Single-channel source separation using EEMD and

Bounded Component Analysis in rotating machines

C.A. Aguirre-Echeverry, E.F. Sierra-Alonso, and G. Castellanos-Dominguez*
Universidad Nacional de Colombia, Signal Processing and Recognition Group
{caaguirreec,efsierraa,cgcastellanosd}@unal.edu.co

Abstract. Among condition monitoring approaches, source separation is a prominent area which has been studied in the last decade, but only with an statistical
independence framework, which is restricted to most of the mechanic dynamics
into a rotating machine. Furthermore, in real industry applications, the accessibility for installing several mechanical vibration transducers is limited which
entail few measurement channels, or even just one. Several methodologies study
the single channel approach as SSA, WAVELET, and EMD. The proposed paper
introduces a novel methodology for single channel source separation in rotating machines using jointly an enhancement EMD for pseudo channel generation
and Bounded Component Analysis for source separation with a wider approach
than ICA. Both synthetic and real signals are tested showing the improvement of
the proposed methodology and several comparison are made with other standard
current techniques. The current methodology proves to be suitable for rotating
machines source separation and its application for other topics as classification,
or feature extraction is promising.
Keywords: Single channel, Bounded Component Analysis, Ensemble Empirical
Mode Decomposition

2
3
4
5
6
7
8
9
10
11
12

1 Introduction
Nowadays, condition monitoring of rotating machines becomes an important issue since
it significantly allows reducing the maintenance costs in the industry. Among condition monitoring techniques lately developed for failure detection of rotating machines,
source separation is an encouraging area that allows accurate analysis of multiple mixed
components from which failure condition may be better inferred. The Independent
Component Analysis (ICA) is the most used source separation technique where statistical independence is assumed in relationship among mixed components. By instance,
authors in [6, 4, 8, 10, 13] discuss the use of ICA in machinery fault diagnosis.
There are several ICA approaches that differs from each other in terms of the involved optimization framework. Yet, their application is somehow restrained because
the presence of signals that are clearly correlated or even dependent each other. This

Sistema de monitoreo de condicion en turbinas de generacion electrica cod Hermes 23079,

cod QUIPU 20101008725 de la convocatoria: Fomento de una cultura de la innovacion a la
Universidad Nacional de Colombia 2013-2015- sedes Manizales y Palmira

35

situation usually happens in rotating machines where the joint dynamics of mechanical components and the latent dependency between vibration signals are frequently
present [12]. Besides, most of machinery provides reduced accessibility for installing
several mechanical vibration transducers on the machine that is limited either by the
physical space or the high cost of sensor networks. In consequence, measurement of
the vibration signals must be carried out by means of a single sensor. To cope with this
issue, the source separation framework relying on correlated signals, termed Bounded
Component Analysis (BCA), has been developed [3]. However, the baseline BCA technique assumes multichannel vibration recordings and is not suitable for decomposition
of single-channel signals. To overcome this drawback, decomposition of the singlechannel signals is to be considered like singular spectrum analysis, wavelet transform,
and Empirical Mode Decomposition (EMD) [2, 11]. Hence, the EMD offers advantages
over other techniques since it usually does not require any parameter tuning and preserves the quasi-orthogonality condition among decomposed components.
We introduce a novel methodology of Bounded Component Analysis that is based
on the EMD decomposition for dealing with source separation of single channel vibration signals. Also, the Ensemble version of EMD (termed EEMD) is introduced to
alleviate the mode mixing problem of the carried out decomposition [7]. The current
methodology is tested on both synthetic and real signals, and compared with some classic ICA algorithms as fastICA, SOBI, JADE and BEFICA. Obtained results on the
tested real-world database shows that the proposed methodology successfully estimates
fault related sources even with high levels of noise being competitive with other methods reported in the literature.

