection close to 100%
of the cases tested from the smallest crack size with a 1,77% of false alarms.

Key words:
Cracked shaft detection, Wavelet transform, Intelligent Classification Systems,
Condition monitoring, Artificial Neural Networks

1 Introduction

The main objective of condition monitoring of rotating machinery is to detect

faults before a catastrophic failure occurs. Besides, detection must arrive early
enough to have time for programming a stop at the most convenient moment.
This kind of maintenance has a lot of advantages, such as the avoidance of
stopping and dismount the elements of the machine to check its status, and
∗ Corresponding author: [email protected]

Preprint submitted to Reliability Engineering & System Safety 1 April 2016

the increasing of probability of fault detection. Condition monitoring improves
safety and costs of processes, reason why it has focused a lot of attention in
lasts decades.

In an industrial process, when a defect is detected, three main stages are passed
trough; the first one is detection, the second one is diagnosis and the third one
is intervention to correct undesired effects. Condition monitoring is usually
based on the control of certain parameter that is considered fault indicator,
and when its value exceed certain limits it is considered that a fault exists.
Nowadays, vibration analysis is the most extended technique for condition
monitoring [1].

A vibration signal obtained from a rotating machine has a complex structure

and comprises a high number of data. The key of the process is the feature
extraction. Feature extraction consist on finding the proper processing for a
vibration signal to get parameters easier to handle, and also representative
of the real status of the machine. Features must content reliable information
about the fault.

After the features extraction, a classification system is needed to automate the

diagnosis process. The classification system automatically decides the output
of the system avoiding human factors. The process of this type of methods is
schemed in figure 1

Input Signal processing Classification Output

(Vibration signal) (feature extraction) system (Diagnosis)

Fig. 1. Scheme of current processes of condition monitoring

Many fault diagnosis algorithms have been proposed both for features extrac-
tion and for classification.

Regarding feature extraction, several approaches have been used for signal
analysis. Fourier theory (FT) and techniques derived from it, as Fast Fourier
transform (FFT) and Hilbert transform (HT), have traditionally been used
to observe changes in the response when a fault appears [2,3]. However, FT
and most techniques derived from it are inappropriate to treat non stationary
signals, that are commonly obtained from rotating machinery, due to the ab-
sence of temporary information. The short time Fourier transform (STFT) is
suitable to treat non stationary signals, however its main disadvantage is that
the frequency resolution obtained is constant for the whole signal, since the
window applied is the same. Therefore, new techniques working both in time
and in frequency domain have appeared, such as Hilbert-Huang Transform
(HHT) [4] and Wavelet Transform (WT) [5].

2
Specifically, WT is a especially effective tool in treating non-stationary signals
and has become one of the most widely used techniques for signal processing.
Currently, applications of the WT are increasing, and they are now used for
speech recognition [6,7]; denoising [8]; electrocardiographs [9]; and diagnosis of
cracked rotating elements [10] as bearings [11], gears [12], and beams [13,14].

Regarding classification systems, these include artificial neural networks (ANN),

fuzzy logic (FL), genetic algorithms (GA), bayesian classifier (BC) and support
vector machines (SVM), among others. Specifically ANNs have been widely
used for different applications, such as residual life estimation of mechanical
elements [15], degradation prediction [16], diagnose accident scenarios as in
[17] or [18], diagnosis of cracked shafts detection [19,20].

The main advantage of ANNs is their flexibility and capability of learning

complex nonlinear relationships between input and output. The critical stage
for ANNs is to choose the training parameters used to learn the structure of
the problem [21]. On the other hand, the main drawback of the use of ANNs is
the need of a large amount of historical data, not only extracted from healthy
condition, but also from faulty condition, to facilitate the proper training of
the system [22].

Fault detection is specially critical for shafts, due to the high loads they sup-
port. The dynamical behavior of a cracked rotor has focused a lot of interest
among researchers [23–26]. A crack in a shaft can cause a failure with costly
processes of reparation. Cracked shaft detection has used combination of WT
and ANN for signals coming from models in cases such as [27,28]

The present work details an integrated system for maintenance based on the
combination of energy features and a trained ANN. The feature extraction is
calculated from a vibration signal by means of the Wavelet Packets Transform
(WPT). The technique is applied to experimental signals, showing that the
reliability of the technique is high.

