You are on page 1of 11

Automated cepstral editing procedure (ACEP) as a signal

pre-processing in vibration-based bearing fault diagnostics


Agusmian Partogi Ompusunggu1

1 Flanders
Make
Celestijnenlaan 300, 3001 Heverlee (Leuven), Belgium
{agusmian.ompusunggu}@flandersmake.be

Abstract
The cepstral editing method for removing discrete components in vibration signals has been automated in our
previous work [1]. In the automated cepstrum editing procedure (ACEP), the narrow impulses in the cepstrum
domain originating from the discrete components are removed by applying a comb lifter to the real cepstrum.
The comb lifter is automatically generated from the peak-enhaced version of the real cepstrum obtained after
applying the spectral subtraction method to the high-pass liftered cepstrum. The automated method has been
applied and validated to vibration signals measured on a realistic gearbox set-up. In this paper, the performance
of the ACEP method for bearing fault diagnostics is quantitatively compared with the performance of the
traditional cepstral editing method. The results show that the automated method outperforms the traditional
cepstra editing method.

1 Introduction
Detecting bearing faults on rotating machinery based on vibration signals is often a challenge due to the
high energy, dominating signals originating from various machine elements like gears, screws, and shafts that
can mask weak, non-deterministic (random) signals generated by bearing faults. These dominant signals are
deterministic and appear as discrete components in the frequency domain. For bearing faults detection it is
therefore required to remove these discrete components prior to applying further signal processing.
Randall and Sawalhi [2, 3, 4] have shown quite recently that it is possible to remove discrete components
in a vibration signal by means of editing (liftering) the real cepstrum of the signal. This method is called
the cepstral editing procedure (CEP). A number of comparative studies about the performance of the CEP
method with other established discrete component removal (DCR) methods have been recently reported in
the literature [5, 6]. In Randall et al. [5], a qualitative comparison of four different methods including (i) the
time-synchronous averaging (TSA), (ii) the self-adaptive noise cancellation (SANC), (iii) the discrete/random
separation (DRS) and the (iv) CEP, is presented. It showed that the CEP method enhances bearing fault related
signals more significantly compared with the other methods. In Kilundu et al. [6], the performance of the
CEP method in enhancing bearing fault related signals is quantitatively benchmarked with two DCR methods,
namely (i) the TSA and (ii) the SANC method. This study demonstrated, through analysis on a number of
experimental vibration signals, that the CEP method outperformed the other two methods.
So far, the editing procedure in the CEP method is performed manually by using some information about
the shaft speed and machine configuration: transmission ratio, number of gear teeth, etc. However, for devel-
opment of a continuous, online condition monitoring system, manual editing is thus not a workable solution.
To solve this problem, an automated editing procedure for cepstrum editing, where shaft speed information and
machine configuration are no longer required, is proposed for the first time in this paper. The automated editing
procedure, which is from now on referred to as the automated cepstrum editing procedure (ACEP), is based on
a comb lifter that removes narrow impulses and the rahmonics. Here, the comb lifter is automatically generated
based on all detected peaks on the peak-enhanced cepstrum.
The remainder of this paper is structured as followed. A short introduction about the cepstrum theory, the
CEP method, and the proposed ACEP method is presented in Section 2. The test rig and the procedure to
generate experimental data used for demonstrating the validity and the effectiveness of the ACEP method are

1
presented in Section 3. Analysis on the experimental data and the results are discussed in Section 4. Finally,
some important remarks taken up from this study are concluded in Section 5.

2 Cepstrum editing
2.1 The cepstrum and its application
Despite various definitions, the commonly used definition of the cepstrum of a signal x(t) is the ”inverse
Fourier transform of the log spectrum”, which can be mathematically written as follows:

c(τ) = F −1 [log (X( f ))] , (1)