36

2 Methods

37

2.1 Ensemble Empirical Mode Decomposition of Single Channel Signals

13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

38
39
40
41
42
43
44
45

Generally, the single channel signal, measured by just one vibration transducer, makes
the linear system solution in (2) be an ill-conditioned problem. To this end, we make
use of the Ensemble Empirical Mode Decomposition that is an adaptive method decomposing time series into a set of the intrinsic mode functions (IMF). The IMF set must
account for: i) the amount of local extremes and the zero crossing differs at most by
one, ii) at any point the mean value between the superior envelope defined by the local
maxima and the inferior envelope defined by the local minima is zero. Thus, a given
time series x(t), with tT, can be represented by EMD in the form:
X
x(t) =
ck (t) + r(t), tT
(1)
kK

46
47
48
49
50
51
52

where {
ck (t)} is the extracted IMF set holding K components and r(t) is the remainder
term. The first IMF is related to the highest frequency, whereas the last one to the lowest.
However, EMD faces the mode mixing problem because of reached low orthogonality between neighboring IMFs. This issue is overcome by the use of the ensemble
empirical mode decomposition (EEMD) that takes advantage of the additive white gaussian noise (AWGN) cancelation property within dyadic filter bank EMD structures [7].
The EEMD is sequentially carried out as follows:

53
54
55
56
57
58
59
60
61

62

63
64
65
66
67
68
69
70

1. An input time series x(t) is contaminated with AWGN as much as J times, i.e.,
xj (t)=x(t) + j (t), tT, j=1, . . . , J, being j (t), each j-th trajectory of the
randomly generated AWGN,
2. Afterwards, obtained xj (t) is decomposed, using the conventional EMD, into the
corresponding IMF set, {
ck,j (t) : k=1, . . . , K}.
3. Lastly, an averaged version of ck (t) is obtained as: ck (t)=E {
ck,j :jJ} , tT,
where notation E {} stands for the expectation operator. Generally, J should be
large enough to cancel the AWGN since there is a directly proportional relation
between the standard deviation of the AWGN and the amount J.
2.2 Single-channel Bounded Component Analysis
Bounded component analysis is a blind and geometric technique for the linear decomposition that is grounded on the hypothesis of the compactness and cartesian decomposition of the convex support of the component signals. Unlike the ICA algorithms,
BCA does not need the independence assumption to work. So, it can extract successfully sources that are somehow dependent [5].
Provided the n linear mixture set {xn (t) : nIN}, that is measured by N transducers
producing the corresponding N independent source set {sn (t) : nIN}, then each linear
mixture can be expressed as follows:
x = As

71
72
73
74
75
76
77
78
79
80
81
82
83
84

(2)

where x=[x1 (t), . . . , xN (t)] , with xIRN 1 , is the vector observations or channels, s=[s1 (t), . . . , sN (t)] sIRN 1 , is the vector of sources, and the square matrix
AIRN N is the mixing model with elements aij , with i=1, . . . , N and j=i.
BCA solution lies in the fact that the normalized convex perimeter of the estimation
output is a suitable cost function for the successful extraction of the sources. Thus, one
can blindly estimate one of the sources by confining a particular combination of the observations into a convex set of minimum perimeter [3]. The convex bodies employed in
BCA are: i) the Bounding hyper-rectangle, z , that is the box defined by the cartesian
product of the support sets of the individual components. That is; the minimum volume
box containing all samples and aligning with the coordinate axes, ii) Principal hyperellipse, Z , that is the hyper-ellipse principal semi-axis directions are determined by
the eigenvectors of the covariance matrix and the principal semi-axis lengths are equal
to principal standard deviations, i.e., the square roots of the M eigenvalues z(k), with
kM, computed from the covariance matrix in the form:
Rz = W Rx W 1

85
86
87
88
89
90
91

where Rx IRN N is the covariance matrix associated to x, which satisfies Rx =

ARs A1 , and Rs IRN N is the covariance of the sources s, a more detailed description can be found in [5].
Nonetheless, after applying the EEMD methodology on measured single-channel
vibration signals, one may employ each extracted IMF ck (t) (see Eq. (1)) as linear
input mixture within the linear separation model (see Eq. (2)). Thus, the new model can
be rewritten as c=As.

