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

net/publication/315939366

A Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals

Article  in  IEEE Access · March 2017

DOI: 10.1109/ACCESS.2017.2688467

CITATIONS READS

11 978

1 author:

Yanli Yang
Tianjin Polytechnic University
18 PUBLICATIONS   121 CITATIONS   

SEE PROFILE

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

Investigation on the Theory and Methods of Empirical Mode Decomposition View project

All content following this page was uploaded by Yanli Yang on 26 June 2017.

The user has requested enhancement of the downloaded file.

Received February 12, 2017, accepted March 22, 2017, date of publication March 31, 2017, date of current version May 17, 2017.
Digital Object Identifier 10.1109/ACCESS.2017.2688467

A Signal Theoretic Approach for Envelope

Analysis of Real-Valued Signals
YANLI YANG
Tianjin Key Laboratory of Optoelectronic Detection Technology and Systems, Tianjin Polytechnic University, Tianjin 300387, China
(yyl070805@163.com)
This work was supported in part by the National Natural Science Foundation of China under Grant 61401305 and in part by the Natural
Science Foundation of Tianjin, China, under Grant 15JCYBJC16500.

ABSTRACT Analytic envelope is the most prevalent definition of envelopes of real-valued signals. However,
analytic signals are not adopted by some envelope detectors in application. This paper investigates the
envelopes of real-valued signals from a signal processing perspective. We show that the upper and lower
envelopes of signals can be obtained by signal reconstruction after extrema sampling on signals. We prove
that extrema sampling is a sub-Nyquist sampling. We then conclude that the envelopes of real-valued
signals contain two parts, some low-frequency components of the original signal and some new components
generated by sub-Nyquist extrema samplings. Some examples are presented to compare with the analytic
envelope of signals.

INDEX TERMS Envelope of real-valued signals, analytic signal, extrema sampling, signal reconstruction,
digital signal processing.

I. INTRODUCTION sequences in the empirical mode decomposition (EMD) algo-

As an imaginary curve, the envelopes of a signal are the rithm which is an adaptive non-stationary signal processing
boundary within which the signal is contained. Envelopes method [9]. This paper aims to analyze envelopes of real-
contain some information of signals, though it is an valued signals from the signal processing perspective.
imaginary curve, for example, demodulating amplitude mod-
ulated (AM) signals by them. Envelope detection or demod- II. ENVELOPES DEFINITION
ulation method is a dominant approach in detecting faults in In this section, the definition of analytic envelopes is reviewed
rolling element bearings and gearboxes [1], [2]. firstly. Then, a new definition of envelopes, called interpola-
There are various definitions about signal envelopes, such tive envelopes, is proposed.
as the analytic envelope [3]–[5], the pre-envelope [4], and
the nature envelope [5]. Among these definitions, the analytic A. THE ANALYTIC ENVELOPE
envelope based on the Hilbert transform is famous. The definition of analytic envelope is associated with analytic
However, the concept of the envelope used in technical signals. The complex-valued signal m(t) is termed analytic if
problems is not as clear-cut as the geometrical concept of and only if m̂(t) = −jm(t) where m̂(t) is the Hilbert transform
the envelope of a family of curves [5]. Moreover, the Hilbert of m(t) [10]. It would be of more interest to mention that an
transform is generally employed to obtain the analytic enve- analytic signal contains no negative frequencies [6].
lope of real-valued signals, which is not convenient in prac- For a given real band-limited signal s(t), if s(t) ∈ L2 (R)
tice. Bedrosian [6] pointed out that the envelope of an actual and its Fourier transform is S(f ), then its Hilbert transform is
waveform has physical significance only for narrow-band defined by [11]
signals and cannot be measured precisely except for a pure
sinusoid. Z+∞
In engineering, the detection of envelopes of real signals 1 s(τ )
ŝ(t) = H [s(t)] = P dτ (1)
sometimes does not rely upon analytic signals, though it is π t −τ
−∞
uniquely defined based on analytic signals [7]. Most methods
for envelope demodulation depend on peak detection fol- where P is the Cauchy principal value. Hence, the complex-
lowed by a low-pass filter, typically a RC network, to remove valued signal
the carrier and provide smoothing [8]. The envelopes of real
signals are constructed by spline interpolation from extrema m(t) = s(t) + jŝ(t) (2)

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 5, 2017 Personal use is also permitted, but republication/redistribution requires IEEE permission. 5623
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals

