Mechanical Systems and Signal Processing 86 (2017) 75–85

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Complex variational mode decomposition for signal processing

applications
cross
⁎
Yanxue Wang , Fuyun Liu, Zhansi Jiang, Shuilong He, Qiuyun Mo
Guilin University of Electronic Technology, 541004 Guilin, PR China

A R T I C L E I N F O ABSTRACT

Keywords: Complex-valued signals occur in many areas of science and engineering and are thus of
Variational mode decomposition (VMD) fundamental interest. The complex variational mode decomposition (CVMD) is proposed as a
Complex-valued signals natural and a generic extension of the original VMD algorithm for the analysis of complex-valued
Complex variational mode decomposition data in this work. Moreover, the equivalent filter bank structure of the CVMD in the presence of
(CVMD)
white noise, and the effects of initialization of center frequency on the filter bank property are
Time-frequency analysis
Rubbing
both investigated via numerical experiments. Benefiting from the advantages of CVMD
algorithm, its bi-directional Hilbert time-frequency spectrum is developed as well, in which
the positive and negative frequency components are formulated on the positive and negative
frequency planes separately. Several applications in the real-world complex-valued signals
support the analysis.

1. Introduction

In modern disciplines, complex-valued signals are encountered in a wide variety of applications, such as structural health
monitoring, sensor array processing as well as biomedical sciences and physics. For the case of bivariate (or complex valued) data,
several extensions to traditional real-valued signal decomposition method have been developed so far. For example, bivariate EMD
(BEMD) has been proposed in [1], and it has been used to wind turbine condition monitoring in [2], because of its advantages for
information fusion. Motivated by the bivariate framework in the BEMD method, a complex local mean decomposition algorithm was
proposed in [3]. However, this class of complex extension methods can only be used on the assumption that complex signals are
proper or circular. A proper complex random variable is uncorrelated with its complex conjugate, and a circular complex random
variable has a probability distribution that is invariant under rotation in the complex plane [4]. However, there are many cases where
proper and circular random signals are very poor models of the underlying physics. Therefore, another class of complex-extension of
the EMD (CEMD) was introduced in [5], which cleverly used the relationship between the positive and negative frequency
components. Currently, the CEMD has been applied to perform the fusion of data from multiple and heterogeneous sources [6].
However, CEMD can not guarantee the same number of intrinsic mode functions (IMFs) across real and imaginary data channels,
which is a major requirement in real-world applications.
Variational mode decomposition (VMD) was recently proposed in [7] as an alternative to the empirical mode decomposition
(EMD) for the separation of composite real-valued time series into respective modes. It has been pointed out that VMD is
theoretically much better founded compared to the sequential iterative sifting of EMD, because VMD is based on a clear variational
model and the resulting minimization steps perform concurrent mode extraction in an intuitive way. It was also demonstrated in [7]
VMD over EMD has some advantages on tones separation and is less sensitive to noise and sampling. Based on these characteristics,

⁎
Corresponding author.
E-mail address: Yan.xue.wang@gmail.com (Y. Wang).

http://dx.doi.org/10.1016/j.ymssp.2016.09.032
Received 29 September 2015; Received in revised form 9 August 2016; Accepted 24 September 2016
0888-3270/ © 2016 Elsevier Ltd. All rights reserved.
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

