Applied Acoustics xxx (2017) xxx–xxx

Contents lists available at ScienceDirect

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

Cyclic sources extraction from complex multiple-component vibration

signal via periodically time varying filter
Piotr Kruczek a,⇑, Jakub Obuchowski a, Agnieszka Wylomanska b, Radoslaw Zimroz c
a
Research and Development Centre, KGHM Cuprum Ltd, Sikorskiego 2-8, 53-659 Wroclaw, Poland
b
Faculty of Pure and Applied Mathematics, Hugo Steinhaus Center, Wroclaw University of Science and Technology, Wybrzeze Wyspianskiego, 27, 50-370 Wroclaw, Poland
c
Diagnostics and Vibro-Acoustic Science Laboratory, Wroclaw University of Science and Technology, Na Grobli 15, 50-421 Wroclaw, Poland

a r t i c l e i n f o

Article history: Abstract: The problem of local damage detection is widely discussed in the literature. There are many
Received 12 August 2016 methods which can be applied, however there is still a need for new techniques addressing specific diag-
Received in revised form 3 March 2017 nostic issues. In particular, the case of complex multiple-component vibration signal is a challenging
Accepted 14 May 2017
problem. In this paper we focus on such a problem related to a gearbox operating in industrial conditions.
Available online xxxx
Our method consists of several stages. First we transform signal to time-frequency domain using spectro-
gram. Then for each frequency bin we apply a novel procedure which indicates location of cyclic impulses
Keywords:
in given time series. This algorithm is based on the periodically distributed local maxima detection and
Semi-blind source extraction
Time-frequency adaptive filter
quantification of their significance. Such procedure requires a priori known fault frequency. If the
Rotating machinery machine might reveal multiple fault, the procedure has to be calculated separately for each fault fre-
Local maxima quency. A time-varying filter is designed using the indicated local maxima comprised in the score matrix.
Then, signals representing each fault frequency are obtained using inverse short-time Fourier transform
algorithm. The method is illustrated with application to simulated and real data from complex mining
machine - heavy duty gearbox.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction gearbox vibration signal. Another approach to involve ICA in condi-

tion monitoring is presented in [9] where its combination with
Fault detection on early stage can prevent the machinery from kurtogram is proved as a reliable method for source identification.
unexpected breakdowns. Furthermore, it gives a possibility to Once the data from several channels (sensors) is available, the
schedule the repair for the most convenient time. Thus, the principal component analysis can be applied to provide more com-
machine performance can be enhanced in terms of availability prehensive information than from a single sensor [10]. In order to
and reliability. It has been already described by many authors that assess condition of the rotating machine one can also extract fea-
damage diagnosis based on vibration signal analysis can provide tures related to its condition and detect anomalies instead of mak-
satisfactory results [1–5]. The signal acquired on the machine often ing a decision upon the entire signal [11]. Furthermore, signals
consists of many components related to different machine ele- from complex machines are mixture of several sources. For
ments. Moreover, the environment in which the machine is located instance, planetary gearbox consists of many components assem-
influences the data as well. It is especially important in case of the bled in a compact form. Therefore, vibration signal from such
mining industry, where machines work in harsh conditions [4,6,7]. machine requires advance processing tools [12–14]. An interesting
In such case the standard fault detection methods might not per- approach was introduced in [15], namely the logarithm is applied
form properly. In order to overcome this issue a source extraction in order to stabilize the variance. It provides the new log-envelope
method can be applied taking into account specific characteristics procedure for signal analysis. In case of the highly impulsive con-
of external sources. In [8] the kernel independent component anal- tamination of the vibration signal one can extract signal related
ysis (ICA) was used to extract the fault-related component of the to damage via cancelation of the impulses, which are not related
to the damage (are not cyclic). As a result the signal with more vis-
ible and easier to detect cyclic impulses is obtained. In [16] the
⇑ Corresponding author.
authors presented the method for cancelation based on the regime
E-mail addresses: [email protected] (P. Kruczek), jobuchowski@
cuprum.wroc.pl (J. Obuchowski), [email protected] switching model. Blind extraction of source signal related to dam-
(A. Wylomanska), [email protected] (R. Zimroz). age might be also performed by blind deconvolution methods

http://dx.doi.org/10.1016/j.apacoust.2017.05.013
0003-682X/Ó 2017 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
2 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

