See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328434279

Fast Multivariate Empirical Mode Decomposition

Article  in  IEEE Access · October 2018

DOI: 10.1109/ACCESS.2018.2877150

CITATIONS READS
5 417

7 authors, including:

Lang Xun Alexander Horch

Zhejiang University ETH Zurich
11 PUBLICATIONS   19 CITATIONS    81 PUBLICATIONS   1,377 CITATIONS   

SEE PROFILE SEE PROFILE

Hongye Su
Zhejiang University
549 PUBLICATIONS   9,336 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

quantized and event-triggered control View project

Oscillation detection and diagnosis View project

All content following this page was uploaded by Lang Xun on 30 October 2018.

The user has requested enhancement of the downloaded file.

IEEE ACCESS
Fast Multivariate Empirical Mode Decomposition
Xun Lang, Qian Zheng, Zhiming Zhang, Shan Lu, Lei Xie, Alexander Horch,
and Hongye Su, Senior Member, IEEE
Abstract—The multivariate empirical mode decomposition
(MEMD) has been pioneered recently for adaptively process-
ing of multichannel data. Despite its high efficiency on time-
frequency analysis of nonlinear and nonstationary signals, high
computational load and over-decomposition have restricted wider
applications of MEMD. To address these challenges, a fast
MEMD (FMEMD) algorithm is proposed and featured by the
following contributions: (i) A novel concept, pseudo direction-
independent Multivariate Intrinsic Mode Function (IMIMF)
which allows the interchange of sifting and projection operations,
is defined for the purpose of developing FMEMD, (ii) FMEMD
is computationally efficient. Compared with MEMD, the number
of time-consuming sifting operations reduces from K · p to K for
each iteration, where K and p denote the number of projection
directions and signal dimension, respectively, (iii) FMEMD is
consistent with EMD in terms of the dyadic filter bank property
and (iv) FMEMD is more effective in working at low sampling
rate. Validity of the raised approach is demonstrated on a wide
variety of real world applications.
Index Terms—Multivariate empirical mode decomposition,
multivariate intrinsic mode function, fast MEMD, computational
load, filter bank.
I. I NTRODUCTION
E MPIRICAL mode decomposition is one of the most
powerful time-frequency (T-F) tools and has been ex-
tensively studied and widely applied in numerous engineering
applications [1]. Unlike traditional techniques such as Fourier
and Wavelet transform that project signal onto predefined
basis functions, the bases of EMD are derived from the
data and thus can be nonlinear and nonstationary [2]. By
processing on the local characteristic time scales, the input
data is decomposed into a complete and finite set of localized
amplitude/frequency modulated (AM/FM) components called
intrinsic mode functions (IMFs) [3]. The IMFs indicate the
natural oscillatory modes embedded in the signal and serve as
the basis functions, which are determined by the signal itself
rather than preset kernels [4]. Due to the ability of EMD to
analyze nonlinear and nonstationary processes, the number of
This work was supported in part by Natural Science Foundation of Zhejiang,
China under Grant LR17F030002, Science Fund for Creative Research Groups
of the National Natural Science Foundation of China under Grant 61621002,
and National Natural Science Foundation of P.R. China under Grant 61673358.
Xun. Lang, Qian. Zheng, Zhiming. Zhang, Lei. Xie and Hongye, Su are
with the State Key Lab of Industrial Control Technology, Institute of Advanced
Process Control, Zhejiang University, Hangzhou 310027, China (e-mail:
langxun@zju.edu.cn; 21732059@zju.edu.cn; zhangzhimingzju@gmail.com;
leix@iipc.zju.edu.cn; hysu@iipc.zju.edu.cn).
Shan Lu is with the Shenzhen Polytechnic, 518055 Shenzhen, China (e-
mail: lushan@szpt.edu.cn).
Alexander Horch is with the HIMA Paul Hildebrandt GmbH, Albert-
Bassermann-Str. 28, 68782 Bruehl bei Mannheim, Germany (e-mail:
a.horch@hima.com).
Corresponding authors: Shan Lu (lushan@szpt.edu.cn) and Lei Xie
(leix@iipc.zju.edu.cn).
1
publications on EMD has been increasing steadily over the
past decade [5].
In recent years, advances in signal acquisition tools have
highlighted the requirement for synchronous processing of
multichannel data [6]. As a result, the standard EMD has
been extended to various versions including those suitable for
the bivariate [7], trivariate [8] and multivariate signals [9].
Bivariate EMD (BEMD) [7] estimates the bivariate local mean
by mapping the input to a number of real-valued projections
and then averaging the respective local mean obtained from
these projected signals. Following the same idea, ur Rehman
et al. [8] extended the EMD algorithm to Trivariate EMD
(TEMD) by taking the three-dimensional signal as a pure
quaternion. Similarly, the multivariate EMD (MEMD) [9]
further generalizes this concept for multivariate signal and
efficiently solve the uncertainty of IMFs arrangement [10].
The MEMD performs well in many applications from
various fields such as process control [11], mechanical health
diagnosis [10], biomedicine [12], data denoising [13], etc.
However, the major drawback of MEMD is that it requires a
long computation time especially for high-dimensional signals.
More precisely, a total number of K · p sifting operations are
required for each MEMD-iteration on a p–variate signal with
K projection directions [14]. Note that sifting operations is
the most time-consuming step in EMD since it involves the
extrema-based cube spline interpolation to obtain the signal
envelope. Recently, several dynamical sampling schemes [15],
[16], [17] have been provided to improve the computational
performance of BEMD/MEMD. Notice that these improve-
ments are continuous and incremental, without avoiding the
envelope interpolation on multidimensional space.
Another drawback of MEMD relates to the realization of
the sifting process. According to [9], the sifting process for
a multivariate signal does not stop until all projected signals
fulfill the stoppage criterion of standard EMD. This is, how-
ever, unnecessary since different dimension of the investigated
signal may possess distinct T-F features. Enforcing the same
number of shifting operations for all signal channels will
inevitably induce over-decomposition effect and corrupt the
T-F information encapsulated in multivariate IMF (MIMF).
The contributions of this paper are fourfold, (i) The concept
of MIMF defined in [7], [8], [9] is extended to direction-
independent MIMF (IMIMF) and its pseudo version which fa-
cilitate the development of FMEMD. Since the pseudo IMIMF
allows the interchange of projection and sifting operations,
FMEMD firstly extracts a set of IMFs from the projected
univariate signals, then the desired pseudo IMIMF is obtained
by solving an overdetermined system of linear equations. (ii)
FMEMD is computationally much more efficient than MEMD,
i.e. the number of sifting operation for each iteration reduces
IEEE ACCESS
from K · p to K. (iii) The filter bank property of FMEMD is
analyzed with the study on white Gaussian noise (WGN). It
reveals that FMEMD is consistent with EMD in the sense that
it also acts as a dyadic filter bank. (iv) The proposed FMEMD
avoids the over-decomposition problem and is more effective
in working at low sampling rate, since the IMFs for different
projections are obtained independently.
The rest of this paper is organized as follows. Traditional
EMD and MEMD are first reviewed in Section II. Section
III details the proposed FMEMD approach. In Section IV,
the filter bank properties of FMEMD are analyzed by de-
composing multichannel white Gaussian noise. A comparative
study between FMEMD and MEMD is presented in Section V.
Section VI studies the FMEMD performance with three real
world applications. It is followed by conclusions in Section
VII.
II. P RELIMINARIES
A. Empirical Mode Decomposition
The standard EMD aims to adaptively decompose a signal
into a finite set of oscillatory components known as intrinsic
mode functions (IMFs) [3]. The decomposition is achieved
by sifting out rapidly oscillating components from the data,
through iteratively subtracting the slower oscillatory modes.
More specifically, for a real valued p–dimensional signal x (t),
the application of EMD yields M sets of IMFs, denoted as
M
{di (t)}i=1 , and a monotonic residue r (t), so that
M m (t) =
X K
x (t) = di (t) + r (t) , (1)
i=1
where the residual signal r (t) represents the trend within the
signal. The procedures used for extraction of IMFs from a
real-valued signal x (t) are summarized in Algo. 1.
Algorithm 1 Empirical mode decomposition.
Input: x1 (t) = x2 (t) = x (t), i = 1
1: Find the locations of all the extrema of x1 (t);
2: Interpolate (typically using cubic spline interpolation) all
the maxima (minima) to obtain the upper (lower) enve-
lope, emax (t) (emin (t));
3: Find the local mean, m (t) = [emin (t) + emax (t)]/2;
4: Subtract the mean from the signal to obtain the oscillatory
mode, s (t) = x1 (t) − m (t);
5: If s (t) obeys the stoppage criterion, di (t) = s (t) be-
comes an IMF, go to step 6. Otherwise set x1 (t) = s (t)
and repeat the procedures from step 1;
6: Subtract the derived IMF from x2 (t), so that x2 (t) =
x2 (t) − di (t). If x2 (t) becomes a monotonic function, or
does not contain enough extrema to form a meaningful
envelope, stop the sifting process with r (t) = x2 (t).
Otherwise, x1 (t) = x2 (t), i = i + 1 and go to step 1;
M P
7: return {di (t)}i=1 and r (t);
2
B. Multivariate Empirical Mode Decomposition
The standard EMD is only applicable for univariate signal.
For multivariate signals, however, the local maxima and min-
ima cannot be defined directly, and the notion of oscillatory
mode defining a multivariate IMF is not straightforward either
[7]. To tackle this problem, multiple p-dimensional envelopes
are generated by taking signal projections along different
directions in p-dimensional space, and subsequently interpo-
lating their extrema [9]. These envelopes are then averaged to
obtain the local multivariate mean.
Let x (t) = [x1 (t) , x2 (t) , . . . , xp (t)] be a p-variate
   K
signal, and vθk = v1k , v2k , . . . , vpk k=1 denotes the kth
projection vector alongio the direction given by angle
n h K
θk = θk1 , θk2 , . . . , θkp−1 on an (p − 1) sphere1 , the
k=1
MEMD algorithm is summarized in Algo. 2.
Algorithm 2 Multivariate empirical mode decomposition.
Input: x1 (t) = x2 (t) = x (t), i = 1;
1: Generate a K uniformly distributed θk .
2: Calculate the kth projection q θk (t) of the input x1 (t) as
q θk (t) = x1 (t) · (vnθk )To, for all k (k = 1, 2, . . . , K);
θ
3: Find time instants ti k , which correspond to maxima of
θk
the projected h signals q i (t) for all k;
θk θk
4: Interpolate ti , x1 ti to obtain the multivariate en-
θk
velope e (t);
5: For a set of K direction vectors, the mean of envelope
curves, m (t), is given by
K
1 X θk
e (t). (2)
k=1
6: Extract the detail signal s (t) using s (t) = x1 (t) − m (t).
If the detail signal satisfies the stopping criterion of MIMF,
di (t) = s (t), go to step 7. Otherwise x1 (t) = s (t), go
to step 2;
7: Update x2 (t) = x2 (t) − di (t). If x2 (t) becomes mono-
tonic, or does not contain enough extrema to form a
meaningful multivariate envelope, stop the above sifting
process, and obtain the trend r (t) = x2 (t). Otherwise,
let i = i + 1 and x1 (t) = x2 (t), go to step 2.
M
8: return {di (t)}i=1 and r (t).
As a result, MEMD uses a vector-valued form of standard
EMD to decompose a p-variate signal x (t) as
M
X
x (t) = di (t) + r (t) , (3)
i=1
M
where the p-variate MIMFs, {di (t)}i=1 , contain scale-aligned
intrinsic joint rotational modes [2].
1 A (p − 1) sphere (hypersphere) is an extension of the ordinary sphere
to an arbitrary dimension and is represented mathematically in R =
p
(xi − Ci )2 , where R denotes the radius and Ci is the ith coordinate
i=1
value of center point C. This work adopts the terminology that a (p − 1)
sphere resides in a p-dimensional Euclidean coordinate system.
IEEE ACCESS
III. FAST M ULTIVARIATE E MPIRICAL M ODE
D ECOMPOSITION
This section introduces the proposed multivariate extension
of EMD algorithm. The main content of this section includes:
(i) definitions of the direction-dependent MIMF (DMIMF) and
the direction-independent MIMF (IMIMF), (ii) methods for
computing the DMIMF, (iii) pseudo-IMIMF and its calcu-
lation, (iv) the specific algorithm of FMEMD and (v) some
supplementary statements.
A. Definitions around MIMF
The standard EMD is implemented based on the intuitive
definition of IMF that naturally relates to local extrema.
According to [3], an IMF is a function that satisfies two
conditions: (1) in the whole data set, the number of extrema
and the number of zero crossings must either equal or differ
at most by one; and (2) at any point, the mean value of the
envelope defined by the local maxima and the envelope defined
by the local minima is zero. In the multivariate case, BEMD
[7], TEMD [8] and MEMD [9] compute the undefined MIMFs
through iteratively subtracting multivariate local means from
the original data, where the multi-dimensional local mean
is defined as an average of multiple envelopes obtained by
interpolating the extrema that are extracted from multiple
projections. However, an accurate definition of MIMF is still
missing which renders the analysis of above methods difficult.
To resolve the problem, this section proposes to define
the DMIMF and IMIMF based on notions of standard IMF
and multivariate extrema. Although an intuitive definition
of multivariate extrema is not straightforward, ref [9] has
illustrated that they can be extracted corresponding to specified
projection directions. For instance, a bivariate signal is adopted
for describing the geometric characteristic of bivariate (multi-
variate) extrema as shown in Fig. 1. If the projection is selected
as the vertical direction, point A will be considered as a local
maximum of the given signal according to the definition2 in
[18]. Likewise, a local minimum is also collected, which is
given by point B. Accordingly, [7], [8], [9] utilized the obser-
vation that the multivariate extrema on a specified direction
are in one-to-one correspondence to the univariate extrema,
which are identified from the projected signal. Therefore, in
this work, a general definition of multivariate extrema is given
as follows,
Definition 1. Given a specified direction vector, the local ex-
tremum of a multivariate signal is a point whose projection on
that direction is a univariate extremum that can be extracted
from the same projected signal.
The above definition constructs a relationship between
the multivariate extrema and the univariate extrema that are
extracted from the corresponding projections. According to
this definition, it can be easily proved that the multivariate
2A point c in the interior of the domain of a function f (x) is a point of
local maximum if there exists a value δ > 0 such that f (x) ≤ f (c) for all
x satisfying the condition that the distance between x and c is less than δ. In
this work, the function f (x) is set as the projected distance (projected on the
projection direction) between the observer and the point on the given signal.
3
A
Y
projection
direction
signal
direction
B tangent
direction
X
Fig. 1. Illustration of bivariate local maximum and minimum.
extrema are in one-to-one correspondence to the univariate
extrema. More precisely, given a specified direction vector,
the multivariate signal reaches a local maximum (minimum)
if and only if its projected signal reaches a local maximum
(minimum). Similarly, the univariate IMF can also be extended
to multivariate case in combination with a specified direction,
which yields the definition as,
Definition 2. Given a specified direction vector, the direction-
dependent multivariate intrinsic mode function (DMIMF) is
a function whose projection on that direction satisfies two
conditions:
(1) symmetric upper and lower envelopes (zero mean);
(2) the number of zero-crossings and extrema must either
equal or differ at most by one.
This definition is naturally extended on the basis of the
local multivariate extrema that are defined with specified
directions. Moreover, conditions (1) and (2) are borrowed
from the conditions of IMF for standard EMD in order
to guarantee a consistent development of the multivariate
EMD. As a result, such definition is able to simplified to
standard IMF in univariate case, and their consistency will
be demonstrated further by the filter bank property analysis
in Section IV. Note that a multivariate signal may produce
an infinite number of DMIMFs if considering all directions
defined on the hyperspace.
For the purpose of FMEMD application, an unique syn-
thesized MIMF should be developed, which means a repre-
sentative MIMF that is independent of the projection direc-
tion is preferentially required. More precisely, a direction-
independent MIMF (IMIMF) should satisfy the characteristics
of DMIMFs from all directions.
Definition 3. The direction-independent multivariate intrinsic
mode function (IMIMF) of a multivariate signal is a function
whose projection on each of the directions satisfies two con-
ditions:
(1) symmetric upper and lower envelopes (zero mean);
IEEE ACCESS 4