VMD was applied to detect rub-impact fault of the rotor system [8]. Nevertheless, the VMD has been presently developed only for
real-valued data, its wider ranges of applications in signal processing and related fields are limited. Motivated by the work [5], a new
extension of the VMD designed to handle complex-valued time series is developed in this work. To further shed light on its
performance, the behavior of CVMD in the presence of white Gaussian noise is also analyzed, which will be of great benefit to
applications of the CVMD.
Numerous signals in science and engineering exhibit time-varying oscillatory behavior that is not possible to characterize
adequately by conventional Fourier analysis. Time-frequency analysis can identify the signal frequency components, reveals their
time variant features, and is an effective tool to extract machinery health information contained in nonstationary signals. Various
time–frequency (TF) analysis methods have been proposed and applied to machinery fault diagnosis [9]. Moreover, it should be
noted that interest in time-frequency analysis of multichannel data has also recently been growing with multivariate data driven
algorithms that directly exploit multichannel interdependencies [10] and [11]. A multivariate TF algorithm based on synchrosqueez-
ing transform that generated a compact TF representation of multichannel signals was developed in [12]. Motivated by the Hilbert-
Huang spectrum [13], the full Hilbert time-frequency spectrum of CVMD is thus developed in this work as well, in which the positive
and negative frequency components are formulated on the positive and negative frequency planes separately.
The rest of this work is organized as follows: Section 2 briefly recalls the VMD algorithm developed in [7]. The proposed complex
extension of the VMD and the corresponding filter bank property as well as its Hilbert spectrum are all proposed in Section 3. Three
applications based on the CVMD method and its Hilbert spectrum are addressed in Section 4. Conclusion is presented in Section 5.

2. Variational mode decomposition

The VMD can non-recursively decompose a real-valued multi-component signal f into a discrete number of quasi-orthogonal
band-limited sub-signals uk with the specific sparsity properties of its bandwidth in the spectral domain [7]. Each mode is compact
around a center pulsation ωk and its bandwidth is estimated using / 1 Gaussian smoothness of the shifted signal. For the
convenience of following discussions, let us commence by calling these modes obtained by VMD as band-limited IMFs (BLIMFs) in
this work. BLIMFs are different from IMFs defined in EMD technique according to the numbers of extrema and zerocrossings as well
as zero-mean binding conditions. VMD first turns the real valued mode uk into an analytic signal uk+ with single sided frequency
spectrum
⎛ j⎞
uk+(t )=⎜δ (t )+ ⎟ *uk (t )
⎝ πt ⎠ (1)
The single sided spectrum is then shifted down to 0-frequency baseband, and eventually the effective bandwidth of such a signal
is obtained using the L 2 -norm of the time derivative. Actually, the generalization in the proposed CVMD is to decompose the
complex signal into two single-sided frequency spectra using an ideal band-pass filter, which can be found in detail in the following
section. The VMD method is written as a constrained variational problem [7]:
⎧K ⎪
⎫ ⎪
K
min ⎨∑ ∂t [uk+(t ) e−jωk t ] 22 ⎬,s ubjectto ∑ uk (t )=f (t )
⎩
{uk },{ωk }⎪ ⎭
⎪

k =1 k =1 (2)
The constraint in Eq. (2) can be addressed by introducing a quadratic penalty and Lagrangian multipliers λ (t ). The augmented
Lagrangian is given as follows
K
3 ({uk }, {ωk }, λ )=α ∑ ∂t [uk+(t ) e−jωk t ] 2
2
k =1

K 2 K
+ f (t ) − ∑ uk (t ) + λ (t ), f (t )− ∑ uk (t )
k =1 2 k =1 (3)
where α denotes the balancing parameter of the “data-fidelity” constraint. The corresponding unconstrained problem in Eq. (3) is
then solved using the Alternate Direction Method of Multipliers (ADMM) [14]. All the modes with i < k gained from solutions in the
Fourier domain are updated in nature by Wiener filtering using a filter tuned to the current center frequency on the positive part of
the spectrum (i.e.,ω≥0 ), that is
λˆ (ω)
fˆ (ω) − ∑i < k uˆin +1 (ω) − ∑i > k uˆin (ω) + 2
uˆkn +1 (ω)=
1 + 2α (ω − ωkn )2 (4)
where the center frequency ωkn
is accordingly updated as the center of gravity of the corresponding mode’s power spectrum
uˆin+1 (ω)(ω ≥ 0). Due to the embedded Wiener filtering in the algorithm, VMD is much more robust for sampling and noise. The ωkn+1
is computed as follows
∞
∫0 ω uˆkn +1 (ω) 2 dω
ωkn +1= ∞
∫0 uˆkn +1 (ω) 2 dω (5)

76
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Center frequencies ωk of the modes are initialized in two ways in this work: the uniform distribution and zero. Generally, different
results may be achieved when a different initialization is applied. The parameter α is also specified with an appropriate value.
Extensive investigation of those parameters and their effects is beyond the scope of this paper, and interesting readers can refer to
[15].
The complete algorithm of VMD was referred to [7] for detail. It should be mentioned that VMD was only defined for real-valued
data. Complex-valued signals rise in many areas of science and engineering, such as mechanics, electromagnetics, optics, and
acoustics, and are thus of fundamental interest. Consequently, it would be highly beneficial to extend the conventional VMD to the
complex domain, in view of the rapidly increasing number of the applications of complex-valued signal processing.