e.g. minimum entropy deconvolution, maximum correlated kurto- analyzed fault frequency has to be set. In case of testing several dif-
sis deconvolution etc. [17–21]. Moreover, it was shown in [22] that ferent fault frequencies the following operation have to be repeated
a-stable distribution can be applied to extract the signal of interest. for each fault frequency separately. Thus, the proposed methodol-
Furthermore, the filter based on the a-stable distribution for signal ogy allows to extract separate sources related to different modula-
enhacment can be designed [23]. On the other hand, in case of the tion frequencies. On the other hand, indicators of the informative
early stages of the damage the tempered stable distribution is suit- frequency band, e.g. spectral kurtosis, infogram, protrugram etc.
able [24]. The sources in a vibration signal from rotating machine would indicate both carrier frequency bands as containing informa-
might be also considered as cyclostationary. The procedure for tion about local damage [31–34]. As a result, signals related to each
extraction of second-order cyclostationary sources is illustrated fault frequency would not be separated.
in [25]. The described method requires to predefine the fault fre- In the following step for each frequency bin in the time series
quencies. Another cyclostationary source extraction algorithm is Specð:; f Þ local maxima are found. The crucial parameter in this
presented in [26], where a subspace decomposition of the signal step is the range in which the maximum is founded. Low range
using periodic statistics is investigated. In [27] authors illustrate leads to large number of local maxima unrelated to local damage.
the application of the discrete-random extraction to the non- On the other hand, wide range could result in some significant fault
stationary signal. In case of local damage the fault component signatures omitted. Therefore, we propose to relate the range with
j k
reveals in the signal as a periodic pulse train, contrary to non- fs
considered fault frequency, i.e. r ¼ 23 f Nw ð1O . Such r translates
vÞ f
cyclic impulsive noise that might arise due to specific operation
the period related to fault frequency into the spectrogram time
of a machine. This property has been already used for damage
detection in [28]. axis. The factor 23 is responsible for slight reduction of the range,
In this article we propose a novel method for source separation since subsequent fault-related local maxima might occur on the
which is based on a periodically time varying filter. An example of boundary of the period 1=f f . The binary function that indicates if
application of periodically correlated structure to the real data is the time point ti reveals the local maximum might be defined as:
PAR time series. In [29] the method was applied to the data from (
1; if Specðt i ; f Þ ¼ max fSpecðt k ; f Þg
the energy marked. Moreover, in [30] the spectral properties of Mðt i ; f Þ ¼ ir6k6iþr
the PARMA sequences were analyzed. Filter coefficients are cyclic 0; otherwise:
and the filter design reflects the periodic nature of impulses related
to damage. In the first step the signal is represented in time- Thereafter each subband in M is considered separately. Given f,
frequency domain. Then data in each frequency bin is considered a score, which quantifies periodicity, is assigned to each ti . It eval-
as time series and local maxima with given time axis range are uates the average of M values at time points
indicated. Each local maximum is assigned with a value propor- ð. . . ; t i  2T; ti  T; t i ; ti þ T; t i þ 2T; . . .Þ, namely:
tional to the number of equally spaced local maxima indicated P
16t i þkT6bN=Tc Mðt i þ kT; f Þ
by the previous step. The requested time interval between these scoreðt i ; f Þ ¼ ;
bN=Tc
local maxima is related to the investigated fault frequency. Finally,
the obtained two dimensional score values are considered as the j k
fs
where T ¼ f f Nw ð1Ov Þ
is the fault-related period (in samples) and
time-varying filter with cyclic coefficients. Different fault frequen-
cies will result in a different filter design. The performance of pro- k 2 Z. The score matrix indicates the average number of local max-
posed algorithm is tested on the simulated and real data recorded ima in time points spaced by T. Thus, it might be considered as a
on a heavy-duty gearbox operating in an underground mine. time-varying filter with periodic coefficients, since
The paper is organized as follows. Section 2 describes the scoreðti ; f Þ ¼ scoreðti þ kT; f Þ.
methodology, including details of the filter design. In Section 3 In order to return to the time domain the STFT is multiplied
application to simulated data is discussed. Application to real data element-wise by the score matrix. Then, the inverse short-time
is presented in Section 4. Section 5 summarizes the results of this Fourier transform algorithm is applied to such STFT with modified
work and draws conclusions. amplitudes and filtered signal yðtÞ might be further analyzed [35]:
Z 1 Z 1