(2) the number of zero-crossings and extrema must either fulfilled after l times sifting, a well-developed DMIMF can be
equal or differ at most by one. computed as
dθk (t) = D| ·{z
· · D} x (t) . (4)
Remarks:
l
1) The defined IMIMF is an intersection of all sets of
DMIMFs that satisfying conditions from definition 2. It It can be proved that if the standard EMD algorithm is
requires that the projection of IMIMF on each of the meaningful on the projected signal, there exists a positive
directions is a well-defined IMF. As a result, the physical integer l so that Eq. (4) meets the requirements for defining a
meaning of the univariate IMF has been extended to DMIMF.
multivariate case by IMIMF in a strict sense. The IMIMF
Lemma 1. The projection of the cubic spline interpolation
is a symmetrical multivariate-oscillation with local zero
curve of a multi-dimensional point set is equivalent to the
mean, for which the instantaneous frequency can be
interpolation that is calculated from the projected point set.
defined everywhere. Similarly, although [7], [9] have not
provided a clear definition of the rotation (MIMF), it can A mathematical proof of the proposed lemma is presented
be inferred that the projection of rotation for any given in Appendix. A.
direction is also a zero-mean oscillation. Therefore, the
defined IMIMF and the rotation defined by MEMD are Theorem 1. Given a specified direction vector, there exists
essentially the same notion. Both of them are naturally a positive integer l that makes D
| ·{z
· · D} x (t) satisfying the
extended from the definition of univariate IMF (oscilla- l
tion). On the other hand, they realized the computation conditions of DMIMF if and only if EMD is convergent on
of MIMF/IMIMF in different ways, and the next few the projected signal.
sections will show that the proposed IMIMF is a more Proof. Let x (t) denotes the investigated multivariate signal,
effective extension. xθk (t) represents the projection of x (t) along vθk , the rela-
2) Extending the sifting process for computing a standard tionship between x (t) and xθk (t) is given as
IMF to IMIMF calculation is unrealizable in practice
due to following difficulties: (i) Finding all multivariate x (t) · vθk = xθk (t) . (5)
extrema requires projecting the original signal onto infi-
nite directions; (ii) Determining whether a given signal is According to definition 1, the extracted multivariate extrema
IMIMF also needs to check its projections onto infinite of x (t) and the univariate extrema of xθk (t) fulfill the
directions. following equation set
3) A pseudo-IMIMF that defined only on K directions will
 max θ
xm · v k = xmax m , m = 1, 2, . . . , M
be given in Section III-C as an alternative, where K > p . (6)
xmin
n · v θk
= x min
n , n = 1, 2, . . . , N
is a finite positive integer.  min N
M
where {xmaxm }m=1 and xn n=1 denote the multivariate
max M
 N
B. Interchangeability for Computing DMIMF extrema, {xm }m=1 and xmin n n=1
are univariate extrema
of xθk (t) respectively. Lemma 1 shows that projection of
Before addressing the problem of calculating (pseudo) IM-
the cubic spline interpolation curve of a multi-dimensional
IMF for the purpose of FMEMD development, the direction-
point set is equivalent to the interpolation calculated from the
dependent MIMF (DMIMF) is first computed in combination
projected point set. Therefore it can be concluded that
with a theoretical demonstration of the interchangeability

S (xmax ) · vθk = S xmax θk max

between (multivariate) sifting process and direction-base pro- m
m · v
=S (xm
)
min θk min θk min , (7)
jection. Given a specified direction vector vθk , the sifting S xn · v = S xn · v =S xn
process of standard EMD (step 1 to step 4 in Algo. 1) can
where S(·) denotes the spline interpolation operator, which
be extended multivariable for decomposing DMIMF from a
yields
p-variate signal x (t), as given by: 
eu (t) · vθk = eθuk (t)
1) Find the locations of all maxima {xmax
M , (8)
 min N m }m=1 and minima el (t) · vθk = eθl k (t)
xn n=1 on direction vθk .
2) Interpolate (using cubic spline interpolation channel-wise in which eu (t) (eθuk (t)) and el (t) (eθl k (t)) represent the upper
[9]) all the maxima (minima) to obtain the upper (lower) and lower envelopes of x (t) (xθk (t)), respectively. On the
envelope, eu (t) (el (t)). basis of Eq. (8), the relationship between the local mean
3) Compute the local mean, m (t) = [eu (t) + el (t)]/2. (m (t)) of x (t) and the local mean (mθk (t)) of xθk (t) is
4) Subtract the mean from the signal to obtain the oscillatory given by
modes, o (t) = x (t) − m (t). [eu (t) + el (t)] θk
5) If o (t) obeys the two conditions in definition 2, a m (t) · vθk = ·v
2
DMIMF is obtained. Otherwise, repeat step 1 to 4 on
h i . (9)
eθuk (t) + eθl k (t)
o (t). = θk
= m (t)
Let D denotes the operation from step 1 to 4, then Dx (t) = 2
x (t) −m (t). By supposing that conditions for definition 2 are As a result, the so obtained oscillatory modes Dx (t) and
IEEE ACCESS
Dxθk (t) show the following relationship
θk θk
Dx (t) · v = [x (t) −m (t)] · v
. (10)
= xθk (t) −mθk (t) = Dxθk (t)
Eq. (10) proves that operations ’multivariate sifting +
projection’ and ’projection + univariate sifting’ satisfy in-
terchangeability. This useful property can be exploited for
FMEMD development, which will be further discussed in
Section III-C.
Since the operator D does not change the projection rela-
tionship, a new equation is obtained further
 