should be a local minimum located between two adjacent

local maxima and a local maximum between two adjacent
local minima [20]. Hence, the minima sampling rate is equal
FIGURE 1. Block diagram of the process of forming envelopes. to the maxima sampling rate for a continuous signal. About
the extrema sampling, we have the following theorem.
Theorem 1: Consider x(t) ∈ L2 (R) a T -periodic signal that
is an analytic signal [6]. Then, (2) can be rewritten as [11] satisfies the Dirichlet conditions. Then, the maxima (minima)
rate of x(t) is less than or equal to the highest frequency
m(t) = |m(t)| ejφ(t) (3) of x(t).
where Proof: Let fe denote the maxima (minima) rate of the
signal x(t). With no loss of generality, assume that f1 denotes
ŝ(t)
φ(t) = arctan , (4) the highest frequency of x(t). The periodic signal x(t) can be
s(t) expressed by a finite Fourier series, in the form [21]:
h i1/2
|m(t)| = s2 (t) + ŝ2 (t) . (5) M
X
The analytic signal m(t) is referred to the pre-envelope x(t) = cn ejnt (6)
n=−M
of s(t) [3], [4] or the complex analytic envelope of s(t) [5].
The modulus |m(t)| is termed the envelope of s(t) [3], [4] or where
the analytic envelope of s(t) [5]. 2π
= (7)
T
B. THE INTERPOLATIVE ENVELOPE
1 T /2
Z
In the EMD algorithm, the upper and lower envelopes of cn = x(t)e−jnt dt. (8)
T −T /2
signals are obtained by using two steps: 1) identify all the
local extrema; 2) interpolate between extrema by a cubic We have cn = 0 when n > T /T1 , where T1 = 1/f1 because of
spline [9]. The interpolative envelope includes the upper and the orthogonality of the basis functions of the Fourier series.
lower envelopes. The first step is essentially a process of Then, for the sake of simplicity, let M = T /T1 .
extrema sampling, and the second step is used to recover the If x(t) takes an extremum at t = te , then its first deriva-
signal from the extrema samples [12]–[16]. Then, the upper tive is
and lower envelopes of real-valued signals can be given as M
follows.
X
x 0 (te ) = jncn ejnte = 0. (9)
Def. 1: An upper (or lower) envelope of a real signal is a n=−M
function s : R → R; t → s(t) used for its reconstruction with
its local maxima (minima) sequence. In fact, x 0 (t) can be regarded as a polynomial in ejt of
Based on Def. 1, an upper or lower envelope of a signal degree 2M . Using in (6) the mapping U = ejt , we derive
can be obtained by two steps: extrema sampling and signal M
reconstruction, as shown in Fig. 1. For the method of envelope
X
x 0 (U ) = jncn U n . (10)
demodulation based on RC network, the peak detection is just n=−M
the process of extrema sampling, and the RC network is just
used to recover signals. It is clear that the upper and lower The number of zeros of x 0 (U ) is 2M and equals the trigono-
envelopes can be formed by the same procedure. Without loss metric polynomial’s degree within a period T [22]. Therefore,
of generality, only the upper envelope is analyzed in the rest we have
of this paper. T
According to the Shannon’s sampling theorem for uniform N0 = 2M = 2 (11)
T1
sampling and the folk theorem [17], or alternatively, the
Beurling-Landau theorem [18] for non-uniform sampling, where N0 denotes the number of zeros of x 0 (t) within a
a component can be perfectly recovered when its frequency period T . Let Ne represent the number of extrema of x(t)
is not greater than half of the maxima rate. As far as spline within a period T . It is shown in [21] that not all the zeros
interpolation is concerned, it is reported in [19] that the spline point of x 0 (t) are the extrema point of x(t). Then we have
interpolator of order n approaches an ideal sinc interpolator Ne ≤ N0 . (12)
as n goes to infinity. Therefore, the content of upper envelopes
is determined to a large extend by maxima samplings. For the continuous signal x(t), the numbers of maxima and
minima are equal in a period T . Then, we obtain
III. ANALYSIS ON ENVELOPES
Ne /2 N0 /2 1
A. EXTREMA SAMPLING fe = ≤ = = f1 (13)
Extrema sampling includes the maxima and minima sam- T T T1
plings. From the viewpoint of mathematical analysis, there which brings the desired result.

5624 VOLUME 5, 2017

Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals

B. COMPONENTS OF ENVELOPES where ai , fi , and ϕi are, respectively, the amplitude, frequency

Theorem 1 shows that the maxima and minima sampling are and phase of the component xi (t), 1 ≤ i ≤ 2. Without loss
sub-Nyquist sampling. The highest frequency component of of generality, assume f1 > f2 . Then, we have fe ≤ f1 which
signals can not be reconstructed perfectly by the upper and implies that envelopes will contain some new components, as
lower envelopes. However, the components whose frequency shown in Fig. 2.
is no more than fe /2 can be recovered completely in the
envelopes. As a direct consequence, the upper and lower
envelopes will contain two parts, that are: components of the
original signal as well as news components generated by the
sub-Nyquist extrema sampling.
Considering a multicomponent signal in the form
n
X
x(t) = xi (t), (14)
i=1

we can rewrite it as
x(t) = xh (t) + xl (t)
where xh (t)and xl (t) denote the components whose frequency
is greater or lesser than fe /2, respectively. Let eu (t) and ed (t)
denote the upper and lower envelopes of a signal, respectively.
Based on the above analysis, the upper envelope of the mul-
ticomponent signal x(t) can be expressed as
(15)
FIGURE 2. Illustration of new components in the upper envelopes. (a) The
two tones signal with f1 = 100 Hz, f2 = 90 Hz, a1 = 1, and a2 = 1. (b) The
spectrum of the upper envelope shown in (a). (c) The two tones signal
with f1 = 100 Hz, f2 = 60 Hz, a1 = 1, and a2 = 1. (d) The spectrum of the
upper envelope shown in (c).