2. Methodology yðtÞ ¼ STFT x ðt; f Þscoreðt; f Þe2pift dtdf :

1 1

The proposed methodology consists of six main steps. In the The numerical computation of the inverse short-time Fourier
first one the vibration signal xðtÞ is converted into the time- transform can be applied with weighted Overlap-add method
frequency domain. In this paper we propose to use the spectro- [36]. As the result both the filtered and raw signals are of the same
gram, which is based on the short-time Fourier transform, length. Amplitude of yðtÞ reflects not only presence of periodic
although other decompositions might be beneficial as well. The local maxima in raw signal spectrogram but also real amplitudes
formula for spectrogram is presented in Eq. (1): related to them. Thus, the proposed method takes into account
 2 presence of periodic amplitude modulation in each subband sepa-
XK1  rately, and the corresponding energy. It is worth mentioning that
 
Specðt; f Þ ¼ jSTFTðt; f Þj ¼  wðt  mÞxðmÞe2ipfm=K  ;
2
ð1Þ
m¼0  in case of real signal, which is cyclostationary of order 2 (CS2),
yðtÞ is also CS2. Therefore, the standard envelope based method
where wðÞ is a W-long window function, t ¼ 1; . . . ; N is a time point, can be applied. On the other hand, proposed filtration provide
f is a frequency bin and K P M is the number of points in which the the signal, which contain only the cyclic components, thus the
Fourier Transform is calculated. Given the time-frequency represen- proper harmonics should be more visible in the envelope spec-
tation of the data, time series related to each frequency bin are trum. In Fig. 1 the flowchart of the proposed algorithm is
examined for having cyclic properties. Periodic amplitude modula- presented.
tion, shared by several frequency bins, might be a signature of local Clearly, real signals might consist of two sources with different
damage. It starts with spectrogram-based decomposition of the sig- modulation frequencies, which are observed in different carrier
nal on subbands with following parameters: Nw - window length, frequencies. In most of such cases the algorithm is expected to sep-
Ov - overlap (percentage of overlapping windows), fs - sampling fre- arate signal sources appropriately. The problem might occur once
quency and NFFT - the number of FFT points. In the next step the the first modulation frequency is the multiple of the second and

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx 3

Fig. 1. Flowchart of the semi-blind source extraction algorithm.