3. Complex variational mode decomposition

3.1. CVMD algorithm

A new extension of the VMD destined to handle complex-valued time series is proposed in this subsection. Generally, a simple
way to extend a real-valued signal decomposition method to the complex domain would be to apply this method separately to the
real and imaginary parts of a complex-valued signal. However, the complex, bivariate time series with some existing mutual
information between the real and imaginary parts is mapped onto two independent real-valued univariate time series, where this
mutual information is lost [5]. In addition, BEMD can only be used for the time series, which have the relationship in Cartesian
coordinates [1].
For a real data, only positive frequencies of an analytic signal are required for the frequency-shifting and / 1-norm measure of
bandwidth in the VMD algorithm (described in Eq. (2)). However, for a given non-analytic complex-valued signal {x (t )}, its Fourier
transform X (e jω ) is not Hermitian. Thus, an extension of the VMD for the complex-valued data is proposed inspired by the technique
proposed in [5], which applies the relationship between the positive and negative frequency components of their spectra. The
corresponding positive and negative frequency components can be extracted using an ideal band-pass filter specified by
⎧1, 0 ≤ ω < π
H (e jω )=⎨
⎩ 0, −π ≤ω<0 (6)
This complex sharp band-pass filter can be well constructed using the digital filters designing methods, such as frequency
response masking [16]. A simple method to construct this filter in terms of FIR filter designing and shifting operation was adopted in
this work. Fig. 1 shows a designed band-pass filter which will be used in the following sections. We can find the frequency response
(see Fig. 1(b)) of this band-pass filter is well consistent with Eq. (6)..
Based on the above designed ideal band-pass filter, two analytic signals can then be generated, i.e, H (e jω ) X (e jω ) and
H (e jω ) X * (e−jω ) in frequency domain. Owing to well-known properties of the analytic signals, we can deal with their real parts,
only, of the analytic signal, without of loss information. Thus, two real signals can be deduced using Eqs. (7) and (8) below
x+(t ) = R {-−1(H (e jω ) X (e jω ))} (7)

x−(t ) = R {-−1(H (e jω ) X * (e−jω ))} (8)

where * denotes complex conjugate operation and R{∙} means the operator that extracts the real part of a complex function, and
-−1(∙) is the inverse Fourier operator. The original information can be captured by x+(t ) and x−(t ), because of analytic characteristics
of H (e jω ) X (e jω ) and H (e jω ) X * (e−jω ). The reconstruction of the original complex-valued signal is achieved by x+(t ) and x−(t ) as well as
their Hilbert transform pairs, which is written as follows
x (t ) = (x+(t ) + jℏ(x+(t )))+(x−(t )+jℏ(x−(t )))* (9)
1 +∞ x (u )
whereℏ(x )= π ∫
−∞ t − u
du represents the Hilbert transform using the Cauchy principal value integral.

Fig. 1. The comparison between the original and the recovered signal.