D · · · D x (t) · vθk = D · · · D xθk (t) .
| {z } | {z }
l l
Part I: If the EMD is not convergent on the projected
signal, none of the IMFs can be extracted from xθk (t).
Therefore, it is impossible to find such a positive integer l
that causes D · · D} xθk (t) to be a proper IMF. In accordance
| ·{z
l K
with definition 2 and Eq. (4), dθk (t) does not exist due to the
infeasibility of l.
Part II: If the EMD algorithm is meaningful on the projected
signal, the projected signal must yield an univariate IMF after
integer times sifting. By supposing it costs l0 times sifting for
decomposing the current IMF (dθk (t)) from xθk (t), we have
dθk (t) = D · · D} xθk (t) .
| ·{z
l0
Combine Eq. (11), Eq. (12) with definition 2, the DMIMF
dθk (t) that defined on vector vθk can be properly addressed
after l0 times multivariate sifting, which satisfies the following
relationship
 
dθk (t) · vθk = D · · D} x (t) · vθk
| ·{z
l0 . (13)
=D · · D} xθk (t) =dθk (t)
| ·{z
l0
This equality proposes a complete scheme for computing
the DMIMF on such specified direction. The projective rela-
tionship between the calculated DMIMF and the univariate
IMF extracted from the projected signal has been proved
theoretically.
The presented theorem 1 also shows the irrationality of the
original MEMD method [9]. For each specified direction vec-
tor, the T-F feature of the current projection is different from
other projected signals, therefore, distinct numbers of itera-
tions are required for extracting different direction-dependent
MIMFs. As a result, running of the original MEMD, which
enforces the same number of sifting for all data channels, may
induce the effect of over-decomposition.
C. Raise of Pseudo-IMIMF
Inspired from theorem 1, it is observed that processes
’DMIMF extraction + projection’ and ’projection + IMF
5
extraction’ satisfy interchangeability. Therefore, the ideal
direction-independent MIMF (IMIMF) as an intersection of
all meaningful DMIMFs should also meet the same property.
Mathematically, it requires the theoretical IMIMF id (t) of the
analyzed signal x (t) satisfies the following equation
K
id (t) · vθk =dθk (t) · vθk = dθk (t) k=1 , K → ∞, (14)

K

where dθk (t) k=1 denote all sets of DMIMFs calculated
 K  K
upon direction vectors vθk k=1 , and dθk (t) k=1 are cor-
responding IMFs extracted from the projected signals. Since
K is an infinite positive integer, the above system of overde-
termined linear equations is unsolvable in engineering. To
(11) address the problem, a pseudo definition of IMIMF with
relaxed constraint is presented by limiting K into a finite
positive integer, which is,
 K
Definition 4. For any given set of direction vectors vθk k=1 ,
where K is a limited and positive integer. The pseudo-IMIMF
id (t) of a multivariate signal satisfies the following equation:
id (t) · vθk =dθk (t) k=1 ,

(15)
 θ K
where d k (t) k=1 are corresponding IMFs extracted from
the projected signals..
If a p-variate signal x (t) is projected into K directions
to yield K projections (K ≥ p, and it is generally set as
K > p for reducing noise), a set of overdetermined linear
(12)
equations can be obtained after the corresponding univariate
IMFs being extracted. Accordingly, the first pseudo-IMIMF
can be computed by solving Eq. (15), and the first slower
mean, s (t), is therefore subtracted into
s (t) =x (t) − id (t) . (16)
The procedures can be re-applied by setting the mean signal
s (t) as a new input, then the iteration continues until a
monotonic trend r (t) is obtained. In consequence, an indi-
rect multivariate extension of the EMD algorithm (namely,
FMEMD) is realized. For convenience, the rest of this paper
uses IMIMF instead of the defined pseudo-IMIMF.
D. Algorithm of FMEMD
Definition 4 enables FMEMD to be generalized indirectly
by making full use of the decomposed results from the
univariate EMD. Based on solving an overdetermined system
of linear equations (Eq. (15)), the p-variate IMIMF of the input
data can be simply computed, which yields the slower mean
as: mean = signal - IMIMF (Eq. (16)).
Consider a sequence of p-dimensional signal x (t) =
[x1 (t) , x2 (t) , . . . , xp (t)] which represents a multivariate sig-
   K
nal with p components, and vθk = v1k , v2k , . . . , vpk k=1
denote ha set of vectors alongio the directions given by angles
n K
θk = θk1 , θk2 , . . . , θkp−1 on a (p − 1) sphere, the pro-
k=1
posed FMEMD decomposes signal x (t) into
M
X
x (t) = idi (t) + r (t) , (17)
i=1
IEEE ACCESS 6

M
where the p-variate IMIMFs, {idi (t)}i=1 , contain scale- The Halton and Hammersley sequence, however, does not
aligned intrinsic joint components. The algorithm of FMEMD usually provide suitable samplings. Consider that a direction
T
is summarized in Algo. 3. vector [1, 0] is selected to process a bivariate signal, x (t) =
[sin (3πt) , cos (6πt)], then the univariate IMF extracted from
Algorithm 3 Algorithm of FMEMD. the projected function would be sin (3πt). More specifically,
Input: x1 (t) = x (t), i = 1 the direction vector with a zero ingredient obscures one
1: Generate a suitable point set for uniform projection; component of the bivariate input, which causes the higher
2: Calculate the kth projection q θk (t) of the input signal frequency component x2 (t) = cos (6πt) to be projected into
x1 (t) along direction vector vθk , for all k (i.e. k = zero. Therefore, a vector whose direction is very close to an
1, 2, . . . , K, where K gives the whole set of vectors); unit coordinate vector will be harmful to the estimation of
3: Extract the first univariate IMF dθk (t) of the projected IMIMF. To resolve this issue, cosine similarity is introduced
function q θk (t) for all k using the standard EMD algo- for further filtering of the invalid vectors. Given two vectors
rithm; of attributes, A and B, the cosine similarity, cos(θ), is
 K
4: Combine all IMFs dθk (t) k=1 with their corresponding represented using a dot product and magnitude as
 θ K
direction vectors v k k=1 , the p-variate IMIMF idi (t) n
P
ai bi
can be obtained by solving the following overdetermined A·B
equations cos (θ) = = s i=1 s , (19)
kAk2 kBk2 n n
a2i b2i
P P
 θ K  K
d k (t) k=1 = idi (t) · vθk k=1 . (18) i=1 i=1

where ai and bi are coefficients in vector A and B respec-

5: The slower mean is calculated by subtracting the yielded tively. In practice, two vectors with the same orientation have
IMIMF from the current input signal s (t) =x1 (t) − a cosine similarity of 1, two vectors at 90◦ have a similarity
idi (t); of 0, and −1 denotes opposed vectors. According to our
6: If s (t) becomes monotonic, or does not contain enough experience, the absolute similarity of a direction vector to the
extrema to form meaningful multivariate envelopes, stop unit coordinate vector should be confined less than 0.8.
the iterative process and obtain the trend, r (t) = s (t). h iT
Otherwise, update the current input as x1 (t) = s (t) and Once a specified set of angles θk = θk1 , θk2 , . . . , θkp−1
i = i + 1, then go to step 2; is calculated based on Hammersley T sequences, the projection
M

7: return {idi (t)}i=1 and r (t); vector vθk = v1k , v2k , . . . , vpk can be simply given in
 k
 v1 = cos(θk1 )
Remarks:  v k = sin(θ1 ) cos(θ2 )

 2 k k


1) The stopping criterion used for univariate IMF extraction ..
. . (20)
in step 3 is borrowed from [19], where the sifting is  k 1 p−2 p−1
v = sin(θ ) · · · sin(θ ) cos(θ )


stopped when the number of zero crossings and extrema  p−1 k k k


is the same number for S successive sifting steps. Typ- vpk = sin(θk1 ) · · · sin(θkp−2 ) sin(θkp−1 )
ically, a value of S = 5 has proved successful as the
default stopping criterion. 2) Solution to overdetermined equations: To enhance the
2) The standard EMD is a special form of the proposed reconstruction of the proposed IMIMFs, while attenuate some
FMEMD algorithm. FMEMD will be simplified to EMD unexpected noise, the number of projections is chosen greater
T
when the direction vector is set to [1] . than the dimension of input (K ≥ 6p is adopted in this
3) The decomposition of a multivariate signal via FMEMD work). As a result, Eq. (18) is usually designed into an
exhibits mode alignment property, whereby common fre- overdetermined system of linear equations. For convenience,
quency modes from distinct channels are aligned in a the above system is rewritten into the form of y = Sx
single IMIMF. This property will be further demonstrated for each time instance t, where matrix S ∈ RK×p is the
in Section IV and VI. selected direction set and vector y ∈ RK×1 denotes all K
IMFs, while the vector x ∈ Rp×1 is unknown. In many
real world applications, the overdetermined system is usually
E. Supplementary Statements inconsistent, thus it is desired to find a best approximation. The
1) Direction vectors selection: IMIMF can be considered well known approach ([21]) is to find a selection of vector xo
as an approximation of all solutions satisfying Eq. (18) along so that the `2 norm (sum of squared errors criterion) of the
multiple projections, therefore, the accuracy of this approxima- residual error vector
tion for real-world data depends on the validity of the direction
e = y − Sxo (21)
vectors, especially for a limited member of directions. Ac-
cording to [9], Halton and Hammersley sequence [20] can be is minimized. One of the main benefits accrued in employing
conveniently utilized to generate a set of uniformly distributed a minimum `2 norm criterion is the existence of a closed form
points on a (p − 1) sphere, and hence can be equivalently solution to this approximation problem. According to [21], if
represented as the corresponding direction vectors. matrix S has full column rank, there exists a unique solution
IEEE ACCESS
to the normal equations as given by
−1
xo = ST S ST y,