related series of excitations are consistent in phase. The extracted have values around 0:02 and the later one are approximately
signal related to the lower modulation frequency would contain 0:04. For both components the white noise with amplitudes
some of the excitations related to the higher modulation fre- 0.01 was added. The spectrogram parameters are as follows:
quency. This problem can be overcomed with the proper local N w ¼ 250; Ov ¼ 96%. According to envelope spectrum (Fig. 2c) of
maxima range r selection. It is worth mentioning that once the the raw simulated signal only harmonics for 17 Hz can be
range r value is too high for given fault frequency, the periodic local observed. The amplitude of impulses with modulation frequency
maxima would not be detected. One can observe that it is of great 5 Hz is significantly lower, thus corresponding harmonics are hid-
importance to set the proper r. It is also of high importance if the den. In this study we applied the proposed source separation algo-
proposed methodology requires pre-whitening in order to appro- rithm for two different fault frequencies in order to illustrate its
priately deal with high-energy components. basic properties. Due to spectrogram parameters, the range r corre-
sponding to fault frequency 5 Hz is equal to 164. Binary maps M for
both cyclic frequencies are calculated and score maps are derived
3. Application to simulated data
upon them. The score maps are presented in Fig. 4a and b. One
can notice some noise in these maps, i.e. above 1500 Hz and below
Firstly, the methodology introduced in Section 2 is applied to
2000 Hz for 5 Hz and 17 Hz cyclic frequencies, respectively. This
the simulated data. Such application gives a possibility to check
noise is not related to the periodic amplitude modulation, since
if the algorithm performs appropriately for the entirely known sig-
there are no clearly visible lines along the frequency axis in these
nal. We would like to analyzed three cases. In particular, first one
frequency bands. Time-varying filter coefficients are represented
consists of two different components corresponded to the fault.
by two-dimensional maps in Figs. 4c and d. Recall that these are
In order to test the performance in case of variable period of the
raw spectrograms multiplied element-wise by score maps. One
damage the signal with jitter is simulated. Finally, in the last case
can notice that the random noise in score maps does not influence
the signal without any cyclic component is analyzed with proposed
the filter characteristics. Non-zero scores not related to periodic
method. It would inform how the proposed method is sensitive for
modulations result in low filter coefficient, while multiplied by
the false alarm.
the spectrogram value. On the other hand, high values of the score
map are related to periodic modulation and multiplication by the
3.1. Signal with multiple fault source spectrogram values results in a high value of the time-varying filter
coefficients. According to the results, the novel algorithm automat-
Let us analyze firstly the signal with multiple cyclic compo- ically indicates points in the spectrograms related to given fault
nents. It corresponds to the machine, in which two damages occur. frequencies.
One can be interested if the method is able to extract both cyclic Given two weighted spectrograms the inverse short-time Four-
components. The length of the signal is 2.5 s and the sampling fre- ier transform might be applied in order to return to the time
quency is 8192 Hz. The signal is presented in time and time- domain. In case of the simulated signal, there is a possibility to
frequency domain in Fig. 2. compare the raw components and the extracted time series repre-
There are two periodic impulse trains with modulation frequen- senting sources. The comparison is presented in Fig. 5. Clearly, the
cies 5 Hz and 17 Hz, respectively, and random noise added to each extracted signal components are similar to simulated. The
pulse train. Corresponding time plots are presented in Fig. 3, where impulses reveal in the same time points, with similar amplitudes
periodic impulses can be observed. These components represent of corresponding impulses. Moreover, the noise level is lower for
local fault, which can occur in a rotating machine. The carrier fre- extracted components, thus the impulses are even more visible.
quency bands are 700–1300 Hz and 2300–3200 Hz for modulation Noise level in Fig. 5a is about 0.3 while in the extracted signal it
frequencies 5 Hz and 17 Hz, respectively. The former amplitudes is close to 0.2. In case of 17 Hz amplitude modulation the noise

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
4 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

Fig. 2. Time plot (a), spectrogram (b) and envelope spectrum (c) of the vibration simulated signal with two damages. The spectrogram parameters are as follows:
N w ¼ 250; Ov ¼ 96%.

level is close to 0.2 and 0.1 for raw and extracted signal, respec- 1.3 Hz and the sampling frequency is 8192 Hz. One cyclic pulse
tively. Presented results for the simulated data sample confirm that train with modulation frequency 5 Hz is added and it has a jitter
the proposed methodology can be applied for cyclic source extrac- effect. The carrier frequency bands are 700–1300 Hz. The ampli-
tion from the raw signal that consists of two pulse trains, each with tude of the impulses are around 1. Furthermore, the Gaussian ran-
different carrier and cyclic frequencies. In Fig. 6 the envelope spec- dom noise with mean 0 and standard deviation 4=3. The waveform
tra for each analyzed signal component and the corresponding of the signal is presented in Fig. 7a. The impulses are barely visible
extracted source signals are provided. One can observe that in both on the time-frequency map (Fig. 7b). Furthermore, on the envelope
cases several harmonics related to fault frequencies are indicated. spectrum the harmonics correspond to fault frequency cannot be
This confirms that the extracted signals follow the cyclic modula- observed (Fig. 7c). Thus, it is a complicated signal with invisible
tion from the raw signals. It is wort mentioning that in the raw sig- fault component, where the standard envelope methods for fault
nals only the harmonics corresponding to cyclic frequency 17 Hz detection fail. Therefore, we would like to apply proposed method
were detected (Fig. 2c). for cyclic source extraction. The extracted cyclic normalized signal
component is presented in Fig. 8a. One can observe that the level of
3.2. Signal representing bearing jitter effect noise is smaller and the impulses are clearly detectable. As it was
expected on the envelope spectrum the harmonics related to the
One can be interested how the proposed method is performing fault frequency are observed, namely 7 harmonics are visible. It
in case of small variation of fault frequency. In particular, we would is shown that the proposed algorithm is performing well in case
like to analyze the signal in which the period of the fault occur- of the signal with slightly varying fault frequency. Obviously, the
rence is not constant. Therefore, the signal, which resemble the results depend on the level of fault frequency variances. However,
bearing with jitter is simulated. The rotational speed is equal to in case of the reasonable changes of the period of fault occurrence

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx 5

Fig. 3. Time plots of the normalized impulse trains with modulation frequencies 5 Hz (a) and 17 Hz (b).