77
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Then, VMD is used to decompose two real signals x+(t ) and x−(t ), respectively. Unlike the case in CEMD [5], CVMD always
achieves the same number of the BLIMFs for a different analyzed signal x+(t ) and x−(t ), because the decomposition level is required
as a priori parameter in the VMD procedure. Suppose the decomposition level is N, VMD is separately adopted to each real signal
x+(t ) and x−(t ), which is written as
N
x+(t )= ∑ xi (t )
i =1 (10)

−1
x−(t )= ∑ x i (t )
i =− N (11)

where {xi (t )}iN=1

and {xi (t )}−1
denote sets of BLIMFs corresponding to x+(t ) and x−(t ), respectively. Thus, desirable information
i =−N
inherent in the modes can be captured using these decompositions.
Based on the Eqs. (9)–(11), the developed complex VMD for a complex time series x (t ) is finally written as
−N
x (t ) = ∑ z i (t )
i =− N , i ≠0 (12)

in which zi (t ) denotes the ith complex BLIMF which can be described below
⎧ xi (t ) + jyi (t ), i=1, …, N
z i (t ) = ⎨
⎩ (xi (t )+jyi (t )*, i=−N , …, −1 (13)

with yi =ℏ(xi ). As can be seen in the recovering process given in Eqs. (12) and (13), x+ and x− parts are separately coupled with their
respective Hilbert transform pairs. The coupling center frequencies for both x+ and x− parts is not considered in the proposed
method.
A non-analytic complex-valued signal is used to demonstrate the reconstruction of the original back from its two sub-part x+ and
x− (specified in Eq. (9)). Provided x (t )=0. 5 sin(2πf1 t )+j cos(2πf2 t ), we set f1 =30Hz , f2 =40Hz and t∈[0,0.512], individually.
Subsequently, the sub-part x+(t ) and x−(t ) can be achieved using Eqs. (7) and (8) based on our designated ideal band-pass filter.
The Hilbert transform pairs of x+(t ) and x−(t ) are also shown in Fig. 2(a) and (b), respectively. Spectra of the generated H (e jω ) X (e jω )
and H (e jω ) X * (e−jω ) are illustrated in Fig. 3(a), where their analytic characteristics have been well demonstrated. By rearranging the

Fig. 2. The impulsive and frequency response of the designed filter. Fig. 1(a) and (b) the decomposed components and their Hilbert transform, Fig. 1(c) and (d) the
real and image parts of the original signal and recovered signal.

78
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Fig. 3. (a) Two generated analytical signals, (b) spectrum of the original and the recovered signal.

real- and imaginary-part resulting from(x+(t )+jℏ(x+(t )))+(x−(t ) + jℏ(x−(t )))*, the final recovered real- and image-part are also
displayed in Fig. 2(c) and (d) (see red solid lines). The original real- and imaginary-part of the simulated complex-valued signal
(denoted as the blue dashed lines) are also illustrated in Fig. 2(c) and (d). It can be seen the recovered real- and imaginary-part are
well consistent with the original ones (only a little shifting of the phase of the signal caused by the Hilbert operation). In addition, it
can be seen in Fig. 3(b) spectra of the simulated signal are well reconstructed using the proposed CVMD technique. Thus, this
example visually demonstrates the soundness of Eq. (9). Based on the proposed CVMD algorithm, its behavior of complex filter bank
is further studied via numerical experiments in the following subsection...

Fig. 4. CVMD equivalent filter bank with different initialization of center frequencies. In Fig. 3(a), center frequencies of BLIMFs were initialized with uniformly
spaced distribution; center frequencies of BLIMFs were initialized with zero in Fig. 3(b).

79
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

3.2. CVMD equivalent filter bank

It has been demonstrated that IMFs obtained by EMD provide frequency responses similar to those dyadic filter banks [17]. The
behavior of multivariate EMD in the presence of white Gaussian noise was analyzed in [18], where similar filter bank structure was
presented across multiple channels for a multivariate white noise input. In this subsection, the equivalent filtering of the CVMD is
essentially investigated as well. A set of 5000 independent Gaussian complex-valued time series of 1024 samples each were
randomly generated for the numerical experiment. The center frequencies of all BLIMFs were initialized in two ways: uniformly
spaced distribution and zero, while the α mentioned in Eq. (3) was set to 2000. The resulting six spectra were averaged and shown in
Fig. 4, which indicated a fine-to-coarse multi-resolution of BLIMFs. It can be seen CVMD equivalent filter bank structure depends on
the initialization method of the center frequencies. As is shown in Fig. 4(a), when center frequencies are initialized with the
k−1
uniformly spaced distribution ωk0= 2N ,k = 1, …, N , the CVMD equivalent filter banks behave almost constant bandwidths (except the
spectra of the x ± 1 (t ) components), while it exhibits much more adaptive bandwidth when all the center frequencies start from zero
(shown in Fig. 4(b)). Thus, the initialization of the center frequencies to zero will be used for the subsequent practical applications in
Section IV. In order to comprehensively interpret the physical meaning of CVMD technique, CVMD Hilbert spectrum was developed
using the totality of the decomposed components..