(22)
where xo is the unique optimal solution to x. To guarantee the
existence of ST S, a mechanism to determine whether ST S is
full rank is preferentially required. If such condition cannot
be satisfied, the starting point of Halton and Hammersley
sequence should be changed by a new random integer to obtain
a new set of direction vectors. Actually, the constructed matrix
ST S is usually full rank due to the fact that K ≥ 6p and vθk
is generated in a uniform way.
IV. F ILTER BANK P ROPERTY A NALYSIS
In order to verify the consistency between the proposed
FMEMD and standard EMD algorithm, the decomposition
performance of FMEMD, which is called as the filter bank
property [22], is evaluated in comparison with EMD. More
precisely, the characterization of FMEMD is carried out in
frequency domain from a stochastic perspective to understand
how a full spectrum process is split into its IMIMFs3 .
It has been shown in [22] and [23] that IMFs obtained
from the standard EMD present frequency response similar
to that of a dyadic filter bank, and the Fourier spectra of
the IMF components are identical in shape, which cover the
same area on a semi-logarithmic period scale. Analogously,
this section investigates whether FMEMD presents the similar
filter bank structure. More exactly, it needs to check whether
the decomposed sub-noise satisfy the following properties:
1) Mode alignment. This property indicates that FMEMD
should ensure the overlapping of the frequency response
of corresponding IMFs from different channels which are
associated with the same index.
2) Dyadic filter. A dyadic filter is capable of separating
the white noise into IMIMF components having mean
frequency exactly half of the previous one.
3) Unbiasedness. Since the contribution to the total spectrum
energy comes from each Fourier component of a white
Gaussian noise is uniform and equivalent, the spectral
shapes associated with the corresponding IMIMFs from
multiple indexes should exhibit some self-similarity.
In consequence, this section sets out to investigate the filter
bank structure of FMEMD from three aspects: (i) the averaged
spectra of WGN realizations, (ii) regression on number of the
zero crossings in each IMIMF, and (iii) self-similarity analysis.
The frequency response of FMEMD is first illustrated by
applying the proposed algorithm on totally 500 different four-
channel WGN realizations with each length N = 1, 000. The
number of the direction vectors is chosen as K = 64. The
power spectra of the first eight IMIMFs are then computed
and ensemble averaged to yield an averaged power spectra as
3 Itis mainly due to the fact that both EMD and FMEMD are presently
defined only as the output of iterative algorithms, with no analytical definition
that could be used for performance evaluation. In addition, their high- versus
low-frequency discrimination applies only locally and corresponds by no way
to a predefined subband filtering. Therefore, numerical experiments based
on white Gaussian noise is applied on FMEMD as an alternative, with the
objective of obtaining a empirical and statistical knowledge of its filtering
behavior in signal processing.
7
Averaged spectra of white noise realizations from FMEMD
IMIMF 8 7 6 5 4 3 2 1
Spectrum
0 1 2
10 10 10
Frequency(Log)
Zero crossings(log 2)
10
FMEMD approximate slope = −1.0104
8
6
4
2
1 2 3 4 5 6 7 8
IMIMF index
Fig. 2. Top: averaged spectra of IMIMFs (IMIMF1-IMIMF8 from right to
left) obtained from 500 realizations of four-channel white Gaussian noise by
FMEMD. Bottom: averaged number of base-2 logarithm of zero crossings
plotted versus the IMIMF index.
shown in the top of Fig. 2. It is evident that the frequency
bands of same-index IMFs from different noise channels are
prominently overlapped, which confirms the mode alignment
property of the proposed FMEMD. Additionally, this graph
also reveals a number of striking features that are in agreement
with the results from the standard EMD [22], [23]: (i) The first
IMIMF possesses the characteristics of a high-pass filter while
higher order modes behave similarly to band-pass filters. (ii)
The energy balance among the different modes reflects the
behavior of the global spectrum, as a flat spectrum in case of
WGN.
The nature of FMEMD as a filter bank with respect to the
IMIMF index can be quantified further using the number of
zero crossings. According to [24], all extrema appear as an
alternation of local minima and maxima separated by only one
zero-crossing. Finding the average number of zero-crossings in
a mode is, therefore, a rough indication of its mean frequency,
and the manner this number varies from mode to mode is a
further indication of the hierarchical structure of the equivalent
filter bank. For standard EMD, it is illustrated empirically [22],
[23] that IMFs shown linear relationship (slope close to −1)
between the base-2 logarithm of number of zero crossings
and the IMF index. Likewise, the bottom of Fig. 2 gives
the fitting result obtained from the previous 500 four-channel
WGN realizations by using the base-2 logarithm of number
of the zero crossings against the IMIMF index. As shown
in this figure, the slope of approximately −1.0104 for all
four channels verifies the quasi-dyadic filter bank nature of
FMEMD. Therefore, the number of zero crossings z [k] is a
decreasing exponential function of the IMIMF index k,
z [k] ∝ ρ−k , (23)
with ρ very close to 2. In practice, parameter ρ can be
calculated from the slope of the straight line. The empirical
relationship between the number of zero crossings and the
bandpass filter index can be approximated by
log2 z [k] = mk + c, (24)
IEEE ACCESS
−10
−20
Spectrum(db)
−30
−40
−50
−60 0
1 2
10 10 10 10
Frequency
Fig. 3. Fourier spectra of IMIMFs from the proposed FMEMD with modes
3 − 8 shifted so as to overlap with the spectrum associated with mode 2.
where m denotes the slope, c represents the y-intercept of the
straight line and is dependent on the length of the data been
investigated. Accordingly, the base-2 exponent of Eq. (24) is
given below
z [k] = 2mk · 2c = 2−m·(−k) · 2c . (25)
In this simulation, the approximated slope is m = −1.0104.
Combining the above equation with Eq. (23), parameter ρ can
be therefore computed as
ρ = 2−m = 21.0104 = 2.0145. (26)
This regression reveals that the proposed FMEMD is em-
pirically a dyadic filter capable of separating the WGN into
IMIMF components having mean frequency exactly half value
of the previous components.
Another important finding gained from the filter bank analy-
sis of the standard EMD is the self-similarity of its constituent
bandpass filters [22], [24]. To illustrate that the power spectra
of IMIMFs obtained from the proposed FMEMD also exhibit
this self similar property, the averaged spectra corresponding
to IMIMFs 2 − 8 are properly renormalized into a curve-
cluster with the same center frequency based on coefficient ρ.
Fig. 3 gives the redrew curves of these spectra. It is observed
that the power spectra of all IMIMFs can be collapsed onto
a single curve, when ρ based renormalization is conducted.
This diagram is in agreement with the EMD-base findings
reported in [22], [23], [24]. To this end, let Pk (f ) denotes
the frequency response of the kth IMIMF, self-similarity of the
bandpass filters for both EMD and FMEMD can be described
by
Pn (f ) = Pk ρn−k f ,


(27)
where n ≥ k ≥ 2. The numerical analysis in this section
supports the claim the proposed FMEMD is a consistent
generalization of the standard EMD.
V. C OMPARATIVE S TUDY BETWEEN MEMD AND FMEMD
The proposed FMEMD is a well developed extension of the
univariate EMD for time-frequency analysis of multivariate
data. Unlike traditional MEMD, the raised method is able to
not only remove the dependence of the computational com-
plexity on dimension of the input signal, but also intuitively
present a more desired decomposition result. To verify the
8
statements, a comparative study around the decomposition
performance of MEMD and FMEMD is elaborated.
These two methods are first compared based on deriving
their computational cost in extracting a single MIMF/IMIMF.
In addition, the noise-based filter bank property analysis of
MEMD is conducted, which is dedicated to confirm that
FMEMD acts more similar to the standard EMD than that
of MEMD. Finally, a low-sampling rate signal is synthesized
3
to compare them further. The Matlab codes of MEMD and
EMD for developing the proposed FMEMD are available
online4 . Throughout this work, the Halton and Hammersley
sequence is used for generating a set of K = 64 direction
vectors. The stoppage criterion is S = 5, and end effects
of all investigated signals are restrained in advance before
FMEMD/MEMD being applied.
A. Computational Cost Analysis
In this section, a simple comparison between the compu-
tational loads of MEMD and FMEMD is presented. It is
noteworthy that some other fixed-point arithmetic operations
are negligible compared with the sifting procedure5 . Similarly,
the infrequent matrix operation, Eq. (22), which is utilized for
the solution of IMIMF, also has very small effect on the total
computational time due to two reasons:
 −1
i) The time-consuming operation ST S in Eq. (22)
requires computing the inverse of a p × p matrix, ST S. In
contrast, the computation of cubic spline interpolation needs
to solve the inverse of a (n − 1)×(n − 1) matrix, Q (Eq. (45)),
where n denotes the number of maxima/minima extracted from
the input signal. It is clear that the value of n of a real-world
signal will far exceed the dimension of the signal p. Therefore,
the computational time of solution xo is much smaller than
the time cost of a single cubic spline interpolation.
ii) Unlike MEMD or FMEMD  that−1
repeats the cubic spline
interpolation for many times, ST S ST is computed only
once during the whole process of the FMEMD decomposition,
which greatly saves the computational time.
Based on the above statements, the sifting time based
comparison is deducted between MEMD and FMEMD, to
facilitate the complexity analysis.
Let us suppose a single operation of the univariate sifting
process for standard EMD requires T0 time. According to [9],
the multivariate envelope associated with a specified projection
is computed by applying the univariate cubic spline interpo-
lation channel-wise. Notice that the cubic spline interpolation
in a single sifting process occupies most of the computation
for both EMD and MEMD. Consequently, focus on each of
the direction vectors, the sifting time of MEMD of a p-variate
input is about p times of the univariate sifting time, T0 . If
4 The code of MEMD is obtained from: http://www.commsp.ee.ic.ac.uk/
∼mandic/research/emd.htm, and the code of EMD can be found in: http://
perso.ens-lyon.fr/patrick.flandrin/emd.html.
5 The sifting process includes all of the large computational operations,
such as extrema identification and cubic spline procedures, where the cubic
spline interpolation occupies most of the computation. In contrast, the rest
of MEMD/FMEMD only involves some simple operations like addition,
subtraction, multiplication and division.
IEEE ACCESS 9

totally K direction vectors are considered, the time cost of
MEMD for extracting a single MIMF is given by
Tm = nm · p · K · T0 ,
where Tm denotes the computation cost, and nm is the number
of sifting times consumed for one MIMF extraction. In contrast
to MEMD, the p-variate input of FMEMD is projected and
then operated univariable using the standard EMD. Therefore,
the total consumption of FMEMD for a p-variate IMIMF
computation only involves
the stopping criteria. However, not every channel of the real
world signal needs to have the same number of iterations. As
a result, enforcing the same number of iterations for all data
(28)
channels may excessively sift more extrema to the following
modes, thus causes the phenomenon of over-decomposition.
(ii) According to ref [9], the condition for equality of the
number of extrema and zero crossings is not imposed for stop-
ping of the MEMD based sifting process, as extrema cannot be
properly defined for multivariate signals. This may also lead
to incomplete rationality in the decomposition process.