92
93
94
95
96
97

Although the vector cIRN 1 contains the observed component set, the desired
output is the source vector sIRN 1 . Therefore, the needed linear solution is the form:
s=W c, where W =A1 . Now, the goal is to train the separator W IRN N based on
the available observations of mixtures {c1 , c2 , ..., cK } . To this end, it is performed maximization of the relative ratio between the volume of the hyper-ellipse and the volume
of the hyper-rectangle as given in [3]:
( p
)
det(Rz )
max {J1 (W )} = max
zK
z2 . . . R
z1 R
R
zi is the range operator defined for a vector zi = [zi (1), . . . , zi (M )] as:
where R
zi =
R

98

99
100

101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125

max zi (k)

k1,...,M

min

k1,...,M

zi (k)

3 Experimental setup
Validation of the proposed methodology for single-channel source separation is carried
out on synthetic and real signals.
Testing on real vibration signals: Database of the Case Western Reserve University [9]
holds vibration signals measured in the following test stand: a 2 hp motor, a torque
transducer/encoder, a dynamometer, and control electronics. The test bearings support
the motor shaft. Microcracks were introduced in different locations of selected rolling
bearing (drive end bearing or fan end bearing) by using electro-discharge machining.
The ball, inner and outer races have a single point fault created by electro-discharge machining. Bearings with microcracks were reinstalled into the test motor under different
engine speeds to simulate diverse faulty conditions. Vibration signals were sampled by
accelerometers that were attached to the housing with magnetic bases. The fault size is
0.007 inches in diameter and 0.011 inches in depth for ball, inner and outer races. The
shaft rotating speed is 1797 RPM (29.95 Hz). The data were collected at 12 kHz [4].
Here, a single channel is formed by mixing the outer race, inner race and ball failure
of drive end bearing. The acquisition time was 10 s. For the bearing database, threedifferent failures corresponding with the outer race, inner race, and ball bearing failures are mixed in a single channel signal. The theoretical characteristic frequencies for
each failure are 104.56 Hz for Outer race failure, 157.94, Hz for Inner race failure, and
137.48 Hz for ball bearing failure.
The primary goal of condition monitoring is to isolate different failures of bearings in a rotating machine using a single channel recording. To this end, the proposed
methodology is focused on the generation of pseudo channels with meaningful information. We make use of the EEMD decomposition because of its adaptative nature and
high performance in decomposition of rotating machinery signals [6]. Still, we also select the most meaningful IMFs by using the principal component analysis applied to
the whole of the decomposition components. Then, the envelope analysis is performed
for extracting relevant features of the mixture in narrow band signals through Hilbert

5
Amp

Amp

5
0
5

0.05

0.1
Time(s)

0.15

0.2

0.05

0.1
Time(s)

0.15

0.2

0.4
Amp

Amp

2
1
0

0
5

1000

2000
3000
Frequency(Hz)

4000

5000

0.2
0

1000

2000
3000
4000
Frequency(Hz)

5000

6000

(a) Synthetic
(b) Real
Fig. 1. Examples of testing synthetic and real single-channel vibration signals.
126
127
128
129
130
131
132
133
134

transform. Finally, the bounded component analysis is carried out in order to separate
the failures.
For the comparison sake, some conventional ICA algorithms are carried out: fastICA, SOBI, JADE and BEFICA . Several measures are used for testing the separation
results named correlation, mutual information (MI) and spectral correlation measure
(SCM) based on cross power spectral density [1]. As recommended in [7], the parameters employed for EEMD are 0.2 for white gaussian noise, which correspond to signal
noise relation of 6.9 dB and 500 iterations. Regarding BCA, 1000 iterations are implemented.

148

Testing on synthetic signals: A set of simulated vibration signals, provided in [?], is

used. Examples of both, synthetic and real, signals and their frequency representation
are shown in the Fig. 1.
In the first experiment, the separation is performed for synthetic signals without
noise and the results are shown in the Fig. 2. In the second experiment, white Gaussian noise injected to the synthetic single channel. Part of the results are shown in the
Fig. 3. Regarding the real bearing signals, using the proposed methodology, the separated sources are evaluated with theoretical frequency failures. White Gaussian noise is
injected in the single channel signal in order to validate the stability of the separation
techniques. Results can be seen in Fig. 4.
Although the Fig. 3(a), Fig. 3(b), and Fig. 3(c) only shows the separation behaviour
under several levels of noise for one source, the dynamic is the same for the rest of
sources. For each source, mostly separation algorithms show similar high separation
levels, proving EEMD as a valid methodology for pseudo-channel generation.