3.3. CVMD Hilbert spectrum

As is well known, each time-frequency signal processing method has its own framework for building a time-frequency plot. The
Hilbert-Huang spectrum (HHS) analysis pioneered by Huang et al. [13] applies the Hilbert transform to find the analytical signal
representation of a real-valued signal which can then be used to compute the instantaneous frequency. Moreover, Hilbert transform
is also adopted in the Hilbert spectrum via wavelet projections [19] for the analysis of multi-component non-stationary signals. As
the instantaneous frequencies and amplitudes have already been derived using the local mean decomposition algorithm, the
instantaneous time-frequency spectrum was constructed directly without Hilbert spectrum in [20]. It’s easy to find all these methods
derive the corresponding time-frequency plots based on their respective characteristics. The proposed CVMD algorithm yields the
analytic BLIMFs of the positive and negative frequency components, its Hilbert spectrum can be deduced under Hilbert transform.
More precisely, provided a complex-value signal consists of two analytic sub-signals (the discrete version of derived complex BLIMFs
is given in Eq. (13)), it can be further represented in the polar coordinates as follows
⎧ a (t ) e jϕi (t ), i=1, …, N
zi (t )=⎨ i −jϕ (t )
⎩ ai (t ) e i , i=−N , …, −1 (14)
⎛ y (t ) ⎞
with the instantaneous amplitude ai (t )= xi (t )2 + yi (t )2 and instantaneous phase ϕi (t )=arctan ⎜ xi (t ) ⎟ . Thus yi (t )the instantaneous
⎝ i ⎠
frequency can be defined as
sgn(i ) dϕi (t )
fi (t )= , i=−N , …, −1, 1, …, N
2π dt (15)
The CVMD Hilbert spectrum of the complex-value data x (t ) is expressed in the following form:
N
⎛ ⎞
x (t )= ∑ ai (t )exp ⎜ j
⎝
∫ 2πfi (t ) dt ⎟⎠
i =− N , i ≠0 (16)

Fig. 5. Time sequences of wind speed and direction and the CVMD decomposition. The real- and imaginary-part of the original data are respectively drawn in blue
dashed and red solid lines, and four pairs of BLIMFs corresponding to the positive (blue dashed lines) and negative frequency (red solid lines) components are shown
subsequently.

80
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

In a time-frequency plot, the positive and negative components are formulated on the positive and negative frequency planes,
respectively, which thus comprehensively indicate the underlying time-varying signatures and the correlation between them. Like
the HHS, CVMD Hilbert spectrum is also based on the instantaneous frequencies resulting from the BLIMFs of the signal being
analyzed; thus it is not constrained by the uncertainty limitations with respect to the time and frequency resolutions. Applications of
CVMD and its Hilbert spectrum of real-world complex signals are evaluated in the following section.

4. Applications

4.1. Analysis of wind signal

The real-world wind measurement can be represented by a complex variable using wind speed v (t ) and wind direction
φ (t ):φ (t )=v (t ) e jφ (t )/2π . This complex-valued representation of the wind signal provides a better model than a univariate real-valued
one [21]. If VMD is separately applied to the wind speed and wind direction signal, the relationship between them will be lost.
However, the proposed CVMD algorithm can well preserve this information. Fig. 5 shows the wind speed and direction used in this
example, where the real- and imaginary-part of the derived complex-valued wind representation are also illustrated. BLIMFs with
regard to the positive/negative frequency components using the proposed CVMD are depicted in Fig. 5. This data was also utilized by
CEMD in [5]. The results of this work clearly demonstrate an advantage of CVMD over CEMD is that the number of BLIMFs in the
negative and the positive frequency domain is identical. This nice property is much more applicable and reasonable to analyze
physical data. Fig. 6 indicates the achieved CVMD Hilbert spectrum where the physical meaning and representation of the
instantaneous frequencies and the power of the positive/negative components can be observed...