Fig. 4. Score matrices for fault frequency 5 Hz (a) and fault frequency 17 Hz (b). The weighted spectrogram (time-varying filter coefficients) for fault frequency 5 Hz (c) and
fault frequency 17 Hz (d).

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
6 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

Fig. 5. The normalized components of the simulated signal with modulation frequency 5 Hz (a) and 17 Hz (b). The normalized extracted signal from weighted spectrogram
for fault frequency 5 Hz (c) and 17 Hz(d).

Fig. 6. The envelope spectrum of the simulated signal with modulation frequency 5 Hz (a) and 17 Hz (b). The extracted signal envelope spectrum from weighted spectrogram
for fault frequency 5 Hz (c) and 17 Hz(d).

the method is able to extract the signal component related to the signal after source extraction method application is presented in
fault. Fig. 10a. One can observe that it is similar to signal before the pro-
cessing (Fig. 9c). In the envelope spectrum the harmonics related to
the modulation frequency 17 Hz are not observed (Fig. 10b). There-
3.3. Signal without cyclic component
fore, the proposed method do not create the false alarms. It means
that it can be applied to the machine without the local fault and
Finally, we would like to test the performance of the method for
the healthy signal. Indeed, the signal without any cyclic compo- the extracted signal would not contain the cyclic impulses. Pre-
sented analysis on the simulated signal ensure that the method
nent is simulated. The length is equal to 2.5 s and the sampling fre-
quency is 8192 Hz. The signal consists of background Gaussian is powerful in case of detection the cyclic impulses.
random noise with mean zero and variance equal to 0.2. Further-
more, in the carrier frequency bands 700–1300 Hz and 2300– 4. Application to real data
3200 Hz additional random noise with mean zero and variance
0.2 is added. Therefore, it is signal like in Section 3.1 with ampli- Once the performance of the algorithm was tested on the sim-
tude of impulses equals to 0. In the analysis the proposed cyclic ulated data, it can be applied to the real signal. In this section
source extraction method was tested for the modulation frequency results of the methodology application to the vibration data
17 Hz. It is expected that as a result, the signal without fault pat- recorded on the belt conveyor gearbox are presented. Analyzed
tern is obtained. It would ensure that the proposed method do machine is located in the underground copper ore mine, thus it
not create the cyclic impulses. In Fig. 9a the signal without cyclic works in harsh environment. Kinematic scheme of the analyzed
component is presented. Furthermore, the spectrogram is shown gearbox is presented in Fig. 11. The gearbox is equipped with
in Fig. 9b, where no impulses are visible. In the envelope spectrum two stages - the first bevel pair is followed by the spur pair. The
no harmonics of modulation frequency 17 Hz are observed. The electric engine rotor speed is around 996 rpm, thus the first shaft

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx 7

Fig. 7. Time plot (a), spectrogram (b) and envelope spectrum (c) of the vibration simulated signal with bearing jitter effect. The spectrogram parameters are as follows:
N w ¼ 250; Ov ¼ 96%.