2 Wavelet Packets transform

The WT is especially efficient at performing local analysis of non-stationary

signals. The discrete wavelet transform (DWT) is more commonly used than
the continuous form of the WT because signals usually comprise discrete data
and the computational cost is lower. Following Mallat definitions [29], the
DWT can be implemented using filters. The signal is decomposed being passed
through a low-pass filter g to obtain the wavelet approximation (A), and a high
pass filter h to obtain the wavelet detail (D), as in figure 2. DWT decomposi-
tion halves the frequency band of the input; thus, following the Nyquist rule,

3
it is justified to downsample by two [30].

High pass filter

D

f=[π/2-π]
N/2 points

f=[0-π]
N samples X

f=[0-π/2]
N/2 points
Low pass filter
A

Fig. 2. DWT decomposition of a signal (S) in approximation information (A) and

detail information (D) using filters

Several types of analysis are performed by the recursive application of the

DWT, such as Multiresolution Analysis (MRA) and Wavelet Packets Trans-
form (WPT). The wavelet packets transform (WPT) was used to prevent
problems associated to MRA, where the downsampling process can only be
performed for the A information. WPT allows A and D information to be
decomposed recursively until the desired resolution, as shown in figure 3.

W (k, j) represents coefficients of the signal in each packet, k represents the

decomposition level, and j represents the position of the packet within the
decomposition level. Each correlation vector W (k, j) has the structure given
in Eq.1:

W (k, j) = {w1 (k, j), ..., wN (k, j)} = {wi (k, j)} (1)

The calculus of energy using WPT is similar to the used in the FFT [5]. The
energy of a certain packet j within the decomposition level k can be obtained

W(0,0)

W(1,0) W(1,1)

W(2,0) W(2,1) W(2,2) W(2,3)

W(3,0) W(3,1) W(3,2) W(3,3) W(3,4) W(3,5) W(3,6) W(3,7)

Fig. 3. WPT decomposition process until decomposition level 3

4
as the sum of the squares of its coefficients, as in eq. 2:

Ek,j = {wi (k, j)}2 (2)
i

3 ANN classification systems

Currently, ANNs focus great interest both at academic and industrial level
[31]. ANNs are an example of learning and automatic processing inspired in
the way that the human nervous system works. It comprises an interconnecting
system of neurons that collaborate to produce an output. ANNs are widely
used as an effective, cost-effective and efficient automated indicator of health
of modern engineering systems [32].

There are two main phases in the use of an ANN; the phase of learning or
training, and the phase of validation and use. For the most common case of
supervised learning, the stage of training consists on the presentation of a set
of input patterns, which output is known. The network adjusts the connections
in an iterative process of error minimization, until it gets the admissible level
of error. However, the accuracy of ANNs depends upon the features selected,
ANN architecture, training parameters and training set size.

The radial basis function (RBF) architecture constitutes one of the more
widely used ANN in the diagnosis of defects area [33]. It was created with
the main purpose of working on real time applications. The name of radial
basis function derives from the fact that the function is symmetric; the output
is the same for inputs that are at the same distance of the center.

RBF networks are constituted by at least three layers of neurons; one at the
input, one hidden and one at the output, as shown in figure 4. These layers are
characterized by having local character, each neuron activates in a different
region of the space of the input patterns.

The use of RBF architecture offers a lot of advantages such as a fast train-
ing, and easy optimization due to the low number of design parameters [34].
The design parameters of an ANN-RBF are the activation function and the
stopping criteria for the training. The most common used activation function
is the Gaussian, thus the spread of the function must be selected. The most
common stopping criteria is a goal sum-squared error (SSE) between the de-
sired output and the real output. The ANN creates one neuron in the hidden
layer at each iteration. When SSE falls beneath the goal error or a maximum
number of neurons (to be selected) has been reached the training stops.

5
 Hidden layer
Input layer Output layer

Fig. 4. ANN basic structure

4 Experimental Setup

The experimental measurements are obtained from a machine simulation fault

created by SpectraQuest that can be observed in figure 5.

Fig. 5. Rig used for experimental setup

The rig comprises a motor M arathon , with maximum speed of 10.000 r.p.m
and a power of 0, 75KW , that drives the shaft by means of an elastic coupling.
The shaft rotates with the load of its own weight. The speed of the motor is set
using a regulator S1 Delta S1, and controlled using an optical tachometer
Banner . The shaft is supported by two ball bearings ER10 Rexnord .