4.2. Analysis of float drift data

Another typical application of the proposed CVMD algorithm is shown in Fig. 7. The data were the trajectory of a freely drifting
oceanographic float as it followed the ocean currents for hundreds of kilometers in the eastern subtropical Atlantic [22], which were
available online from the World Ocean Circulation Experiment Subsurface Float Data Assembly Center (WFDAC). The looping
trajectories can be clearly observed in Fig. 6 which show the intense swirling currents. As can be seen in Fig. 7, results of CVMD are
different from those achieved using BEMD [1]. We have demonstrated that CVMD behaves an equivalent filter bank structure, as
shown in Fig. 2. The BLIMF components will be physical, for the corresponding scales are physical. Actually, each BLIMF associated
with the real/imaginary-part indicates a different physical pathway in the evolution of a vortex mentioned in [23]. That is the
intention of CVMD Hilbert spectrum developed in this work. Traces of each decomposed mode pair of the oceanographic float signal
are displayed in Fig. 8. The local energy and the instantaneous frequencies derived from the BLIMFs of the positive- and negative-
frequency components can give us a full energy-frequency-time distribution of a complex valued data, which can comprehensively
show the physical meaning of the totality of the decomposed components. It can be seen in Fig. 9 that CVMD Hilbert spectrum shows
a wealth of information regarding to the time-varying variability of the signal, among which is the reflection of instantaneous
frequency variations in an underlying oceanic vortex structure....
The BEMD [1] was proposed for bivariate signal, which is widely used in signal processing applications. As did in BEMD, VMD
can be separately used to analyze the real- and imaginary part of the original signal. The trajectory of a freely drifting oceanographic
float is conducted as a complex-valued series in this work, while it was also analyzed as a bivariate data in [1] and [24]. We here want
to use this example to compare the differences between the proposed CVMD and “bivariate VMD” (BVMD). It should be mentioned
that BVMD, like BEMD, can only handle some special bivariate time series where the two components can be assimilated to the
Cartesian coordinates of a point moving in a two-dimensional space [1]. In addition, if VMD is simply applied to the real and

Fig. 6. The CVMD Hilbert spectrum of the complex-valued wind data.

81
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Fig. 7. The freely drifting oceanographic float signal and its CVMD decomposition. The lines of signals and BLIMFs represent the same objects with those explained
in Fig. 4.

imaginary part of a complex-valued signal, separately, the complex, bivariate quantity with some existing mutual information
between the real and imaginary part is mapped onto two independent real-valued univariate quantities where this mutual
information is lost [5], which can be seen in the Fig. 10. It should be mentioned that bivariate VMD can directly illustrate the trends
of the original signal and the “rotating” component, only in terms of the decomposed some sub-components. Nevertheless, a
limitation of CVMD is that some information (e.g., the low-frequency trend) can not be intuitively found from the individual
decomposed sub-component, compared with the results obtained by BEMD in [1]. The reason is that CVMD is used on the band-
pass filtered data, rather than on the original one. In order to represent the trend signature, CVMD needs to combine the x+ and x− as
well as their Hilbert transform components, as is given in Eqs. (12) and (13)..

4.3. Analysis of mechanical rubbing signal

In recent years, time-frequency analysis technique, such as HHS, has been investigated in feature extraction from nonstationary,
transient vibration signals [25], with machine health monitoring as a target application area. The CVMD is utilized to detect the
rubbing signatures in this subsection. Rubbing experiments were carried out on a SpectraQuest Machinery Fault Simulator (MFK-

Fig. 8. The traces of each decomposed mode pair of the oceanographic float signal.

82
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Fig. 9. The CVMD Hilbert spectrum of the freely drifting oceanographic float data. /label{TIM_7.