rotates with the frequency close to 16.61 Hz. Variability of the quencies are presented in Fig. 14a and b. Some barely visible
rotational speed does not exceed 1.5 rpm. The rotational fre- periodic patterns might be noticed in both figures. The pattern
quency of the second shaft is 4.2 Hz, since the first stage ratio is related to 4.2 Hz is located in carrier frequency band lower than
equal to 91:23. The vibration signal was acquired on three different 1000 Hz. The second pattern reveals in almost every frequency
spots of the machine (Fig. 12a). Moreover, the rotational speed was band, including the lowest frequencies. Thus, this signal is more
also acquired in order to ensure that it is almost constant during challenging than the simulated data, since in a single frequency
the experiment. Sampling frequency is 17066 Hz and the signal band two periodic modulations occur and one of them reveals in
duration is equal to 5 s. The exemplary local fault in the gear is a relatively narrow frequency band. Moreover, two different
illustrated in Fig. 12b. phases can be detected in case of 4.2 Hz modulation frequency
The signal reveals two faults in the machine related to the rotat- (see arrows in Fig. 14a), i.e. two different components with modu-
ing speed of first and second shaft. The signal in time domain is lation frequency close to 4.2 Hz can be noticed. The related carrier
presented in Fig. 13a. The raw spectrogram (Fig. 13b) reveals some frequencies are 0–400 Hz and 500–700 Hz, respectively. Such phe-
frequency bands with high energy. These are related to gear mesh nomenon might be caused by two different local faults occurring
frequencies (379.5 Hz and 183.3 Hz for first and second stage, on the second shaft, for instance one on the bevel gear and second
respectively) and their harmonics. There are lots of wideband exci- on the spur gear. Therefore, carrier frequency bands might be dif-
tations, although it is difficult to state if they are related to any cyc- ferent due to different transmission path (Fig. 15).
lic pattern. Thus, it is worth to apply the proposed source Weighted spectrograms are illustrated in Figs. 14c and d. One
extraction algorithm. can observe that the level of background noise is relatively small
As in the simulated data case, two fault frequencies are consid- and the periodic excitations are with relatively high amplitudes.
ered, namely, 4.2 Hz and 16.61 Hz. Score maps for two fault fre- In order to transform weighted spectrogram to time domain the

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
8 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

Fig. 8. Time plot (a) and envelope spectrum (b) of the extracted cyclic normalized signal component with bearing jitter effect.

Fig. 9. Time plot (a), spectrogram (b) and envelope spectrum (c) of the vibration simulated signal without the fault component. The spectrogram parameters are as follows:
N w ¼ 250; Ov ¼ 96%.

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx 9

Fig. 10. Time plot (a) and envelope spectrum (b) of the extracted cyclic normalized signal component for the signal without the fault component.

Fig. 11. Kinematic scheme of the analyzed gearbox. The number of teeth in each stage: z1 ¼ 23; z2 ¼ 91; z3 ¼ 44 and z4 ¼ 155.

Fig. 12. The data acquisition system during the experiment (a) and exemplary local damage in the gear (b).

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
10 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

Fig. 13. Time plot (a), spectrogram (b) and envelope spectrum (c) of the real signal recorded on the belt conveyor’s gearbox. The spectrogram parameters are as follows:
N w ¼ 250; Ov ¼ 96%.

inverse short-time Fourier transform can be applied. Therefore, the modulations of 16.61 Hz. Hence, even such sources might be sepa-
cyclic component is extracted. The results are presented in Fig. 16a rated and damage detection can be performed with introduced
and b. Two different in phase pulse trains with modulation fre- novel method.
quency 4.2 Hz might not be clearly visible, since the corresponding
score values are relatively low. Excitations related to the second
fault are clearly visible. Given the extracted components, envelope 5. Conclusions
analysis might be performed in order to check periodicity of the
extracted signals. According to Fig. 16c more than 10 harmonics The damage detection in rotating machinery is one of the key
of 4.2 Hz might be noticed which corresponds to impulsive charac- problems in today’s industry. It is expected from machines to work
ter of the source signal. It is worth to notice that in the carrier fre- without unnecessary breaks. Thus, the analysis of the vibration sig-
quency band 0–1000 Hz both modulation frequencies are revealed. nal with multiple components is needed. Especially, in the mining
Moreover, filter coefficients related to 16.61 Hz are the highest in industry, where the machines operate in harsh environment and
the band from 0 Hz up to 2000 Hz, among both filter characteristics recorded signal can be influenced by many different effects. In this
- it ensures that there are significant local maxima with related article we proposed a novel method for cyclic source component
range. Nevertheless, the signal filtered using the weighted spectro- extraction. Introduced methodology involves time-frequency
gram calculated for 4.2 Hz does not contain significant amplitude decomposition and a time-varying filter with periodical coeffi-

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx 11

Fig. 14. Score matrices for fault frequency 4.2 Hz (a) and 16.61 Hz (b). The weighted spectrograms in the log scale for fault frequency 4.2 Hz (c) and 16.61 Hz (d).