Tf = n̄f · K · T0 (29) Averaged spectra of white noise realizations from MEMD

MIMF 8 7 6 5 4 3 2 1
time, where Tf denotes the time cost, while n̄f represents

Spectrum
K
1
P
an average of the sifting times, namely, n̄f = K ni and
i=1
ni , i = 1, 2, . . . , K represents the respective sifting times for
standard EMD applying on the corresponding projected func- 10
0
10
1
10
2

tions. As discussed in [9], the stoppage criterion of multivariate Frequency(Log)

Zero crossings(log 2)
10
IMFs is similar to that applied for the standard IMF, thus it is
8
concluded that nm = max {ni , i = 1, 2, . . . , K}. To this end,
6
combine Eq. (28) with Eq. (29), a new formulation is given MEMD approximate slope = −0.74087
4
Tm ≥ p · Tf . (30) 2
1 2 3 4 5 6 7 8
MIMF index
This inequality illustrates that running of the proposed
FMEMD is p times or more faster than that of the MEMD. In
Fig. 4. Top: averaged spectra of MIMFs (MIMF1-MIMF8 from right to left)
practice, according to our experiments, the multiple is much obtained by MEMD. Bottom: averaged number of base-2 logarithm of zero
larger than p. crossings plotted versus the MIMF index.

In summary, it can be concluded from this simulation that

B. Comparison on Filter Bank Property
In order to demonstrate that FMEMD outperforms MEMD
in multivariate signal decomposition, the frequency responses
and corresponding filter bank properties of MEMD are inves-
tigated as a comparison.
Similar to Section IV, MEMD is applied on 500 realizations
of four-channel white Gaussian noise with each of length
N = 1, 000. Fig. 4 gives the averaged spectra (top) and
fitting results using the base-2 logarithm of number of the
zero crossings against the MIMF index (bottom). It seems that
MEMD is able to properly align the bandpass filters associated
with the corresponding MIMFs for different noise channels.
However, MEMD fails to provide its frequency responses
similar to that of EMD. More precisely, zero crossings in these
IMFs cannot follow the structure of a dyadic filter bank, which
shows linear (with slope close to −1) relationship between
the base-2 logarithm of number of zero crossings and the
MIMF index. The uneven blue-points in the bottom of Fig.
4 (with slope around −0.7409) have confirmed the statement.
Notice that the approximate slope of MEMD for eight-channel
noise is evaluated as close to -0.92 in [12]. This significant
difference may due to different noise realizations and channels,
or procedure of MEMD, which means the MEMD algorithm
is not very stable for noise decomposition. However, even for
a slope of -0.92, our proposed FMEMD still shows superiority.
There are two potential defects of the MEMD algorithm that
may lead to this poor performance:
(i) In MEMD, the (approximately) zero mean rotational
signal is obtained only after all the projected signals fulfill
the alignment of IMIMF based frequency bands, in case of
FMEMD, is more capable to result in the stabilization of the
dyadic filter bank structure.
C. Low-sampling Rate Signal Processing
The basic idea behind EMD is to evaluate the local en-
velopes by signal interpolation, using the local extrema as in-
terpolation points. Given an specified interpolating technique,
high envelope reconstruction fidelity requires the abundant and
accurate information of extreme points. Under low sampling
rate the extrema cannot be located exactly by searching the
sampling sequence [25]. Therefore, most researchers believe
that a successful application of the EMD algorithm requires
high sampling frequency [26].
Similar to EMD, the operations of both MEMD and
FMEMD rely highly on (multivariate) extrema identification.
As a result, the sampling rate which is required beforehand is
expected to affect the decomposition results of both methods.
Accordingly, the purpose of the section is to investigate
and compare the influences of sampling rate on these two
algorithms by decomposing a synthesized multivariate signal,
which is established as

 x1 = 5 cos (2πf1 t) + 2.5 (cos (2πf2 t) + cos (2πf3 t))



 + sin (2πf4 t)
 x2 = 2.5 (sin (2πf2 t) + cos (2πf3 t) + sin (2πf4 t))


x3 = 2.5 (cos (2πf1 t) + cos (2πf2 t) + cos (2πf3 t)) . (31)
 x 4 = 5 cos (2πf1 t) + 2.5 (cos (2πf2 t) + cos (2πf4 t))


 x5 = 2.5 (cos (2πf2 t) + cos (2πf3 t))




x6 = 2.5 (sin (2πf1 t) + cos (2πf2 t))
IEEE ACCESS
The frequencies of the signal are as follows: f1 = 2Hz,
f2 = 8Hz, f3 = 16Hz, f4 = 32Hz. The number of sampling
points is set to 100, with sampling frequency fs = 100Hz.
According to the conclusion of [25], such sampling rate is
quite low under which the available EMD may not work
efficiently. Both of MEMD and FMEMD are applied on x1
to x6 simultaneously, and decomposition results are shown
in Fig. 5 and Fig. 6, respectively. Obviously, neither method
yields a perfect decomposition result, especially in the first
two modes. Such decomposition performance is expected
because the original signal cannot provide accurate extrema
information due to the low sampling frequency.
On the other hand, the proposed method shows more desired
results than that from the MEMD method. As shown in the
first two MIMFs of Fig. 5, large distortion and leakage are
presented all over the data channel, especially those in d2
of x1 , x2 , x3 and x5 . In contrast to the MEMD method, the
proposed FMEMD is less prone to leakage and sheds better
light on the intrinsic characteristic of modes id1 and id2 . The
explanation to the poor performance of MEMD is similar to
the description in Section. V-B. The MEMD algorithm may
compromise the T-F information in low sampling rate, since it
enforces the same number of iterations for all data channels.
VI. R EAL W ORLD A PPLICATIONS
Due to the ability of MEMD on time-frequency analysis
of multivariate nonstationary signals, it has found a number
of different real world applications from various fields, such
as signal filtering [4], biomedical engineering [27], image
analysis [12], industrial controlling [11] and others. Similarly,
the proposed FMEMD as a more precise extension of the stan-
dard EMD, can also be effectively used in these areas. Since
it has demonstrated that FMEMD outperforms MEMD in
multivariate signal decomposition and operating speed, more
widespread applications of the EMD-base extensions can be
expected. In this section, three representative cases including:
(i) plant-wide oscillation detection in process industry, (ii)
electroencephalography (EEG) analysis and (iii) image fusion
application, are added to verify the above statements.
A. Plant-wide Oscillation Detection in Industrial Process
A typical industrial case is presented to demonstrate the
effectiveness of the proposed method for plant-wide oscillation
characterization. Plant-wide oscillation is one of the most com-
mon abnormal phenomena encountered in process industries
[28]. The presence of oscillatory variables may severely impact
the normal operation of the process plant, therefore, it strongly
motivates the research for techniques to automatically detect
the plant-wide oscillations [29].
The data set under study is provided by a plant of Eastman
Chemical Company [30]. The uncompressed nonlinear and
nonstationary data were sampled from the control system every
20 seconds on each of the indicators. In order to facilitate
the results for displaying, only the last 10 channels data
(denoted by Tag 21 to 30) from total 30 indicators with 2, 000
samples are investigated. The decomposition results obtained
10
by FMEMD are presented in Fig. 7, where each Tag denotes
a single data channel or signal dimension. For simplicity, the
first five adjacent IMIMFs are added together. The top row
of Fig. 7 gives the process output, and the second and third
rows are id1 − id5 and id6 respectively. The fourth row (id7 )
denotes the plant-wide oscillations. id8 is shown in the fifth
row and last row (r) is the remainder of the process data.
Generally, an oscillation is considered to be regular if the
standard deviation σT of its ACF zero-crossing intervals Ti is
less than one third of the mean value µT ([28]). This regularity
index is defined as
1 µT
z= × . (32)
3 σT
A signal is concluded to be oscillatory if its coefficient
variable satisfies z > 1. Table. I depicts the monitoring results
of id6 and id7 , respectively. id8 is not taken into account
since none of the significant components exists in this scale.
Obviously, an unignorable plant-wide oscillation is correctly
detected in mode id7 . Several Tags (22 to 29) show greatly
indexes, while the last two cannot detect obvious oscillation.
A reasonable explanation is that Tag 21 is dominated by
the nonstationary trend, and Tag 30 is caught into some
intensive noise. In addition, combining Table. I with the
time-frequency characteristics of these modes shows that a
slow plant-wide oscillation with a period around 115 min is
presented throughout this plant. The interpretation in [30] that
it is a prominent and widespread disturbance that accounts for
most of the process variability, further confirms the detecting
results accomplished by FMEMD.
B. EEG Data Analysis
Electroencephalogram (EEG) is one of the most efficient
measurements for learning the function and structure of brain
with its activity due to the advantages of non-invasive, con-
venience and low-cost. Since different regions of the brain
may need cooperation with sensory stimulation or process of
cognitive activity, information in local area (single channel)
cannot fully represent the whole perceptual pattern. Therefore,
it is necessary to extract the common information from distinct
regions to analyze the global pattern. In order to demonstrate
the advantages of the proposed FMEMD in multichannel EEG
processing, a set of four-channel data is analyzed with an aim
to extract the synchronized information from each scale of
IMIMFs. The EEG data was recorded from 4 channels (F3, F4,
P3, P4) by a Neuroscan NuAmps Digital Amplifier, referenced
to the left mastoid, and sampled at 1, 000Hz with an analog
filtering pass-band of 0.1 − 100Hz. This configuration of EEG
recorder is very common in clinical and research applications
[31], [32].
Fig. 8 shows results of the decomposed EEG from IMIMF1
to IMIMF6 with a residual. Additionally, the right column
of Fig. 8 gives the power spectra of individuals of each
mode from various channels, where the frequency scopes
of id3 to id6 are narrowed to better illustrate the spectral
meaning of this decomposition. It is worthy to notice that the
measured EEG from all channels contains four main spectral
peaks located in beta (15 ∼ 30Hz), alpha (8 ∼ 14Hz),
IEEE ACCESS 11

x1 x2 x3 x4 x5 x6
10 10 10 10 10 10
0 0 0 0 0 0
−10 −10 −10 −10 −10 −10
5 5 5 5 5 5
d1