Fig. 10. The analysis of real-/imaginary-part of the float data respectively using VMD.

MG) as shown in Fig. 11. The rubbing force was horizontally exerted on the rotor little by little. The original horizontal and vertical
vibration signals were acquired by two orthogonal displacement accelerometers, which are depicted in Fig. 12. Based on these two
signals, synthetic axis orbits are also shown in Fig. 10, where the orbit in blue represents non-rubbing, while orbit in red means the
existence of rubbing. The BLIMFs with regard to the two directional displacement signals using the proposed CVMD are illustrated
in Fig. 12. We can easily find CVMD successfully detect the fundamental rotating components (x ± 1 (t ))and its harmonics
(x ± 2 (t ), x ± 3 (t ), x ± 4 (t )), as well as the impulsive components (x ± 5 (t )) caused by rubbing. Actually, impulsive components clearly
illustrate rubbing begins at about 0.26 s. Moreover, fundamental components can be also found in the traces of each decomposed
mode pair of the rubbing signal, as is illustrated in Fig. 13. Fig. 14 indicates the achieved CVMD Hilbert spectrum where the physical
meaning and representation of the instantaneous frequencies and the power of the horizontal/vertical signals can be observed. For
example, in the negative frequency plane, the fundamental rotating sub-component (horizontal measurement) begins to fluctuate at

Fig. 11. The test rig of rubbing experiment.

83
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

Fig. 12. The orbits of mechanical rubbing signal and its CVMD decomposition. In the left figure, the blue line denotes the non-rubbing state and the red line shows
the rubbing. In the right figures, the vertical and horizontal displacement measurements are shown in blue dashed lines and red solid lines, respectively.

Fig. 13. The traces of decomposed each mode pair of the rubbing signal.

Fig. 14. The CVMD Hilbert spectrum of the mechanical rubbing signal.

84
Y. Wang et al. Mechanical Systems and Signal Processing 86 (2017) 75–85

about 0.26 s. Meanwhile, the impacts can be also found in the region of the normalized ± 0. 3Hz in Fig. 14. Thus, this example
demonstrates that the proposed CVMD can well extract the meaningful signatures buried in the complex-valued signal, while its
Hilbert spectrum can comprehensively indicate this information in the time-frequency domain.....

5. Conclusions

We here have introduced the CVMD method, which represents an extension of the real-valued VMD to the complex domain by
using the relationship between the positive/negative frequency components. Numerical experiments illustrate the behavior of the
CVMD acts as a filter bank structure of white Gaussian series. We also show the effects of the initialization of the center frequencies
of BLIMFs on the alignment of frequency responses. CVMD Hilbert spectrum is developed based on the entire set of BLIMFs, which
can be used to comprehensively reveal time-varying signatures in the time-frequency plot. Three potential applications of CVMD on
real-world complex-valued signals successfully support the analysis.
It should be mentioned that this research is being continued to systematically investigate the suitability and constraints of the
CVMD for complex-valued nonstationary signal analysis applications. We hope this work will encourage more researchers to take full
advantage of the power of complex-valued signal processing.

Acknowledgment

The financial sponsorship from the project of National Natural Science Foundation of China (51475098 and 61463010) and
Guangxi Natural Science Foundation (2016GXNSFFA380008) are gratefully acknowledged. It's also sponsored by Guangxi
Experiment Center of Information Science (20130312) and Guangxi Key Laboratory of Manufacturing System & Advanced
Manufacturing Technology (15-140-30-001Z).