Fig. 15. The zoomed score matrix for fault frequency 4.2 Hz (a) and zoomed weighted spectrogram in the log scale for fault frequency 4.2 Hz (b).

cients is designed using local maxima. The performance of the pro- each other. Even in such a complex case the proposed algorithm
posed algorithm was tested on the signals containing two sources is able to properly extract the signal sources. Furthermore, for
with periodic excitations. In the simulated signal the cyclic compo- the modulation frequency 4.2 Hz two local faults are detected,
nents are with different carrier and modulation frequencies. Fur- since impulses posses different phase. In order to check the sensi-
thermore, the method is applied to the signal with jitter, where tivity for the false alarm, the algorithm was applied to the simu-
the fault frequency is slightly varying. It is shown that in such case lated signal, which does not contain the cyclic component.
proposed method is able to extract information about the cyclic Interestingly, the method can be applied to healthy machine an.
component. The more complicated case is presented in the real sig- The approach is promising and in the future studies its properties
nal. The higher modulation frequency is close to multiple of the in case of many local faults, different noise types and excitations
lower one. Moreover, the carriers of cyclic components overlap with varying cycle could be analyzed.

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013
12 P. Kruczek et al. / Applied Acoustics xxx (2017) xxx–xxx

Fig. 16. The normalized extracted signals from weighted spectrograms for fault frequency 4.2 Hz (a) and 16.61 Hz (b). The envelope spectra of the extracted signals for fault
frequency 4.2 Hz (c) and 16.61 Hz (d).

Acknowledgment [16] Wylomanska A, Zimroz R, Janczura J, Obuchowski J. Impulsive noise