where X( f ) = F [x(t)] is the frequency spectrum of the signal x(t), F and F −1 are respectively referred to
as the Fourier transformation and the inverse Fourier transformation, c(τ) is the signal cepstrum and τ is the
quefrency.
Let us assume that the vibration signal x(t) is noise-free and generated by an excitation signal e(t) applied
to a mechanical system with the impulse response function h(t). In the frequency domain, the relationship of
the three terms is simply written as:
X( f ) = H( f )Y ( f ), (2)
with H( f ) denoting the frequency spectrum of the impulse response function h(t) (i.e. frequency response
function), and E( f ) denoting the spectrum of the excitation signal e(t). Substituting Eq. (2) into Eq. (1), one
can easily show that
c(τ) = F −1 [log (H( f ))] + F −1 [log (E( f ))] . (3)
Eq. (3) clearly shows that the cepstrum of the vibration signal can be regarded as the addition of the cepstrum
of the excitation signal and the cepstrum of the impulse response function. This non-linear transformation
suggests that the effects of the excitation and the system behavior described by the modal parameters (e.g. the
resonance frequency, damping ratio) can be separated in the cepstrum domain. Note that the modal parameters
of the system are concentrated at low quefrency (starting from zero quefrency), while the periodic excitation
components (i.e. discrete components) are typically located at high quefrency.

2.2 Conventional cepstrum editing procedure and its variant


The distinctive property of periodic signals in the cepstrum domain is that they appear as narrow impulses
with a number of rahmonics. Based on this property, Randall et al. [2, 3, 4] have proposed the CEP method to
remove periodic components from vibration signals by editing, i.e. liftering, the magnitude of the real cepstrum
at quefrencies of interest as schematically shown in Figure 1. So far, the liftering procedure is performed
manually by using some information about the shaft speed and machinery configuration including transmission
ratio, number of gear teeth, etc. In the following subsection, an automated procedure for editing the real
cepstrum is presented.

2.3 Automated procedure for cepstrum editing


Figure 2 schematically shows the automated procedure for editing narrow impulse with a number rahmon-
ics on the real cepstrum. First, the real cepstrum is high-pass liftered in order to focus on high-quefrency
components. All the peaks corresponding to discrete components are successively enhanced with the spectral
subtraction method. Finally, all detected peaks are used to automatically generate a comb lifter. The detailed
description of each step is presented below. Let c(τn) be the cepstrum of a discretized vibration signal x(tn),
with τn = n∆τ and tn = nTs respectively denoting the discrete time and discrete quefrency, Ts = ∆τ denoting
the sampling period, n = 0 : N − 1 denoting the time and quefrency indexes, and N denoting the signal length.
For the sake of conciseness, one can also express the discrete spectrum and signal as follows: c(n) and x(n),
respectively. Note that because of the symmetry property of the cepstrum inherited from the symmetry property
of discrete Fourier transform (DFT), the cepstra of discrete signals are therefore only relevant to be analyzed
until half of the signal length, i.e. c(n), n = 0 : bN/2c, with b·c denoting the rounding operation.

2
Edited log
Original time Phase cepstrum
FFT + Exp.
domain signal + Complex
spectrum
Magnitude Log.
IFFT
operator
Log Edited log
magnitude amplitude Enhanced time
cepstrum domain signal
IFFT

Real
cepstrum
Edited
Automated cepstrum
FFT
editing

Figure 1: Shematic diagram of the convential CEP method. The dashed line in the figure indicates that liftering
is performed manually.

Hi-pass Peak-
liftered enhanced
Spectral subtraction cepstrum
Real Hi-pass cepstrum Automated peak
(SS) based peak
cepstrum lifter detector
amplifier

Edited
Comb lifter
cepstrum

Figure 2: Schematic diagram of the automated editing procedure.

2.3.1 High-pass lifter


In order to keep the modal information localized at low quefrency, a high-pass lifter lH (n) needs to be
applied to the real cepstrum c(n). The cut-off quefrency τc of the high-pass lifter is determined based on typical
resonance frequencies excited in rotating machinery. As a rule of thumb, the cut-off quefrency can be set to
τc = j/ fn , with j denoting an integer number and fn = 6 kHz. The quefrency index of this cut-off quefrency
corresponds to Nc = bτc /Ts c. Mathematically, the high-pass lifter lH (n) can be expressed as:

0 n = 0 : Nc
lH (n) = (4)
1 n = Nc + 1 : bN/2c

The high-pass liftered cepstrum cHL (n) can thus be written as:

cHL (n) = c(n)lH (n). (5)

2.3.2 Spectral subtraction


Although the spectral subtraction (SS) is a well known technique (mainly in speech signal processing
community) for signal denoising because of its good performance and low computational complexity [7], it
has received very little attention in condition monitoring community [9]. The SS method works under the
following assumptions: (i) signal of interest and background noise are uncorrelated, (ii) the background noise
is time-invariant, and (iii) the spectra of the noise and signal are different.
The high-pass liftered cepstrum cHL (n) can be seen as the cepstrum conveying information about the