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
d2

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
d3

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
d4

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
0 0 0 0 0 0
r

−5 −5 −5 −5 −5 −5
0 50 100 0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Sample

Fig. 5. Decomposition results of the synthesized signal using MEMD.

x1 x2 x3 x4 x5 x6
10 10 10 10 10 10
0 0 0 0 0 0
−10 −10 −10 −10 −10 −10
5 5 5 5 5 5
id1

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
id2

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
id3

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
id4

0 0 0 0 0 0
−5 −5 −5 −5 −5 −5
5 5 5 5 5 5
0 0 0 0 0 0
r

−5 −5 −5 −5 −5 −5
0 50 100 0 50 100 0 50 100 0 50 100 0 50 100 0 50 100
Sample

Fig. 6. Decomposition results of the synthesized signal using the proposed FMEMD.

TABLE I
O SCILLATION DETECTION RESULTS OF THE DECOMPOSED IMIMF S .

Variable Tag 21 Tag 22 Tag 23 Tag 24 Tag 25 Tag 26 Tag 27 Tag 28 Tag 29 Tag 30
id6 0.3636 0.5686 0.8246 0.8074 0.7499 0.6354 0.5218 0.6373 0.9437 0.6667
id7 0.6114 2.9346 3.0956 3.0562 2.6970 3.0569 1.7429 1.3659 3.0348 0.7561
IEEE ACCESS 12

Tag21 Tag22 Tag23 Tag24 Tag25 Tag26 Tag27 Tag28 Tag29 Tag30
5
0
−5
5
id1−id5
0
−5
5
id6

0
−5
5
id7

0
−5
5
id8

0
−5
5
0
r

−5
0 1000 0 1000 0 1000 0 1000 0 1000 0 1000 0 1000 0 1000 0 1000 0 1000
Sample

Fig. 7. Decomposition results of the process outputs using FMEMD.

theta (4 ∼ 7Hz) and delta (< 4Hz) bands, which are well
explained by id3 , id4 , id5 and id6 respectively. Based on
the aligned common patterns, the underlying brain activity
regarding different research purposes [31], [32] can be further
analyzed. On the other hand, the remaining modes of this de-
composition also have their corresponding physical meanings:
(i) id1 : high frequency components which are not completely
cleared by pass-band filter; (ii) id2 : power-line interference in
EEG recordings.
The FMEMD method enables features of the signal within
the same subband from different channels to be categorized in
the same mode which allows the neural information in differ-
ent frequency bands can be studied across different channels.
The IMIMFs preserve not only the amplitude and frequency
features but also the phase characteristic of synthesized com-
ponents. Unlike other advanced signal processing techniques
(e.g., independent component analysis (ICA) and principal
component analysis (PCA).) utilizing statistical algorithms to
extract common information, only the cross-channel common
components are considered which could ignore the phase
difference among different channels.
C. Image Fusion
Image fusion is the process of registering and combining
multiple images from different modalities to improve the
imaging quality, while reducing randomness and redundancy
such that the new image is more suitable for the purpose of
human visual perception and computer-processing tasks [33].
One of the major requirements for successful image fusion is
that images from the different modalities have to be aligned
correctly. Moreover, as real world images are usually nonlinear
and nonstationary, which render the traditional techniques
invalid, a fully adaptive approach is required [34].
Recently, the uses of multivariate extension of EMD [34],
[35], [36], [37] were proposed for the fusion of multiple out-
of-focus and multi-exposure images. It was shown that the bi-
variate EMD outperforms wavelet and PCA based approaches,
particularly in retaining edge-based information from different
image modalities. In this section, the proposed FMEMD is
adopted to merge two multi-exposure images to a single
improved image as an illustration.
The methodology employed for fusion of both gray-level
and RGB based colored images is similar to the one used
in [34]. More precisely, given two gray scale images, each
considered as a matrix and denoted by X and Y, the fusion
scheme is operated as follows:
(i) X and Y are first converted into row vectors by con-
catenating their rows.
(ii) Two such vectors are combined to form a bivariate
signal. Then the FMEMD method is applied to this data and
resulting in M IMIMFs.
(iii) These bivariate IMIMFs are separated and reconverted
back to image-matrices, which yield M scale images for each
input image, denoted by Xi and Yi , for i = 1, . . . , M .
M M
(iv) {Xi }i=1 and {Yi }i=1 are combined locally, based on
coefficients computed from the local variance estimates at each
spatial point, resulting in a fused image Z as given by
M
X
Z (m, n) = αi (m, n) Xi (m, n) + βi (m, n) Yi (m, n), (33)
i=1
where αi (m, n) and βi (m, n) are weights determined by
location (m, n), on the basis of the relative values between
local and total variance for each mode
var [Xi (m, n)]
αi (m, n) = , (34)
var [Xi (m, n)] + var [Yi (m, n)]
IEEE ACCESS 13

F3 F4 P3 P4
40
0
−40 Power spectrum
10
id1
0
−10
0 50 100 150
10
id2

0
−10
0 50 100 150
10
id3

0
−10
0 15 30
10
id4

0
−10
0 15 30
10
id5

0
−10
0 15 30
10
id6

0
−10
0 15 30
10
0 Frequency/ Hz
r

−10
0 500 1000 0 500 1000 0 500 1000 0 500 1000
Sample

Fig. 8. FMEMD of 1 s-long normal EEG of channel F3, F4, P3 and P4 (left) with corresponding power spectra of the decomposed IMIMFs (right).

and
var [Yi (m, n)]
βi (m, n) = , (35)
var [Xi (m, n)] + var [Yi (m, n)]
where var [Xi (m, n)] (var [Yi (m, n)]) represents the local
variance of Xi (Yi ) calculated across a small block around
(m, n). Similarly, the RGB based color image fusion can be
extended by applying the methodology to three channels (red,
green and blue) of a color image separately [34].
The performance of the proposed method is evaluated on
two multi-exposure images that are shown in Fig. 9 and Fig.
10 respectively. The first image is obtained from a work table,
which is severely underexposed. Observed that the small plate
in the right region is nearly hidden, whereas the left part,
showing details of the book, is well exposed. In the second
input image, on the other hand, the details on the book are
covered due to the strong light, but the right area is exposed Fig. 9. Image of a work table which is obtained from underexposed condition.
properly. Fig. 11 shows the result of the fusion obtained from
applying FMEMD based scheme. As desired, the fused image
retains all the detail of the original images and any spurious
The mode-aligned property of FMEMD contributes to the
fusion artifacts are kept into a minimum.
automatic clustering of similar frequency modes as well as
providing both time and frequency information in nonstation-
VII. C ONCLUSION ary time series. Compared with MEMD, operation of the
A novel multivariate extension of the standard EMD, proposed FMEMD is independent of the dimension of the
namely the fast MEMD (FMEMD), has been provided. multivariate signal, and this approach avoids the problem of
Three key concepts including the direction-dependent MIMF over-decomposition when processing on the white noise as
(DMIMF), the direction-independent MIMF (IMIMF) and the a filter bank. Moreover, numerical simulation shows that the
pseudo-IMIMF are first defined through properly extending the FMEMD method presents better decomposition performance
notion of multivariate extrema. Then FMEMD is generalized in processing data with low sampling frequency. Such superi-
from the univariate EMD by solving an overdetermined system orities can facilitate more widespread applications of the mul-
of linear equations. It is shown numerically that the proposed tivariate EMD method, especially for the implementations in
FMEMD, similar to the standard EMD, follows a dyadic filter analyzing long data and signals in which the dyadic filter bank
bank structure (channel-wise) for multivariate white Gaussian decomposition is relevant. Simulations and several real world
noise. cases have verified the effectiveness of proposed method.
IEEE ACCESS 14

S (α, t) and S (β, t) of point sets α and β satisfy the following

two properties:
(1) Homogeneity: S (kα, t) = k · S (α, t);
(2) Additivity: S (α + β, t) = S (α, t) + S (β, t)
where k ∈ R is an arbitrary real number. Since the three-
moment algorithm with not-a-knot end condition is used in
MATLAB for solving of the piecewise cubic spline interpola-
tion, this section proves the above properties on the basis of
the same configuration.