The tested element is the shaft, under different crack conditions. A first test is
made at healthy state, and then nine different crack levels (a) were induced by
saw cuts. All the cracks were induced without dismounting the shaft from the
machine, because usually cracks appear and grow while the machine rotates

6
and the assembly effects are constant. The values of a in table 1 are expressed
as the ratio between the crack depth d and shaft diameter D, where D =
16mm.
Defect level 0 1 2 3 4 5 6 7 8 9
Value (a = d/D) 0 0.04 0.08 0.12 0.17 0.22 0.28 0.33 0.42 0.5
Table 1
Crack depths a used for the experimental setup, expressed in relative terms with
respect to the diameter of the shaft D.

Test are carried out at steady state at three different rotational speeds, shown
in table 2. Thus, 30 different conditions are tested; 10 different crack conditions
and 3 rotation speeds.
Speed Value (Hz)

Speed 1 20
Speed 2 40
Speed 3 60
Table 2
Rotational speeds used for the tests
For all cases, the cracks are located in the middle section, as shown in figure
6.

(a)

(b)
Fig. 6. Shaft and position of crack induced (a) and detail of an induced crack level
2
Shaft properties were determined experimentally and are shown in table 3.

The accelerometer used is 4374 B&K. The measurement direction is vertical

and is mounted on the housing bearing. The accelerometer is connected to a
signal conditioner B&K NEXUS, and to an acquisition card Keithley KUSB-
100. The acquisition card is connected to a computer with the software called
Btool, that was specifically designed for this purpose in Matlab[35]. The
measurement chain can be observed in figure 7

7
 Mass (M) (kg) 0,378

Damping coefficient (c) (kg/s) 4,58

Stiffness (k) (KN/m) 1544,3

Rho (ρ) (m) 2e-5

Effective length (m) 0,26

Diameter (m) 0,0207

Table 3
Shaft mechanical properties

Fig. 7. Measurement chain

The characteristics of the signals measured and number of signals are shown
in table 4.
Parameter Value

Sampling frequency fs 6 KHz

Number of samples / signal 214
Number of signals /condition 1500
Table 4
Parameters of signals measured

5 Results and Discussion

First, the features extraction stage is carried out by means of WPT energy.
Later, several ANNs are trained trying to optimize their parameters to maxi-
mize success rates and minimize computational cost.

5.1 Feature Extraction and selection using WPT

The energy calculated using WPT is selected for feature extraction due to
its proven effectiveness to process non-stationary signals for diagnosis purpose
[36,34]. The WPT is not a straightforward issue since some parameters must
be selected.

8
The first one is the wavelet function. In this study, the ‘Daubechies 6’ wavelet
function was applied due to the goodness of the results in this area [27]. The
other parameter to be selected is the decomposition level, which determines
the frequency resolution offered by each packet (same for all packets). Using
a decomposition level of k, the number of packets obtained is P = 2k . By
considering the global frequency of the signals as half of the sampling frequency
fs /2 according to the Nyquist theorem, the frequency resolution fr of each
packet is given by Eq.3 [34]:

fs /2
fr = (3)
2k

In this study, the selection of the decomposition level is performed according

to the method proposed in [28]. The method determines the optimal the de-
composition level that maximizes the success rates of an ANN. WPT energies
from 2 (P = 4packets; fr = 1, 500Hz) to 9 (P = 512packets; fr = 5.86Hz)
are evaluated in the present work. Features considered are single level.

After preliminary studies, it was concluded that using all the packets of a
single level for decomposition levels from 6 to 9 the computational cost is high.
Therefore, for decomposition levels from 6 to 9, the packets that experiment
higher changes of energy when a crack appears are selected. Mean values of the
1,500 signals measured by each condition are obtained. For each case of speed,
the mean value of the healthy condition is subtracted to the mean value of
each crack condition. The 10 packets that show higher differences in absolute
value for each crack size are selected, according to figure 8.

Since 9 crack levels are tested, a number of packets between 10 and 90 could be
selected for each speed, depending on the number of packets that are coincident
for different crack sizes.

An analysis of the frequencies that present higher changes of energy with the
crack is performed in [37], and they seem to be related to structural frequencies
of the shaft.

Thus, input packets for the ANNs training are selected according to table 5.