149

4 Discussion

135
136
137
138
139
140
141
142
143
144
145
146
147

150
151
152
153
154
155
156
157

Initially, to test EEMD ability of canceling white noise, the vibration signal is corrupted
adding different levels of AWGN within the interval [10, 10] dB. As seen in Fig. 4
showing values of the estimated relative error; the enhancing EEMD decomposition
technique (see Fig. 4(b)) outperforms the baseline EMD (see Fig. 4(a)). It is worth
noting that each decomposition influences differently of the achieved accuracy. Namely,
the relative error of the EMD behaves monotonically. That is, the lower the SNR the
highest the error. By contrast, there are some intervals of negative SNR values where
the ensemble principle may improve the decomposition performance.

100

80

0.8
fastICA
SOBI
JADE
BEFICA
BCA
BCAcon
VCA

60

40

fastICA
SOBI
JADE
BEFICA
BCA
BCAcon
VCA

0.6

0.4

20

0.2

(a) Correlation

(b) MI

3500
3000
2500

fastICA
SOBI
JADE
BEFICA
BCA
BCAcon
VCA

2000
1500
1000
500
0

(c) SCM
Fig. 2. Correlation, mutual information and SCM performance without noise of synthetic signals.
Correlation source 3

Mutual information source 3

0.6

fastICA
SOBI
JADE
BEFICA
BCA
BCAconv
VCA

0.8
Separation

Separation

fastICA
SOBI
JADE
BEFICA
BCA
BCAconv
VCA

0.8

0.4
0.2

0.6
0.4
0.2

0
10

0
SNR(dB)

0
10

10

0
SNR(dB)

(a) Correlation

10

(b) MI
Spectral correlation measure source 3

3000

Separation

2500
fastICA
SOBI
JADE
BEFICA
BCA
BCAconv
VCA

2000
1500
1000
500
0
10

0
SNR(dB)

10

(c) SCM
Fig. 3. Correlation, mutual information and SCM performance under several levels of noise.
0.35

0.35
Bearing ball
Outer race
Inner race

0.3

Error(%)

Error(%)

0.25

0.2
0.15

0.2
0.15

0.1

0.1

0.05

0.05

0
10

Bearing ball
Outer race
Inner race

0.3

0.25

0
SNR(dB)

10

0
10

0
SNR(dB)

10

(a) EMD
(b) EEMD
Fig. 4. Frequency failure error estimation with the proposed methodology under different levels
of noise
158
159
160
161
162
163
164

However, the noise cancellation (ensemble procedure) of EEMD affects the contaminated vibration signals differently per class. For the inner race class the noise cancellation works through all the studied noise levels, due to it is characterized by a high
energy cyclic frequency. Also, atypical behavior exists for outer race and balls classes.
In the case of the outer race, there is a valley close to 0 dB. From this point to the right,
the estimation presents an increasing behavior, suggesting a convergence error value for
the ensemble procedure. Starting from 0 dB to 10 dB, it is shown a crescent estimation

197

that is directly proportional to the added noise. Likewise, balls class presents a valley
in the interval [2, 8] dB. From this range ahead, it also shows an increasing behavior
suggesting a convergence value. Nevertheless, the expected increasing curve directly
proportional to the noise it is not presented mainly in a decreasing interval from 8 dB
to 10 dB. As a result, each class reacts differently to the noise cancellation achieved
by the ensemble procedure. This aspect prompts authors to plan, as future work, for
deeply considering the influence of the ensemble method to compute the optimal decomposition parameters (number of ensembles and noise standard deviation).
Among the methods tested for source separation in synthetic signals, BCA shows
higher performance than other methods like fastICA, SOBI or JADE due the last ones
tend to attenuate the separated sources. BEFICA, as efficient fastICA algorithm designed for stationary non-Gaussian sources [8], shows similar behaviour as BCA. This
situation may be explained because of the nature of the present synthetic and real
sources, despite the bearings signals are cataloged as non-stationary nature signals [12].
However, BCA does not search for independent sources and allows a more general
framework for source separation than ICA because it does not depend on the statistical constraints. In that sense, BCA offers a new and more general framework to separate mechanical sources from a single channel. For every source of the synthetic single
channel, separability measures are tested under different levels of noise. As expected,
correlation goes down as SNR is decreased. Regardless mutual information, similar
behaviour is obtained. However, spectral correlation measure maintains higher performance than the other measures even with different SNR levels due this measure works
with cumulative energy that allows higher tolerance to noise and a new interpretation
of similarity between signals.
As real mechanical failure bearing signals can be characterized with a singular frequency magnitude with high energy, the separability is measure comparing the estimation error of the maximum spectrum of each source with the theoretic failure frequency.
Also, Gaussian noise is added, and the error is measured using proposed methodology.
The results show how even in the worst case, the estimation error is no more than 30%.
For the hypothetical case when no noise is assumed, the error estimation of the outer
race and inner race failure is less than 3%. For the erratic behaviour of ball bearing,
reached error is less than 8% showing that the proposed methodology provides high
separability among considered classes.