A. Proof of Homogeneity
Let us suppose s (t) = S (α, t), s∗ (t) = S (kα, t), and
n n
{Mi }i=0 , {Mi∗ }i=0 denote the second derivatives of s (t) and
s∗ (t) on time instants τ = {t0 , t1 , . . . , tn } respectively.
(i) Calculate the related parameters.
Fig. 10. Image of a work table which is obtained from overexposed condition. hi = ti+1 − ti , (36)
f [ti , ti+1 ] = xi+1hi−xi
(
−xi ) , (37)
f ∗ [ti , ti+1 ] = k(xi+1 hi = kf [ti , ti+1 ]
where i = 0, 1, · · · , n − 1. Accordingly, some spline-related
parameters can be further computed
(
hi−1 ∗
µi = hi−1 +hi =µi
∗ , (38)
λi = 1 − µi = λi

f [ti−1 ,ti ]−f [ti ,ti+1 ]
 f [ti−1 , ti , ti+1 ] = ti−1 −ti+1
f [ti−1 ,ti ]−f ∗ [ti ,ti+1 ]
∗ ,
 f ∗ [ti−1 , ti , ti+1 ] = =kf [ti−1 , ti , ti+1 ]
ti−1 −ti+1
(39)

di = 6f [ti−1 , ti , ti+1 ]
, (40)
di ∗ = 6f ∗ [ti−1 , ti , ti+1 ] = kdi
where i = 1, 2, · · · , n−1. According to [38], these parameters
satisfy the following equation
Fig. 11. The fused image obtained by employing the proposed FMEMD.
µi Mi−1 + 2Mi + λi Mi+1 = di , (41)
n
from i = 1, 2, · · · , n − 1. Therefore, moments {Mi }i=0 can
The proposed method may be improved from following be obtained by solving the above system of linear equations,
aspects: (i) a more efficient EMD algorithm, (ii) algorithm which form the spline interpolation function as given by
to solve the overdetermined system of linear equations, (iii) 3 3
method for generating a more suitable set of direction vectors (ti+1 − t) (t − ti )
s(t) = Mi + Mi+1
(EMD cannot decompose invalid projection, for example when 6hi 6hi
, (42)
the energy of the highest frequency component is much Mi hi 2 ti+1 − t Mi+1 hi 2 t − ti
smaller/larger than that of the remained components) and (iv) a + [ai − ] + [ai+1 − ]
6 hi 6 hi
more scientific criterion to determine the number of directions.
where t ∈ [ti , ti+1 ] , i = 0, 1, . . . , n − 1. Notice that s∗ (t) and
n
{Mi∗ }i=0 show the similar relationship.
A PPENDIX (ii) Set boundary conditions.
P ROOF OF THE L EMMA 1 Since Eq. (41) has n + 1 unknown numbers but only n − 1
equations, the two ends M0 and Mn should be developed in
Lemma 1. The projection of the cubic spline interpolation
advance. Notice that the not-a-knot end condition [39] requires
curve of a multi-dimensional point set is equivalent to the
the third derivatives of the spline function at the first and
interpolation that is calculated from the projected point set.
last interior break are continuous respectively. The boundary
Proof. Let ∆ := {a = t0 < t1 < · · · < tn = b} be a parti- conditions upon M0 and Mn can therefore be calculated as
tion of the time interval [a, b], α = {a0 , a1 , . . . , an } and 
M0 = λ11 (M1 − µ1 M2 )
β = {b0 , b1 , . . . , bn } represent two different point sets that 1 . (43)
Mn = µn−1 (Mn−1 − λn−1 Mn−2 )
are defined on time instants τ = {t0 , t1 , . . . , tn } respectively,
then it can be proved that the univariate cubic spline functions Likewise, M0∗ and Mn∗ meet the similar equations.
IEEE ACCESS 15

n n
(iii) Solving of {Mi }i=0 and {Mi∗ }i=0 . (iv) Solving of the moments.
Combining Eq. (38), (41) and (43), a system of linear Substitute Eq. (51) into the corresponding cubic functions
n
equations for moments {Mi }i=0 can be consisted sα (t), sβ (t), sα+β (t) that are defined by Eq. (42), it is
µ1 µ1
2+ 1− obtained that
 
λ1 λ1 
M1
 
d1

 µ2 2 λ2 
 M2   d2 

S(α + β, t) = sα+β (t) = sα (t) + sβ (t) = S(α, t) + S(β, t).

    

. . .
 .   . 
. . .  =  .
    

. . .
 . .
  .   . 



µn−2 2
µn−1
λn−2
µn−1


 Mn−2



 dn−2

 (52)
1− 2+ Mn−1 dn−1
λn−1 λn−1
(44)
Consequently, Eq. (47) with Eq. (52) proved that the cubic
Simplify the above equation into the form of QM = D. spline interpolation operator satisfies the operations of addition
Since matrix Q is symmetric, positive definite, and diagonally (additivity) and scalar multiplication (homogeneity) [41].
dominant, the linear system is uniquely solvable [40], and the
solution is represented as C. Proof of the Lemma
M = Q−1 D. (45) Considering a predefined time grid ∆ :=
{a = t0 < t1 < · · · < tn = b} that is spaced on interval
Eq. (38) and Eq. (40) indicate that λ∗i = λi , µ∗i = µi and [a, b], the relationship between the cubic spline functions of
di ∗ = kdi for i = 1, 2, . . . , n − 1. Therefore, it is concluded a p-variate point set x = [x1,i , x2,i , ..., xp,i ], i = 0, 1, · · · , n
that Q∗ = Q and D∗ = kD, thus the solution of moments and its projected point set along direction vector
n
{Mi∗ }i=0 can be computed further T
v = [v1 , v2 , ..., vp ] is studied.
T T
M∗ = (Q∗ )−1 D∗ = kQ−1 D = kM. (46) Let x1 = [x1,0 , x1,1 , ..., x1,n ] , x2 = [x2,0 , x2,1 , ..., x2,n ] ,
T
..., and xp = [xp,0 , xp,1 , ..., xp,n ] denote the p components
(iv) Constructing s (t) and s∗ (t). of x, the cubic spline interpolation of the current point set can
n n
Applying {Mi }i=0 and {Mi∗ }i=0 to the piecewise cubic be computed channel-wise [9], which gives
functions s (t) and s∗ (t) that are defined into Eq. (42), a new
conclusion can be obtained, which is sx (t)=S (x, t) = [S (x1 , t) , S (x2 , t) , . . . , S (xp , t)]
  . (53)
∗
= sx1 (t) , sx2 (t) , . . . , sxp (t)
S(kα, t) = s (t) = k · s(t) = k · S(α, t). (47)
Project sx (t) onto preset direction vector v
B. Proof of Additivity sx (t) · v=sx1 (t) v1 + sx2 (t) v2 + · · · + sxp (t) vp . (54)
Similarly, let sα (t) = S (α, t), sβ (t) = S (β, t),
sα+β
(t) = S (α + β, By supposing xv represents the projected point set of x on
nt), then
on introduce
n the corresponding
on mo-
the same direction, the formula of the projection is therefore
α n β α+β
ments as {Mi }i=0 , Mi and Mi respectively. denoted as
i=0 i=0
(i) Parameters calculation.
Similar to A-(i), the moment-related parameters show fol- xv = x1 v1 + x2 v2 + · · · + xp vp . (55)
lowing relationships Applying the cubic spline operator on xv , then yields the
 α β α+β spline function as
 µi = µi = µi
α β α+β
λ =λ =λ , (48) sxv = S (x1 v1 + x2 v2 + · · · + xp vp , t) . (56)
 iα+β i α i β
di = di + di
According to A and B, the cubic spline interpolation oper-
where i = 1, 2, · · · , n − 1. ator satisfies additivity and homogeneity, thus Eq. (56) can be
(ii) Boundary conditions. rewritten into
Calculation of the two ends of the moments is the same as
that presented in A-(ii). sxv = S (x1 v1 , t) +S (x2 v2 , t) + · · · +S (xp vp , t)
(iii) Solving of the moments. = S (x1 , t) v1 +S (x2 , t) v2 + · · · +S (xp , t) vp . (57)
By extending from Eq. (44), these moments can be solved = sx1 (t) v1 + sx2 (t) v2 + · · · + sxp (t) vp
as  −1
 Mα = (Qα ) Dα Combine Eq. (54) with Eq. (57), it is concluded that

−1 β
Mβ = Qβ D . (49) sx (t) · v=S (x, t) · v = S (xv , t) = sxv (t) . (58)
α+β −1 α+β
 α+β


M = Q D

This equality proves that the projection of a multivariate
Combine Eq. (44) with Eq. (48), the above matrices satisfy spline function that is calculated from a multivariate point set,
 α+β
Q =Qα =Qβ is equal to the univariate spline function interpolated from
α+β . (50) the projected point set along the same direction vector. The
D =Dα +Dβ
presented lemma has been proved mathematically.
The relationship among M α , M β and M α+β is therefore
given as R EFERENCES

−1 α+β
Mα+β = Qα+β D [1] H. Hao, H. Wang, and N. Rehman, “A joint framework for multivariate