Decomposition level Features selected

Low: from 2 to 5 All packets single level

High: from 6 to 9 10 maximum energy increments for each crack size
Table 5
Features selected for each decomposition level

9
 Crack level 1 Crack level 2 Crack level 9
0.7 0.7 0.18

0.16
0.6 0.6

0.14
0.5 0.5
0.12

0.4 0.4 0.1

0.3 0.3 ....... 0.08

0.06
0.2 0.2
0.04

0.1 0.1
0.02

0 0 0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500

(-)
(-) (-)
0.7
Healthy condition Eje sano

0.6

0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500

|Diff 1| |Diff 2| ......... |Diff 9|

0.4 0.8 0.8

0.35 0.7 0.7

0.3 0.6 0.6

0.25 0.5 0.5

0.2 0.4 0.4

0.15 0.3 0.3

0.1 0.2 0.2

0.05 0.1 0.1

0 0 0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500

10 maximum 10 maximum ......... 10 maximum

Fig. 8. Methodology for selecting significant packets for decomposition levels from
6 to 9
5.2 ANNs trained

Once the features are extracted and selected, the ANNs training is designed.
Each energy of the packets selected represents one neuron in the input layer,
thus the number of input neurons depends on the decomposition level.

The design parameters are selected according to preliminary studies. The num-
ber of neurons at the output layer is the number of possible answers of the
ANN. In this case, ANNs did not offer good results when trying to determine
the crack size, and the best results were obtained when 2 outputs are used:
healthy or cracked.

When designing ANN, after the training stage a testing must come. Testing
must use data not used for the training, so the number of data used for each
stage must be selected. The values at the input and at the output must be
normalized to increase stability of training and testing process [38].

The spread value of the Gaussian function will be selected to optmimize the
success rate of the ANN, thus an interval is proposed. Regarding the stopping
criteria, a maximum value of SSE is proposed. If that value is not reached in
a certain number of iterations or neurons in the hidden layer, that must also

10
be selected, the training stops.

Table 6, show the common parameters selected for ANNs training, according
to the inputs described in the previous section.

Normalization of input values Between [-1;1]

Type of learning Supervised
Number of neurons in output 2
Normalization of output values [-1,1]
Training 75%
Input data distribution
Test 25%
SSE 0.1-0.2
Stopping criteria
Number of neurons at hidden layer 700
Spread 0.2-2
Table 6
Design parameters for ANN-RBF

All parameters are selected to maximize the success rate number. To carry
this, a total number of 456 ANNs were trained. From them, 24 optimal ANNs
are selected, one by each decomposition level and speed.

5.3 Results

Results of the ANNs training are shown. For each speed, 8 optimal ANNs are
selected, one for each decomposition level tested.

Figure 9 presents success rates and number of neurons in the hidden layer
of the 8 ANNs selected at 20 Hz, versus the decomposition level. It can be
concluded that decomposition level 2 offers worst results than the rest, offer-
ing lower success rates and higher number of neurons. The results at other
decomposition levels are similar.

Figures 10 and 11 show results for 40 and 60 Hz respectively. For both cases,
decomposition level number 5 seems to be the optimal one, and the speed of
60 Hz offers slightly better results.

ANNs training at decomposition level 2 always stops due to the maximum

number of neurons in hidden layer neurons. Thus, the computational cost is
higher and the success rate is lower than in other cases. For the rest of de-
composition levels, the results are very similar at low and high decomposition
levels at 20 Hz. On the other hand, at 40 and 60 Hz results at low decomposi-

11
 Speed 20 Hz Speed 20 Hz
100 700

600

500

Neurons in hidden layer

95

Success rates (%)

400

300

90
200

100

85 0
2 3 4 5 6 7 8 9 2 3 4 5 6 7 8 9
Decomposition level Decomposition level

(a) (b)
Fig. 9. Results of ANNs for each decomposition level at 20 Hz representing (a)
success rates and (b) number of neurons in the hidden layer

Speed 40 Hz Speed 40 Hz
100 700

600

500

Neurons in hidden layer

95
Success rates (%)

400

300

90
200

100

85 0
2 3 4 5 6 7 8 9 2 3 4 5 6 7 8 9
Decomposition level Decomposition level

(a) (b)
Fig. 10. Results of ANNs for each decomposition level at 40 Hz representing (a)
success rates and (b) number of neurons in the hidden layer

Speed 60 Hz Speed 60 Hz
100 700

600

500
Neurons in hidden layer

95
Success rates (%)

400

300

90
200

100

85 0
2 3 4 5 6 7 8 9 2 3 4 5 6 7 8 9
Decomposition level Decomposition level

(a) (b)
Fig. 11. Results of ANNs for each decomposition level at 60 Hz representing (a)
success rates and (b) number of neurons in the hidden layer

tion levels (3, 4 and 5) show better results than at high decomposition levels
(6, 7, 8 and 9).