3
discrete-components c̃(n) corrupted by additive background noise ψ(n). This can be expressed as follows:
cHL (n) = c̃(n) + ψ(n). (6)
In the SS method, the processing is carried out on a frame-by-frame basis (i.e. window basis) in the frequency
domain. The method is illustrated in Figure 3. The magnitude of the peak-enhanced cepstrum C̃(k) conveying

Figure 3: Overview of the spectral subtraction (SS) algorithm.

information about discrete components is obtained by multiplying the magnitude of the high-pass liftered cep-
strum |C(k)| with an adaptive gain G(k) which is recursively calculated based on the noise magnitude |Ψ(k)|.
Finally, the peak-enhanced cepstrum c̃(n) can be reconstructed using the magnitude of the peak-enhanced
cepstrum C̃(k) of all frames and the original phase information ∠C(k). For the sake of conciseness, the SS
algorithm is not discussed in detail in this paper. The interested readers are referred to Refs. [7, 8] for the full
description of the method.

2.3.3 Peak detection


The peak-enhanced cepstrum c̃(n) obtained after applying the SS method to the high-pass liftered cep-
strum cHL (n) is then squared to further emphasize all the peaks. A simple decision function, based on a fixed
threshold th calculated from the remaining floor noise in the cepstrum, has been implemented in order to auto-
matically find the peak locations in the cepstrum. The decision function d(n) at an arbitrary quefrency index n
is formulated as follows:
1 c̃2 (n) > th

d(n) = (7)
0 else
where c̃( 2)(n) denotes the squared peak-enhanced cepstrum at the quefrency index n and th denotes the thresh-
old. Here, the threshold is calculated according to the Shewhart process control [10] as formulated in the
following equation:
th = E c̃2 (bN/2c − L + 1 : bN/2c) + κ × std c̃2 (bN/2c − L + 1 : bN/2c) ,
   
(8)
where E [·] denotes the expectation operator, std [·] denotes the standard deviation operator and κ = 6 denotes a
constant. The peak location indexes p can be retrieved based on the value of the decision function d(n) at the
places where its value equals one:
p = {∀n| d(n) = 1} . (9)

2.3.4 Comb lifter generation and comb liftering


Given the peak location indexes p, a comb lifter lC (n), with n = 0 : bN/2c, can be constructed according to
the following expression: 
0 n = p−2 : p+2
lC (n) = (10)
1 else
Finally, the edited cepstrum ĉ(n) is obtained by multiplying the original real cepstrum c(n) with the comb lifter
lC (n):
ĉ(n) = c(n) × lC (n). (11)

4
2.3.5 Computational complexity
The execution time of the ACEP method has been evaluated on synthetic signals with different signal
lengths using a PC with the processor of IntelTM Core® i7-4610M CPU@3 GHz, the RAM of 16 GB, and a
64-bit Windows based operating system. As a benchmark, the execution times of the FFT of the signals are
also calculated and plotted in function of signal length N as shown in Figure 4. The figure shows that the
execution time of the ACEP method regardless of the frame size is proportional to the execution time of the
FFT algorithm. As the complexity of the FFT algorithm is O (N × log(N)), the complexity of the ACEP method
is thus O (λ × N × log(N)), with λ denoting a constant.

Figure 4: Computational time comparison between the ACEP method for various frame sizes (N/2, N/8, N/32)
and the FFT algorithm.

3 Experimental methodology
3.1 Experimental setup
Figure 5 shows the photo and the schematic-top-view of the gearbox setup used in this study. The test setup
consists of two gearboxes connected to each other, namely a parallel-shaft gearbox (3) and a perpendicular-
shaft gearbox (4). The input shaft (10) of the parallel-shaft gearbox is driven through a flexible coupling by
an induction electric mo-tor (2) which is controlled by a variable-frequency-drive (1). The shaft rotational
speed of the motor can be varied from 0 to 3,000 rpm, with either a stationary mode or a transient mode (run-
up/run-down). The output shaft (13) of the perpendicular shaft gearbox is coupled with a magnetic-particle
brake (5), where the torque applied to the output-shaft can be adjusted by the brake controller (8) from 0 to 50
Nm. The control signals to the motor drive (1) and to the brake controller (8) are sent out by a PC (7) using
dedicated Labview programs. The parallel-shaft gearbox (3) was constructed with 3 parallel-shafts as seen in

2 11 10 3 5
12 L
1

13
B#1

B#2

A#2

M G#1

Tach
2
B#3

B#4

5 G#2 G#3 B#8


1
G#6
G#5
B#5

B#6
A#1

G#4

B#7

A#3

7 9 8
4

3 4 6 NI DAQ
6

(a) (b)

Figure 5: Picture (a) and the schematic diagram (b) of the experimental setup.