198

5 Conclusions

165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196

199
200
201
202
203
204
205
206
207

The main contribution of the current paper is the application of jointly bounded component analysis and EEMD to rotating machine source separation under pseudo-stationary
conditions and single channel measurement. Bounded component analysis adds a new
framework for source separability. Moreover, this technique enables separation of independent and dependent (even correlated) sources based on geometric framework unlike
ICA which uses a statistic independence criteria, which is a particular case of BCA.
The current paper enhances BCA application in rotating machines, where the mechanical dependence between sources is latent. The use of BCA in rotating machines is
novel, as its use with single channel restriction approach. The proposed methodology

216

remarks the noise tolerance given the natural filtering of EEMD as pseudo-channel generator and a high noise tolerance of BCA comparing some ICA techniques. Together,
EEMD and BCA offers a new framework for mechanical source separation allowing
blind assumptions about the sources. The constraints of the proposed methodology lie
in the computational cost and the boundedness of the a priori sources. However, the advantages of blind non-parametric analysis with a new separation framework makes the
current methodology suitable for condition monitoring and prognosis. As Future work,
the BCA mapping function can be changed in order to optimize the geometric space
using kernel functions.

217

References

208
209
210
211
212
213
214
215

218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248

1. Spectral correlation measure for selecting intrinsic mode functions. In: Bayro-Corrochano,
E., Hancock, E. (eds.) Progress in Pattern Recognition, Image Analysis, Computer Vision,
and Applications, Lecture Notes in Computer Science, vol. 8827 (2014)
2. Amirat, Y., Benbouzid, M., Wang, T., Turri, S.: Performance analysis of an eemd-based
hilbert huang transform as a bearing failure detector in wind turbines. In: Green Energy,
2014 International Conference on. pp. 193198. IEEE (2014)
3. Cruces, S.: Bounded component analysis of linear mixtures: a criterion of minimum convex
perimeter. Signal Processing, IEEE Transactions on 58(4), 21412154 (2010)
4. Dong, S., Tang, B., Zhang, Y.: A repeated single-channel mechanical signal blind separation
method based on morphological filtering and singular value decomposition. Measurement
45(8), 20522063 (2012)
5. Erdogan, A.T.: A family of bounded component analysis algorithms. In: Acoustics, Speech
and Signal Processing (ICASSP), 2012 IEEE International Conference on. pp. 18811884.
IEEE (2012)
6. Guo, Y., Huang, S., Li, Y.: Single-mixture source separation using dimensionality reduction
of ensemble empirical mode decomposition and independent component analysis. Circuits,
Systems, and Signal Processing 31(6), 20472060 (2012)
7. Huang, N., Wu, Z.: Ensemble empirical mode decomposition: A noise-assisted data analysis
method advances in adaptive data analysis vol. 1 (2009)
8. Koldovsk`y, Z., Malek, J., Tichavsk`y, P., Deville, Y., Hosseini, S.: Blind separation of piecewise stationary non-gaussian sources. Signal Processing 89(12), 25702584 (2009)
9. Loparo, K.A.: Bearing data center, case western reserve university (2003)
10. Miao, F., Zhao, R.Z.: Application of independent component analysis in machine fault diagnosis. In: Advanced Materials Research. vol. 905, pp. 524527. Trans Tech Publ (2014)
11. Miao, Q., Wang, D., Pecht, M.: Rolling element bearing fault feature extraction using emdbased independent component analysis. In: Prognostics and Health Management (PHM),
2011 IEEE Conference on. pp. 16. IEEE (2011)
12. Randall, R.B., Antoni, J.: Rolling element bearing diagnostics a tutorial. Mechanical Systems and Signal Processing 25(2), 485520 (2011)
13. Wang, J., Gao, R.X., Yan, R.: Integration of eemd and ica for wind turbine gearbox diagnosis.
Wind Energy 17(5), 757773 (2014)