12
5.4 POD calculation

The global success rate of an ANN gives important information about the
diagnosis, however it is not accurate. There is a need to test the success of
the ANN depending on the crack condition. There is a large difference if all
the fails in the diagnosis of the ANN are accumulated in false alarms or in the
crack level 1, that is not a critical crack.

Thus, Probability of Detection (POD) curves are calculated using the results
of success rates with the data kept for validation stage, for the three cases
of speed. For all cases, decomposition level 5, using all the energy packets
obtained at that level, offers the best diagnosis results. Regarding the speed,
the case of 60 Hz is the optimal one.

POD curves at 60 Hz are shown in figures 12 and 13. The optimal decompo-
sition level is number 5. High decomposition levels present a high number of
false alarms.
POD at 60 Hz decomposition level 2 POD at 60 Hz decomposition level 3
100 100

90 90

80 80

70 70
Probability of detection (%)

Probability of detection (%)

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Crack level Crack level

(a) (b)
POD at 60 Hz decomposition level 4 POD at 60 Hz decomposition level 5
100 100

90 90

80 80

70 70
Probability of detection (%)

Probability of detection (%)

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Crack level Crack level

(c) (d)
Fig. 12. POD curves obtained from ANNs at 60 Hz at low decomposition levels (a)
level 2 (b) level 3 (c) level 4 (d) level 5

Once the decomposition level 5 at 60 Hz is selected, a new ANN is trained

using all the packets at this level with a SSE = 0.1. Figure 14 shows the

13
 POD at 60 Hz decomposition level 6 POD at 60 Hz decomposition level 7
100 100

90 90

80 80

70 70

Probability of detection (%)

Probability of detection (%)

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Crack level Crack level

(a) (b)
POD at 60 Hz decomposition level 8 POD at 60 Hz decomposition level 9
100 100

90 90

80 80

70 70
Probability of detection (%)

Probability of detection (%)

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Crack level Crack level

(c) (d)
Fig. 13. POD curves obtained from ANNs at 60 Hz at high decomposition levels (a)
level 6 (b) level 7 (c) level 8 (d) level 9

POD curve calculated this case. the ANN offers the best diagnosis results
available with these data, with a number of false alarms of 1.77% and a low
computational cost with only 45 neurons.

5.5 Discussion

The main contribution of the present work is the development of an expert

system that is able to diagnose with reliability a crack in a rotating shaft using
experimental vibration signals. The value of the work lies on the fact that
most of the studies of the dynamical behaviour of cracked rotors, specifically
in shafts, do not involve the inverse process of crack detection and do not use
experimental results [39].

Tests were carried out at different speeds that show that diagnosis results are
improved with the speed. This can be assigned to the fact that when the speed
increases, as the tests were performed at the same conditions, the signal-to-
noise ratio is higher. Then, crack effects are more clearly distinguished when
the speed increases.

14
 POD at 60 Hz at decomposition level 5 SSE=0.1
100

90

80

70

Probability of detection (%)

60

50

40

30

20

10

0
0 1 2 3 4 5 6 7 8 9
Crack level

Fig. 14. POD at decomposition level 5 and goal 0.1 at 60 Hz

The aim of the system proposed is to detect a crack when it appears and grows
in a shaft, while it is rotating and thus, assembly influence is constant. The
constant assembly influence has been reproduced in the present work tests,
as the cracks are artificially induced in the shaft without dismounting it from
the machine. Therefore, differences in signals measured are only caused by
crack effects and no by the assembly or other defects. In order to check if a
mistuned assembly or other defects could hide the crack effects, and decrease
the success rates obtained in this work, it would be necessary to make this
specific tests, proposed as future work.

It would be interesting also to test if multi-level selection of WPT energies

could improve the results in terms of success rates and computational cost. In
this work the ANNs parameters and the decomposition level of the WPT are
evaluated to optimize the success rates of the ANNs. However, all the energies
selected are from the same level of decomposition (single-level).