5
Figure 5(b). Four helical gears are arranged in the parallel-shaft gearbox such that it has two-stage reductions.
Two straight-bevel gears are assembled onto their corresponding shaft in the perpendicular-shaft gearbox (4) in
a one-stage reduction. Details of the gears assembly are summarized in Table 1. With this configuration, the
total reduction factor from the motor shaft (10) to the output shaft (13) of the perpendicular-shaft gearbox can
be easily calculated based on the number of teeth of the gears, that is equal to (100/29) × (36/90) × (40/20) =
2.76.

Component G#1 G#2 G#3 G#4 G#5 G#6


#teeth 29 100 90 36 20 40
Comment Helical Helical Helical Helical Straight-bevel Straight-bevel

Table 1: Details of gear assembly.

Two different bearing types are assembled in the test setup, namely MB ER-14K and MB ER-16K. Two
identical bearings (B#1 and B#2) of the former type are mounted in the parallel-shaft gearbox to support the
input shaft (10) and the other bearings (from B#3 to B#8) of the latter type are mounted to support the other
shafts, see Figure 5(b). They all are deep-groove ball bearings with the same geometrical specifications, thus
the same theoretical fault frequencies as listed in Table 2.

FTF [Hz] BPFO [Hz] BDF [Hz] BPFI [Hz]


0.402 3.572 4.644 5.430

Table 2: Theoretical bearing fault frequencies of the MB ER-16K bearing for the shaft speed of 60 rpm.

3.2 Test procedure


3.2.1 Faulty state simulation
In a faulty state, a MB ER-16K bearing with combined damages, i.e. (i) inner race fault, (ii) outer race fault
and (iii) rolling element fault, was mounted on the second intermediate shaft (12).

3.2.2 Operating condition


The input shaft rotational speed of the gearbox setup was set to 2400 rpm. As the reduction factor between
the input shaft (10) and the second intermediate shaft (12) is of 100/29 × 36/90 (= 0.7250), the rotational
speed of the intermediate shaft where the faulty bearing was mounted, was of 1740 rpm (i.e. 29 Hz). All the
experiments were performed under a constant brake torque of 20 Nm.

3.2.3 Measurement settings


Three accelerometers triaxial piezoelectric (ICP type) accelerometers (A#1, A#2 and A#3) were mounted
at different positions as can be seen in Figure 5(b). The three accelerometers have the same sensitivity of 100
mV/g with ±5 % response deviation in the frequency range of 0.5 Hz to 5 kHz. An optical tachometer system
was used to measure the shaft rotational speed of the motor where the light from the sensor head is directed to
a reflective tape attached on the motor shaft, thus generating one pulse signal per revolution. All the signals
are synchronously acquired by a National Instruments (NI) data acquisition system (6), low-pass filtered with
built-in anti-aliasing filters, and sampled at 51.2 kHz for a few seconds. The recorded digital data are then
processed within Matlab® as will be discussed in the next section.

4 Data analysis and discussion


Vibration signals acquired and recorded using the gearbox setup as described in the previous section have
been analyzed and processed in order to demonstrate the effectiveness and the performance of the developed
method (ACEP) compared with the traditional method (CEP). In the following paragraphs, the results obtained

6
using the ACEP method are first presented and discussed, subsequently followed by the results obtained using
the CEP method.
To better understand the effect of each signal processing step in the ACEP method as discussed in Sec-
tion 2.3, the resulting cepstrum of each step is visually presented. Figure 6 shows a representative automated
generation of a comb lifter lC (n) from a real cepstrum c(n) and the resulting cepstrum from each signal pro-
cessing step. Important note that the background noise of the high-pass liftered cepstrum has been significantly
reduced after applying the SS method, thus enhancing the peaks as can be seen in the panel iii of Figure 6.
Subsequently, the generated comb lifter is applied to the real cepstrum in order to remove all the dominating
peaks for further analysis as discussed below.