−1 β . (51) signal denoising using multivariate empirical mode decomposition,”
−1
= (Qα ) Dα + Qβ D =Mα +Mβ Signal Processing, vol. 135, pp. 263–273, 2017.
IEEE ACCESS
[2] D. P. Mandic, N. ur Rehman, Z. Wu, and N. E. Huang, “Empirical mode
decomposition-based time-frequency analysis of multivariate signals:
The power of adaptive data analysis,” IEEE Signal Processing Magazine,
vol. 30, no. 6, pp. 74–86, 2013.
[3] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-
C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition
and the hilbert spectrum for nonlinear and non-stationary time series
analysis,” in Proceedings of the Royal Society of London A: Mathemati-
cal, Physical and Engineering Sciences, vol. 454, no. 1971. The Royal
Society, 1998, pp. 903–995.
[4] J. Fleureau, A. Kachenoura, L. Albera, J.-C. Nunes, and L. Senhadji,
“Multivariate empirical mode decomposition and application to multi-
channel filtering,” Signal Processing, vol. 91, no. 12, pp. 2783–2792,
2011.
[5] Y. Lei, J. Lin, Z. He, and M. J. Zuo, “A review on empirical mode
decomposition in fault diagnosis of rotating machinery,” Mechanical
Systems and Signal Processing, vol. 35, no. 1, pp. 108–126, 2013.
[6] T. Tanaka and D. P. Mandic, “Complex empirical mode decomposition,”
IEEE Signal Processing Letters, vol. 14, no. 2, pp. 101–104, 2007.
[7] G. Rilling, P. Flandrin, P. Gonçalves, and J. M. Lilly, “Bivariate empirical
mode decomposition,” IEEE Signal Processing Letters, vol. 14, no. 12,
pp. 936–939, 2007.
[8] N. ur Rehman and D. P. Mandic, “Empirical mode decomposition for
trivariate signals,” IEEE Transactions on Signal Processing, vol. 58,
no. 3, pp. 1059–1068, 2010.
[9] ——, “Multivariate empirical mode decomposition,” in Proceedings of
the Royal Society of London A: Mathematical, Physical and Engineering
Sciences. The Royal Society, 2009, p. rspa20090502.
[10] Y. Lv, R. Yuan, and G. Song, “Multivariate empirical mode decomposi-
tion and its application to fault diagnosis of rolling bearing,” Mechanical
Systems and Signal Processing, vol. 81, pp. 219–234, 2016.
[11] X. Lang, D. Zhong, L. Xie, and J. Chen, “Application of improved
multivariate empirical mode decomposition to plant-wide oscillations
characterization,” in Advanced Control of Industrial Processes (Ad-
CONIP), 2017 6th International Symposium on. IEEE, 2017, pp. 601–
606.
[12] N. ur Rehman and D. P. Mandic, “Filter bank property of multivariate
empirical mode decomposition,” IEEE Transactions on Signal Process-
ing, vol. 59, no. 5, pp. 2421–2426, 2011.
[13] A. S. Bahri, D. D. Warnana, F. A. M. Santos, B. J. Santosa et al., “Fast,
simultaneous and robust vlf-em data denoising and reconstruction via
multivariate empirical mode decomposition,” Computers & Geosciences,
vol. 67, pp. 125–138, 2014.
[14] Y. H. Wang, C. H. Yeh, H. W. V. Young, K. Hu, and M. T. Lo,
“On the computational complexity of the empirical mode decomposition
algorithm,” Physica A Statistical Mechanics & Its Applications, vol. 400,
no. 2, pp. 159–167, 2014.
[15] A. Ahrabian, N. ur Rehman, and D. P. Mandic, “Bivariate empirical
mode decomposition for unbalanced real-world signals.” IEEE Signal
Processing Letters, vol. 20, no. 3, pp. 245–248, 2013.
[16] N. ur Rehman, M. W. Safdar, U. ur Rehman, and D. P. Mandic,
“Dynamically-sampled bivariate empirical mode decomposition,” IEEE
Signal Processing Letters, vol. 21, no. 7, pp. 857–861, 2014.
[17] N. ur Rehman, K. Naveed, M. Safdar, S. Ehsan, and K. McDonald-
Maier, “Dynamically sampled multivariate empirical mode decomposi-
tion,” Electronics Letters, vol. 51, no. 24, pp. 2049–2051, 2015.
[18] J. Stewart, Calculus. Cengage Learning, 2011.
[19] N. E. Huang, M.-L. C. Wu, S. R. Long, S. S. Shen, W. Qu, P. Gloersen,
and K. L. Fan, “A confidence limit for the empirical mode decomposition
and hilbert spectral analysis,” in Proceedings of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences, vol. 459,
no. 2037. The Royal Society, 2003, pp. 2317–2345.
[20] H. Niederreiter, Random number generation and quasi-Monte Carlo
methods. SIAM, 1992.
[21] J. A. Cadzow, “Minimum `1 , `2 and `∞ norm approximate solutions
to an overdetermined system of linear equations,” Digital Signal Pro-
cessing, vol. 12, no. 4, pp. 524–560, 2002.
[22] P. Flandrin, G. Rilling, and P. Goncalves, “Empirical mode decompo-
sition as a filter bank,” IEEE Signal Processing Letters, vol. 11, no. 2,
pp. 112–114, 2004.
[23] Z. Wu and N. E. Huang, “A study of the characteristics of white noise
using the empirical mode decomposition method,” in Proceedings of
the Royal Society of London A: Mathematical, Physical and Engineering
Sciences, vol. 460, no. 2046. The Royal Society, 2004, pp. 1597–1611.
[24] P. Flandrin, P. Gonçalves, and G. Rilling, “Emd equivalent filter banks,
from interpretation to applications,” in Hilbert–Huang transform and its
applications. World Scientific, 2014, pp. 99–116.
16
[25] G. Rilling and P. Flandrin, “On the influence of sampling on the empir-
ical mode decomposition,” in Acoustics, Speech and Signal Processing,
2006. ICASSP 2006 Proceedings. 2006 IEEE International Conference
on, vol. 3. IEEE, 2006, pp. III–III.
[26] Z. Xu, B. Huang, and F. Zhang, “Improvement of empirical mode
decomposition under low sampling rate,” Signal Processing, vol. 89,
no. 11, pp. 2296–2303, 2009.
[27] C. Park, D. Looney, N. ur Rehman, A. Ahrabian, and D. P. Mandic,
“Classification of motor imagery bci using multivariate empirical mode
decomposition,” IEEE Transactions on Neural Systems and Rehabilita-
tion Engineering, vol. 21, no. 1, pp. 10–22, 2013.
[28] N. F. Thornhill, B. Huang, and H. Zhang, “Detection of multiple
oscillations in control loops,” Journal of Process Control, vol. 13, no. 1,
pp. 91–100, 2003.
[29] X. Lang, Z. Zhang, L. Xie, A. Horch, and H. Su, “Time-frequency
analysis of plant-wide oscillations using multivariate intrinsic time-scale
decomposition,” Industrial & Engineering Chemistry Research, vol. 57,
no. 3, pp. 954–966, 2018.
[30] N. F. Thornhill, J. W. Cox, and M. A. Paulonis, “Diagnosis of plant-wide
oscillation through data-driven analysis and process understanding,”
Control Engineering Practice, vol. 11, no. 12, pp. 1481–1490, 2003.
[31] T. Zhang, L. Dai, Y. Wang, W. J. Freeman, and G. Li, “Eeg spatiotempo-
ral pattern classification of the stimuli on different fingers,” in Advances
in Cognitive Neurodynamics (IV). Springer, 2015, pp. 147–153.
[32] G. Li, T. Zhang, X. Chen, C. Shang, and Y. Wang, “Effect of intermittent
hypoxic training on hypoxia tolerance based on brain functional connec-
tivity,” Physiological Measurement, vol. 37, no. 12, p. 2299, 2016.
[33] H. Li, B. Manjunath, and S. K. Mitra, “Multisensor image fusion using
the wavelet transform,” Graphical models and image processing, vol. 57,
no. 3, pp. 235–245, 1995.
[34] N. ur Rehman, D. Looney, T. Rutkowski, and D. Mandic, “Bivariate
emd-based image fusion,” in Statistical Signal Processing, 2009. SSP’09.
IEEE/SP 15th Workshop on. IEEE, 2009, pp. 57–60.
[35] D. Looney and D. P. Mandic, “Multiscale image fusion using complex
extensions of emd,” IEEE Transactions on Signal Processing, vol. 57,
no. 4, pp. 1626–1630, 2009.
[36] N. ur Rehman, S. Ehsan, S. M. U. Abdullah, M. J. Akhtar, D. P. Mandic,
K. D. McDonald-Maier et al., “Multi-scale pixel-based image fusion
using multivariate empirical mode decomposition,” Sensors, vol. 15,
no. 5, pp. 10 923–10 947, 2015.
[37] S. M. U. Abdullah, N. ur Rehman, M. M. Khan, and D. P. Mandic, “A
multivariate empirical mode decompositionbased approach to pansharp-
ening,” IEEE Transactions on Geoscience and Remote Sensing, vol. 53,
no. 7, pp. 3974–3984, 2015.
[38] J. Stoer and R. Bulirsch, Introduction to numerical analysis. Springer
Science & Business Media, 2013, vol. 12.
[39] C. De Boor, “Convergence of cubic spline interpolation with the not-
a-knot condition.” WISCONSIN UNIV-MADISON MATHEMATICS
RESEARCH CENTER, Tech. Rep., 1985.
[40] K. E. Atkinson, An introduction to numerical analysis. John Wiley &
Sons, 2008.
[41] R. Bellman, Introduction to matrix analysis. Siam, 1997, vol. 19.
Xun Lang was born in 1994. He received the
B.S. degree in automation from Zhejiang Univer-
sity, HangZhou, China, in 2014. He is currently
pursuing the Ph.D. degree with Zhejiang University,
Hangzhou, China. He is currently with the Institute
of Cyber-Systems and Control, Zhejiang University.
His current research interests include, but are not
limited to, process control performance monitoring,
signal processing, time-frequency analysis.
IEEE ACCESS 17

Qian Zheng was born in 1995. He received a
B.S. degree in the Department of Automation from
Harbin Engineering University, Harbin, China, in
2017. He is currently pursuing the Ph.D. degree
at the Institute of Cyber-Systems and Control in
Zhejiang University, Hangzhou, China. His current
research interests include fault detection and diag-
nosis, image texture analysis, signal processing.
Zhiming Zhang was born in 1994. He received a
B.S. degree in automation from Zhejiang University,
Hangzhou, China, in 2017. He is currently pursuing
the Ph.D. degree at the Institute of Cyber-Systems
and Control in Zhejiang University. His current
research interests include model predictive control,
autonomous vehicle, process control performance
monitoring.
Hongye Su received his Ph.D. degree from Zhejiang
University in 1995. He was appointed as a professor
in Dec. 2000 in the Institute of Advanced Control
in Zhejiang University and as the deputy director of
the Institute of Advanced Control from Oct. 1999
to Aug. 2008. Now, he is the deputy director of the
Institute of Cyber-systems and Control in Zhejiang
University.
His research interests include Process Control &
Optimization Theory and Application. He is a Senior
Member of IEEE, and his IEEE Membership starts
from 2010. He is the author/coauthor of 100 journal papers, 4 books. Professor
Su has undertaken more than ten national and provincial funds and state
key projects as the principal investigator (PI), including National Outstanding
Youth Science Foundation of China, National Natural Science Foundation
of China, Teaching and Research Award Program for Outstanding Young
Teachers in Higher Education Institutions of MOE, P.R.C., and projects for
National High Technology Research and Development Program of China (863
Program), etc.

Shan Lu received the B.S. degree in 2011 and

the Ph.D. degree in 2016 from Zhejiang University,
China. He is currently an assistant professor with
the Institute of Intelligence Science and Engineering,
Shenzhen Polytechnic. His research interests include
modeling and optimization of supply chain schedul-
ing, machine learning, and process modeling.

Lei Xie received a B.S. degree in 2000 and a

Ph.D. degree in 2005 from Zhejiang University, P.R.
China. He is currently a full Professor at College of
Control Science and Engineering, Zhejiang Univer-
sity. To date, his research activities culminated in
over 50 articles that are published in internationally
renowned journals and conferences, 3 book chapters
and a book in the areas of process control, model
predictive control and applied multivariate statistics.
His research interests focus on the interdisciplinary
area of statistics and cyber system control theory.

Alexander Horch received his M. Sc. degree

in engineering cybernetics from the University of
Stuttgart in 1996 and Ph.D. degree in automatic
control from Royal Institute of Technology Stock-
holm, 2000. He worked at ABB Corporate Research
Germany as scientist and project leader in the area
of optimization and asset management in different
process industries. Since 2007, he has held several
management roles in the areas of production op-
timization, industrial and power automation, smart
electrical grids, and industrial control systems. He
is a certified project management professional and a certified automation
professional from ISA. His current interests include the concretization of
cyber-physical systems to electrical power systems and industrial production.

View publication stats