6 Conclusions

For the present work, experimental vibration signals are taken from a rig at
steady state at three different speeds and 10 different crack conditions (healthy
and 9 crack levels from 4% of the shaft diameter to 50%). The signals are pro-
cessed by means of WPT energy using ‘Daubechies 6’ as wavelet function. The
single-level energies obtained are used to train several RBF-ANNs. The ANNs
parameters and the decomposition level of the WPT are evaluated to opti-
mize the success rates of the ANNs. Results seem to improve with the speed.

15
Features extracted at decomposition level number 5 from signals obtained at
the higher speed (60 Hz) offer the best results in terms of success rates and
computational cost. Success rates are used to calculate POD curves and the
number of false alarms is of 1.77% and crack levels above 1 can be detected
with high reliability. Results show that the methodology proposed could be
successfully integrated in industrial equipment for condition monitoring.

7 Acknowledgements

The authors would like to thank the Spanish Government for financing through
the CDTI project RANKINE21 IDI-20101560.

References

[1] D. Pedregal, M. Carnero, Vibration analysis diagnostics by continuous-time

models: A case study, Reliability Engineering & System Safety 94(2) (2009)
244–253.
[2] A. Sekhar, On-line rotor fault identification by combined model and signal based
approach, Noise and Vibration Worldwide 35 (7) (2004) 16–30.
[3] M. Feldman, S. Seibold, Damage diagnosis of rotors: application of hubert
transform and multihypothesis testing, Journal of Vibration and Control 5 (3)
(1999) 421–442.
[4] D. Guo, Z. Peng, Vibration analysis of a cracked rotor using hilbert-huang
transform, Mechanical systems and signal processing 21 (2007) 3030–3041.
[5] Q. Hu, Z. He, Z. Zhang, Y. Zi, Fault diagnosis of rotating machinery based on
improved wavelet package transform and svms ensemble, Mechanical systems
and signal processing 21 (2007) 668–705.
[6] L. Vignolo, D. Milone, H. Rufiner, Genetic wavelet packets for speech
recognition, Expert Systems with Applications.
[7] E. Avci, Z.H.Akpolat, Speech recognition using a wavelet packet adaptive
network based fuzzy inference system, Expert Systems with Applications 31
(3) (2006) 495–503.
[8] P. Mercorelli, A denoising procedure using wavelet packets for instantaneous
detection of pantograph oscillations, Mechanical Systems and Signal Processing
35 (1-2) (2013) 137–149.
[9] M. Tinati, B. Mozaffary, A wavelet packets approach to electrocardiograph
baseline drift cancellation, International Journal of Biomedical Imaging (2006)
1–9.

16
[10] Z. Peng, Application of the wavelet transform in machine condition monitoring
and fault diagnostics: a review with bibliography, Mechanical Systems and
Signal Processing 18 (2004) 199–221.
[11] X. Lou, K. Loparo, Bearing fault diagnosis based on wavelet transform and
fuzzy inference, Mechanical Systems and Signal Processing (2004) 1077–1095.
[12] W. Wang, P. McFaden, Application of wavelets to gearbox vibration signals for
fault detection, Journal of Sound and Vibrationl 192 (1996) 927–939.
[13] A. Gentile, A. Messina, On the continuous wavelet transforms applied to
discrete vibrational data for detecting open cracks in damaged beams, Journal
of Solid and Structures 40 (2003) 295–315.
[14] E. Douka, S. Loutridis, A. Trochidis, Crack identification in beams using wavelet
analysis, International Journal of Solids and Structures 40 (2003) 3557–3569.
[15] M. Herzog, T. Marwala, P. Heyns, Machine and component residual life
estimation through the application of neural networks, Reliability Engineering
& System Safety 94 (2009) 479–489.
[16] O. Fink, E. Zio, U. Weidmann, Predicting component reliability and level of
degradation with complex-valued neural networks, Reliability Engineering &
System Safety 121 (2014) 198–206.
[17] T. Santosh, A. Srivastava, V. S. Rao, A. Ghosh, H. Kushwaha, Diagnostic
system for identification of accident scenarios in nuclear power plants using
artificial neural networks, Reliability Engineering & System Safety 94 (2009)
759–762.
[18] S. Martinez-Martinez, N. Messai, J. Jeannot, D. Nuzillard, Two neural network
based strategies for the detection of a total instantaneous blockage of a sodium-
cooled fast reactor, Reliability Engineering & System Safety 137 (2015) 50–57.
[19] T. Babu, A. Sekhar, Shaft crack identification using artificial neural networks
and wavelet transform data of a transient rotor, Advances in Vibration
Engineering 9(3) (2010) 207–214.
[20] H. Srinivas, K. Srinivasan, K. Umesh, Role of an artificial neural network and a
wavelet transform for condition monitoring of the combined faults of unbalance
and cracked rotors, International Journal of Acoustics and Vibration 15(3)
(2010) 121–127.
[21] H. Dai, B. Zhang, W. Wang, A multiwavelet support vector regression method
for efficient reliability assessment, Reliability Engineering & System Safety 136
(2015) 132–139.
[22] K. Kim, M. Zuo, Two fault classification methods for large systems when
available data are limited, Reliability Engineering & System Safety 92(5) (2007)
585–592.
[23] N. Bachschmid, P. Penacci, Editorial. crack effects in rotordynamics,
Mechanical Systems and Signal Processing 22 (2008) 761–762.