Figure 6: Representative generation of a comb filter from a real cepstrum.

Once the comb lifter has been automatically generated, the edited cepstrum (i.e. discrete-component re-
moved cepstrum) can be calculated by simply multiplying the comb lifter lC (n) with the (real) original cepstrum
c(n). Figure 7 shows comparison between the cepstra before and after automated editing of the vibration signals
under healthy and faulty bearing states. As seen in the middle panels of the figure, a number of narrow peaks
corresponding to the discrete components (e.g. shaft frequencies, gear mesh frequencies, sidebands, and the
harmonics) are evident in the original cepstra. After applying the automated editing procedure to the original
cepstra, most of the dominating peaks have been removed as seen in the edited cepstra (see the lower panels).
Similarly, the CEP method has also been applied to the recorded vibration signals. Figure 8 shows com-
parison between the cepstra before and after manual editing of the vibration signals under healthy and faulty
bearing states. It is evident from the figure that the traditional method is able to remove significant peaks
as shown by the edited cepstra. However, some minor peaks mainly located at low quefrencies are not well
removed by the traditional method. Note that these minor peaks correspond to sidebands of the gear mesh
frequencies. The presence of these sidebands in the signal may pollute the envelope spectrum, which in turn
deteriorates diagnostic features.
The edited cepstra in Figures 7 and 8 have been transformed into the time-domain signals according to the

7
(a) (b)

Figure 7: Cepstra of the (a) healthy and (b) faulty bearing vibration signals before and after automated editing
(ACEP).

procedure described in Figure 1. For better visual representation, the time-domain signals are then transformed
into the angular domain as shown in Figure 9. Note that the angular-domain transformation is carried out based
on the rotation of the second intermediate shaft (refer to item#12 in Figure 5), on which the faulty bearing was
mounted. As seen in Figure 9, the kurtosis values of the signals become higher after removing the discrete
components both with the developed method (ACEP) and the traditional method (CEP). One can notice that
the increase of the kurtosis value of the vibration signal of faulty bearing filtered using the developed method
is more pronounced than the one filtered using the traditional method.
Figure 10 shows the spectra of the healthy and faulty bearing vibration signals before and after editing with
both the developed and traditional method. Here, the spectra are plotted in function of the rotation order of
the second intermediate shaft. In general, one can see that the spectra magnitudes of the dominating discrete
components have been reduced more significantly after applying the ACEP method than after applying the CEP
method to the vibration signals. There are three gear-meshing frequencies (GMFs) that can be identified on the
gearbox setup. Based on the gear assembly and the number of gear teeth listed in Table 1, one can determine
that these GMFs correspond to the order of 20, 36 and 40 in the order domain. The spectra magnitudes before
and after discrete component removal around the GMFs are highlighted by the boxes with dashed-lines in
Figure 10. The figure shows the power of discrete components in the signals has been reduced up to 20 dB with
the ACEP method. In contrast, the CEP method reduces the power of discrete components of up to 10 dB.
The resulting signals after discrete component removal with the ACEP method are further processed using
the envelope analysis technique described in [11]. According to this envelope technique, the resulting signals
are band-pass filtered and demodulated, where the frequency band and the carrier frequency are determined us-
ing the spectral analysis [12]. Finally, the squared envelope spectra (SES) are calculated, that will be eventually
used to extract diagnostic features.
Figure 11 shows the normalized envelope spectra in the order domain calculated from the vibration signals
processed both with the ACEP and CEP methods of the healthy and faulty bearing. In the envelope spectrum
of the healthy bearing obtained using the ACEP method (see Figure 11(a)), no significantly distinctive peaks
can be observed. On the contrary, relatively significant peaks close to the theoretical BDF and BPFI can be
observed in the envelope spectrum of the healthy bearing obtained using the CEP method, which in turn can
affect diagnostic conclusions.
In the faulty case, narrow peaks around the three fault characteristic frequencies, namely BPFO, BDF and
BPFI corresponding to the three seeded faults (outer race, inner race and rolling element faults), are evident in

8
(a) (b)

Figure 8: Cepstra of the (a) healthy and (b) faulty bearing vibration signals before and after manual editing
(CEP).

(a) (b)

Figure 9: Comparison of vibration signals before and after cepstra editing of (a) healthy and (b) faulty bearing
in function of the second intermediate shaft rotation.