cancellation method for copper ore crusher vibration signals enhancement.
IEEE Trans Indust Electron 2016;63(9):5612–21.
This work is partially (P. Kruczek, J. Obuchowski, R. Zimroz) [17] Obuchowski J, Zimroz R, Wylomanska A. Blind equalization using combined
supported by the Framework Programme for Research and Innova- skewness–kurtosis criterion for gearbox vibration enhancement.
Measurement 2016;88:34–44.
tion Horizon 2020 under grant agreement n. 636834 (DISIRE -
[18] McDonald ZQ, Geoff, Zuo M. Maximum correlated kurtosis deconvolution and
Integrated Process Control based on Distributed In-Situ Sensors application on gear tooth chip fault detection. Mech Syst Signal Process
into Raw Material and Energy Feedstock). 2012;33:237–55.
[19] Paajarvi P, LeBlanc J. Online adaptive blind deconvolution based on third-order
moments. IEEE Signal Process Lett 2005;12(12):863–6.
[20] Endo H, Randall R. Enhancement of autoregressive model based gear tooth
fault detection technique by the use of minimum entropy deconvolution filter.
References Mech Syst Signal Process 2007;21(2):906–19.
[21] Sawalhi N, Randall R, Endo H. The enhancement of fault detection and diagnosis
[1] Randall R. Vibration-based condition monitoring: industrial, aerospace and in rolling element bearings using minimum entropy deconvolution combined
automotive applications. John Wiley & Sons; 2011. with spectral kurtosis. Mech Syst Signal Process 2007;21(6):2616–33.
[2] Randall R, Antoni J. Rolling element bearing diagnostics-a tutorial. Mech Syst [22] Zak G, Obuchowski J, Wylomanska A, Zimroz R. Application of ARMA
Signal Process 2011;25(2):485–520. modelling and alpha-stable distribution for local damage detection in
[3] Obuchowski J, Wylomanska A, Zimroz R. Recent developments in vibration bearings. Diagnostyka 2014;15(3):3–10.
based diagnostics of gear and bearings used in belt conveyors. Appl Mech [23] Zak G, Wylomanska A, Zimroz R. Data-driven vibration signal filtering
Mater 2014;683:171–6. procedure based on the a-stable distribution. J Vibroeng 2016;18(2):826–37.
[4] Bartelmus W. Object and operation supported maintenance for mining [24] Wylomanska A, Zak G, Kruczek P, Zimroz R. Application of tempered stable
equipment. Mining Sci 2014;21:7–21. distribution for selection of optimal frequency band in gearbox local damage
[5] Javorskyj I, Kravets I, Matsko I, Yuzefovych R. Periodically correlated random detection. Appl Acoust 2016. http://dx.doi.org/10.1016/j.
processes: application in early diagnostics of mechanical systems. Mech Syst apacoust.2016.11.008.
Signal Process 2017;83:406–38. [25] Bonnardot F, Randall R, Guillet F. Extraction of second-order cyclostationary
[6] Stefaniak P, Zimroz R, Bartelmus W, Hardygra M. Computerised decision- sources-application to vibration analysis. Mech Syst Signal Process 2005;19
making support system based on data fusion for machinery system’s (6):1230–44.
management and maintenance. Appl Mech Mater 2014;683:108–13. [26] Boustany R, Antoni J. A subspace method for the blind extraction of a
[7] Makowski R, Zimroz R. New techniques of local damage detection in cyclostationary source: application to rolling element bearing diagnostics.
machinery based on stochastic modelling using adaptive schur filter. Appl Mech Syst Signal Process 2005;19(6):1245–59.
Acoust 2014;77:130–7. [27] Urbanek J, Barszcz T, Straczkiewicz M, Jablonski A. Normalization of vibration
[8] Li Z, Yan X, Tian Z, Yuan C, Peng Z, Li L. Blind vibration component separation signals generated under highly varying speed and load with application to
and nonlinear feature extraction applied to the nonstationary vibration signals signal separation. Mech Syst Signal Process 2017;82:13–31.
for the gearbox multi-fault diagnosis. Measurement 2013;46(1):259–71. [28] Obuchowski J, Zimroz R, Wylomanska A. Identification of cyclic components in
[9] Guo Y, Wu X, Na J, Fung R-F. Simultaneous faults identification of rolling presence of non-gaussian noise–application to crusher bearings damage
element bearings and gears by combining kurtogram and independent detection. J Vibroeng 2015;17:1242–52.
component analysis. J Vibroeng 2015;17(3):1341–50. [29] Kruczek P, Wylomanska A, Teuerle M, Gajda J. The modified yule-walker
[10] Wodecki J, Stefaniak P, Obuchowski J, Wylomanska A, Zimroz R. Combination method for a-stable time series models. Physica A: Stat Mech Appl
of principal component analysis and time-frequency representations of 2017;469:588–603.
multichannel vibration data for gearbox fault detection. J Vibroeng 2016;18 [30] Wylomanska A. Spectral measures of parma sequences. J Time Ser Anal
(4):2167–75. 2008;29(1):1–13.
[11] Zhang B, Georgoulas G, Orchard M, Saxena A, Brown D, Vachtsevanos G, et al. [31] Antoni J. The spectral kurtosis: a useful tool for characterising non-stationary
Rolling element bearing feature extraction and anomaly detection based on signals. Mech Syst Signal Process 2006;20(2):282–307.
vibration monitoring. In: 2008 16th Mediterranean conference on control and [32] Barszcz T, Jablonski A. A novel method for the optimal band selection for
automation. IEEE; 2008. p. 1792–7. vibration signal demodulation and comparison with the kurtogram. Mech Syst
[12] Popiolek K, Pawlik P. Diagnosing the technical condition of planetary gearbox Signal Process 2011;25(1):431–51.
using the artificial neural network based on analysis of non-stationary signals. [33] Antoni J. The infogram: entropic evidence of the signature of repetitive
Diagnostyka 2016;17(2):57–64. transients. Mech Syst Signal Process 2016;74:73–94.
[13] Bartkowiak A, Zimroz R. Dimensionality reduction via variables selection– [34] Obuchowski J, Wylomanska A, Zimroz R. Selection of informative frequency
linear and nonlinear approaches with application to vibration-based condition band in local damage detection in rotating machinery. Mech Syst Signal
monitoring of planetary gearbox. Appl Acoust 2014;77:169–77. Process 2014;48(1):138–52.
[14] Feng Z, Zuo MJ. Vibration signal models for fault diagnosis of planetary [35] Boashash B. Time–frequency signal analysis and processing: a comprehensive
gearboxes. J Sound Vib 2012;331(22):4919–39. reference. Academic Press; 2015.
[15] Borghesani P, Shahriar MR. Cyclostationary analysis with logarithmic variance [36] Dutoit T, Marques F. Applied signal processing: a MATLABTM-based proof of
stabilisation. Mech Syst Signal Process 2016;70:51–72. concept. Springer Science & Business Media; 2010.

Please cite this article in press as: Kruczek P et al. Cyclic sources extraction from complex multiple-component vibration signal via periodically time vary-
ing filter. Appl Acoust (2017), http://dx.doi.org/10.1016/j.apacoust.2017.05.013