17
[24] Q. Han, J. Zhao, F. Chu, Dynamics analysis of a geared rotor system considering
a slant crack on the shaft, Journal of Sound and Vibration 331 (2012) 5803–
5823.

[25] Q. Han, F. Chu, Parametric instability of a rotor-bearing system with two

breathing transverse cracks, European Journal of Mechanics A/Solids 36 (2012)
180–190.

[26] Q. Han, F. Chu, Parametric instability of a jeffcott rotor with rotationally

asymmetric inertia and transverse crack, Nonlinear Dynamics 73(1-2) (2013)
827–842.

[27] C. Castejón, J. Garcı́a-Prada, M. Gómez, J. Meneses, Automatic detection of

cracked rotors combining multiresolution analysis and artificial neural networks,
Journal of Vibration and Control, D.O.I. 10.1177/1077546313518816.

[28] C. Castejón, M. Gómez, J. Garcı́a-Prada, A. O. nez, H. Rubio, Automatic

selection of the wpt decomposition level for condition monitoring of rotor
elements based on the sensitivity analysis of the wavelet energy, International
Journal of Acoustics and Vibration 20(2) (2015) 95–100.

[29] S. Mallat, A wavelet tour of signal processing, Academic Press, 1998.

[30] A. Jensen, A. la Cour-Harbo, Ripples in Mathematics, Springer, 2000.

[31] H. Dai, B. Zhang, W. Wang, A new evolutionary system for evolving artificial
neural networks, IEEE Transactions on neural networks 8 (1997) 694–713.

[32] H. Dai, B. Zhang, W. Wang, A multiwavelet support vector regression method

for efficient reliability assessment, Reliability Engineering & System Safety 364
(2012) 012023.

[33] O. Lara, C. Castejón, J. Garcia-Prada, Bearing fault diagnosis based on neural

network classification and wavelet transform, WSEAS. Transaction on Signal
Processing 2 (10) (2006) 1371–1378.

[34] M. Gómez, C. Castejón, J. Garcı́a-Prada, Incipient Fault Detection in Bearings

through the use of WPT energy and Neural Networks, 2014, Ch. Advances
in Condition Monitoring of Machinery in Non-Stationary Operations. Lecture
Notes in Mechanical Engineering, pp. 63–72.

[35] M. Gómez, C. Castejón, R. Barber, H. Rubio, J. Garcı́a-Prada, Btool:

a friendly teaching tool to acquire and process vibration signals,
Proceedings of the 14th IFTOMM World Congress, Taipei, Taiwan (2015)
10.6567/IFToMM.14TH.WC.OS5.001.

[36] M. Gómez, C. Castejón, J. Garcı́a-Prada, Crack detection in rotating shafts

based on the 3x energy, Mechanism and Machine Theory 96 (2016) 94–106.

[37] M. Gómez, C. Castejón, J. Garcı́a-Prada, Energy analysis distribution regarding

the crack size in a rotating shaft, Proceedings of the 14th IFTOMM World
Congress, Taipei, Taiwan (2015) 10.6567/IFToMM.14TH.WC.OS14.010.

18
[38] M. Chow, Methodologies of using neural network and fuzzy-logic technologies
for motor incipient fault detection, World Scientific, 1997.

[39] N. Bachschmid, P. Penacci, E. Tanzi, Some remarks on breathing mechanism,

on non-linear effects and on slant and helicoidal cracks, Mechanical Systems
and Signal Processing 22 (2012) 879–904.

19