(a) (b)

Figure 10: Power spectra of the (a) healthy and (b) faulty bearing vibration signals before and after cepstra
editing.

9
(a) (b)

Figure 11: Envelope spectra of the (a) healthy and (b) faulty bearing vibration signals.

the envelope spectra calculated from the vibration signals processed both with the ACEP and CEP methods as
shown in Figure 11(b). However, one can see that the peak values of the envelope spectrum of the ACEP method
are higher than those of the CEP method. For reference, these fault frequencies are listed in Table 2. Notably,
the sidebands spaced at 1×RPM around the BPFI and the harmonics indicate the presence of unbalance on the
shaft.

5 Conclusion
A new automated cepstra editing procedure (ACEP) for discrete component removal in vibration signal
analysis has been developed. The method is based on applying a comb-lifter, which is automatically generated,
to the real cepstrum. The computational complexity of the ACEP method is proportional to the computational
complexity of the FFT algorithm, thus allowing for continuous and on-line implementation. The effectiveness
of the method has been demonstrated through analysis of vibration signals recorded from a complex gearbox
setup and been compared with the traditional method. The results show that the automated method (ACEP)
can reduce the power spectrum of the discrete components more effectively than the traditional cepstral editing
method. In addition, bearing fault features obtained with the ACEP method are more pronounced than those
obtained with the traditional method.

Acknowledgements
This work was performed within the frame of the Top Competence research project of the Smart monitoring
systems research program of Flanders Make.

References
[1] A. P. Ompusunggu, T. A. Bartic, Automated cepstral editing procedure (ACEP) for removing discrete
components from vibration signals, The twelfth International Conference on Condition Monitoring (CM)
and Machinery Failure Prevention Technologies (MFPT), 9 - 11 June 2015, The Oxford Hotel, Oxford,
UK.

[2] R. B. Randall and N. Sawalhi, Signal pre-whitening using cepstrum editing (liftering) to enhance fault
detection in rolling element bearings, Proceedings of the 24th International Congress on Condition Moni-
toring and Diagnostics Engineering Management (COMADEM), Stavanger (Norway), 30th May - 1st June,
2011.

10
[3] R. B. Randall and N. Sawalhi, A new method for separating discrete components from a signal, Sound &
Vibration, May 2011.

[4] R. B. Randall and N. Sawalhi, Cepstral removal of periodic spectral components from time signals, Pro-
ceedings of the 3rd International conference on Condition Monitoring of Machinery in Non-Stationary
Operations (CMMNO), Ferrara (Italy), 8th - 10th May, 2013.

[5] R. B. Randall, N. Sawalhi and M. Coats, A comparison of methods for separation of deterministic and
random signals, The International Journal of Condition Monitoring, Vol. 1 (1), pp. 11 - 19, June 2011.

[6] B. Kilundu, A. P. Ompusunggu, F. Elasha and D. Mba, Effect of parameter settings on performance of
discrete component removal (DCR) methods for bearing faults detection, Proceedings of the European
conference of the Prognostics and Health Management (PHM) Society, Nantes (France), 8th - 10th July,
2014.

[7] S. F. Boll, Supression of acoustic noise in speech using spectral subtraction, IEEE Transactions on Acous-
tics, Speech and Signal Processing, Vol. Assp-27, No. 2, April 1979.

[8] M. Berouti, R. Schwartz and J. Makhoul, Enhancement of speech corrupted by acoustic noise, In Proc.
IEEE ICASSP, Washington DC, pp. 208 - 211, April 1979.

[9] P. Estocq, F. Bolaers and J. P. Dron, Method of denoising by spectral subtraction applied to the detection
of rolling bearings defects, Journal of Vibration and Control, Vol.12 (2), pp. 197-211, 2006.

[10] W. A. Shewhart, Economic control of quality of manufactured product, D. Van Nostrand Company: New
York, pp. 501, 1931.

[11] R. B. Randall and J. Antoni, Rolling element bearing diagnostics - A tutorial, Mechanical Systems and
Signal Processing (MSSP), 25(2), pp. 485 - 520, 2011.

[12] J. Antoni, Fast computation of the kurtogram for the detection of transient faults, Mechanical Systems
and Signal Processing (MSSP), 21(1), pp. 108 - 124, 2007.

11

You might also like