References

[1] G. Rilling, P. Flandrin, P. Goncalves, J.M. Lilly, Bivariate empirical mode decomposition, IEEE Signal Proc. Lett. 14 (2007) 936–939.
[2] W. Yang, R. Court, P.J. Tavner, C.J. Crabtree, Bivariate empirical mode decomposition and its contribution to wind turbine condition monitoring, J. Sound Vib.
330 (2011) 3766–3782.
[3] C. Park, D. Looney, M.M. Van Hulle, D.P. Mandic, The complex local mean decomposition, Neurocomputing 74 (2011) 867–875.
[4] T. Adali, P.J. Schreier, L.L. Scharf, Complex-valued signal processing: the proper way to deal with impropriety, IEEE T Signal Process. 59 (2011) 5101–5125.
[5] T. Tanaka, D.P. Mandic, Complex empirical mode decomposition, IEEE Signal Proc. Lett. 14 (2007) 101–104.
[6] D. Looney, D.P. Mandic, Multiscale image fusion using complex extensions of EMD, IEEE T Signal Process. 57 (2009) 1626–1630.
[7] K. Dragomiretskiy, D. Zosso, Variational mode decomposition, IEEE T Signal Process. 62 (2014) 531–544.
[8] Y. Wang, R. Markert, J. Xiang, W. Zheng, Research on variational mode decomposition and its application in detecting rub-impact fault of the rotor system,
Mech. Syst. Signal Process. 60–61 (2015) 243–251.
[9] Z. Feng, M. Liang, F. Chu, Recent advances in time-frequency analysis methods for machinery fault diagnosis: a review with application examples, Mech. Syst.
Signal Pr. 38 (2013) 165–205.
[10] N. Rehman, D.P. Mandic, Multivariate empirical mode decomposition, Proc. R. Soc. A-Math. Phys. Eng. Sci. 466 (2010) 1291–1302.
[11] J. Fleureau, A. Kachenoura, L. Albera, J.-C. Nunes, L. Senhadji, Multivariate empirical mode decomposition and application to multichannel filtering, Signal
Process. 91 (2011) 2783–2792.
[12] A. Ahrabian, D. Looney, L. Stankovic, D.P. Mandic, Synchrosqueezing-based time-frequency analysis of multivariate data, Signal Process. 106 (2015) 331–341.
[13] N.E. Huang, Z. Shen, S.R. Long, A new view of nonlinear water waves: the Hilbert spectrum, Annu Rev. Fluid Mech. 31 (1999) 417–457.
[14] M.R. Hestenes, Multiplier and gradient methods, J. Optim. Theory Appl. 4 (1969) 303–320.
[15] Y. Wang, R. Markert, Filter bank property of variational mode decomposition and its applications, Signal Process. 120 (2016) 509–521.
[16] Y.C. Lim, Frequency-response masking apprach for the synthesis of sharp linear-phase digital filters, IEEE Trans. Circuits Syst. 33 (1986) 357–364.
[17] P. Flandrin, G. Rilling, P. Goncalves, Empirical mode decomposition as a filter bank, IEEE Signal Proc. Lett. 11 (2004) 112–114.
[18] N.U. Rehman, D.P. Mandic, Filter bank property of multivariate empirical mode decomposition, IEEE T Signal Process. 59 (2011) 2421–2426.
[19] S. Olhede, A.T. Walden, The Hilbert spectrum via wavelet projections, Proc. R. Soc. Lond. Ser. a-Math. Phys. Eng. Sci. 460 (2004) 955–975.
[20] Y. Wang, Z. He, Y. Zi, A demodulation method based on local mean decomposition and its application in rub-impact fault diagnosis, Meas. Sci. Technol. (20)
(2009) [10 pages].
[21] S.L. Goh, M. Chen, D.H. Popovic, K. Aihara, D. Obradovic, D.P. Mandic, Complex-valued forecasting of wind profile, Renew. Energy 31 (2006) 1733–1750.
[22] P.L. Richardson, D. Walsh, L. Armi, M. Schröder, J.F. Price, Tracking three meddies with SOFAR floats, J. Phys. Oceanogr. 19 (1989) 371–383.
[23] J.C. McWilliams, The vortices of geostrophic turbulence, J. Fluid Mech. 219 (1990) 387–404.
[24] J.M. Lilly, S.C. Olhede, Bivariate instantaneous frequency and bandwidth, IEEE T Signal Process. 58 (2010) 591–603.
[25] R. Yan, R.X. Gao, Hilbert-Huang transform-based vibration signal analysis for machine health monitoring, IEEE T Instrum. Meas. 55 (2006) 2320–2329.

85