eu (t) = xl (t) + nu (t) (16) For the upper envelope, we first consider the case that
fe = f1 which has a sufficient condition [23]
where nd (t) represents the new components in the upper
envelope. Correspondingly, the lower envelope of x(t) can be a2 f2 < a1 f1 (21)
expressed as
If f2 ≤ f1 /2, the signal x2 (t) can be reconstructed perfectly.
ed (t) = xl (t) + nd (t) (17) For the sake of simplifying the analysis, assume that extrema
are equally spaced. Using (18) and the property of the sinc
where nd (t) is the new components in the lower envelope. function, we then have
Especially, for the case of extrema spaced uniformly, the ∞
upper and lower envelope of signals can be written as [17]
X t − tu
eu (t) = [x1 (kTe + tu ) + x2 (kTe + tu )] sin c( −k)
∞
Te
k=−∞
t − tu
= a1 a0 + a2 cos(2π f2 t + ϕ2 )
X
eu (t) = x(kTe + tu )sinc( − k) (18)
Te
k=−∞ = a1 a0 + x2 (t) (22)
∞
X t − td
ed (t) = x(kTe + td )sinc( − k) (19) where a0 = cos(2π f1 tu +ϕ1 ). In this case, the upper envelope
Te eu (t) only contains the low frequency signal x2 (t) and a
k=−∞
constant a0 . A simple illustration is shown in Fig. 3. However,
where sinc(t) = sin(π t)/(π t), Te = 1/fe , tu (td ) denotes the
if f1 /2 < f2 < f1 , we have
first maximum (minimum) point, respectively. Some exam-
ples and further discussion are shown in the next section. eu (t) = a1 a0 + xa2 (t) (23)
where xa2 (t) denotes the aliasing component of x2 (t) [17].
IV. ILLUSTRATIVE EXAMPLES
We then consider the case that fe < f1 . If f2 ≤ fe /2, the
In this section, the interpolative envelopes are analyzed
signal x2 (t) can be reconstructed perfectly. Using (18), we
for two tones signals, double-sideband suppressed
then have
carrier (DSB-SC) signals, conventional AM signals, and
non-stationary signals. The interpolative envelopes are also eu (t) = x2 (t) + xa1 (t) (24)
compared with the analytic envelope in these signals.
where xa1 (t) denotes the aliasing component of x1 (t). In this
case, the upper envelope eu (t) contains the low frequency
A. ENVELOPES OF TWO TONES SIGNALS
signal x2 (t) but with a new component xa1 (t). If f2 > fe /2,
A composite two tones signal model can be given by the two components x1 (t) and x2 (t) cannot be reconstructed
x(t) = x1 (t) + x2 (t) perfectly. Then we have
= a1 cos(2π f1 t + ϕ1 ) + a2 cos(2πf2 t + ϕ2 ) (20) eu (t) = xa1 (t) + xa2 (t) (25)

VOLUME 5, 2017 5625

 Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals

It is quite clear from Fig. 4 that the analytic envelope contains

a new component rather than the two components of x(t).
For comparison’s sake, the upper envelope contains the low
frequency component x2 (t).

B. ENVELOPES OF DSB-SC SIGNALS

A DSB-SC signal is expressed mathematically as [24]

s(t) = Ac x(t) cos(2πfc t)

where Ac denotes the carrier amplitude and fc represents the

carrier frequency. The signal x(t) is usually known as the
message signal. Here, we consider x(t) = A cos(2π ft). Then
a simple DSB-SC signal model can be expressed as

s(t) = Ac A cos(2π ft) cos(2π fc t)

FIGURE 3. Illustration of the upper envelopes without new components.
(a) The two tones signal with f1 = 100 Hz, f2 = 20 Hz, a1 = 1, and a2 = 1.
Ac A
= cos[2π (fc − f )t]
(b) The spectrum of the signal shown in (a). 2
Ac A
+ cos[2π (fc + f )t] (27)
2
where fc >> f .
In fact, the signal shown in (27) is a two tones signal.
It is proved in [23] that fe = f1 if a1 f1 > a2 f2 for the two
tones signal expressed by (20). Hence, we have fe = fc + f
for the signal shown in (27). If assuming that extrema space
uniformly, by (18), the upper envelope of the DSB-SC signal
shown in (27) can be expressed as
∞
Ac A X
eu (t) = {A0 + cos[4π f (kTe − t0 )]}
2
k=−∞
t − tu
×sinc( − k) (28)
Te
where A0 = cos[2π (f + fc )tu ] and t0 = (fc − f )tu /(2f ). Then,
we can obtain
Ac A
eu (t) = [A0 + cos(4π ft + ϕ)] (29)
2
FIGURE 4. Example of envelopes of a two tones signal. (a) The original where ϕ is the phase of the upper envelope. The upper
signal and the analytic envelope. (b) The original signal and the upper envelope of s(t) contains the component whose frequency is
envelope. (c) The spectrum of the analytic envelope. (d) The spectrum of
the upper envelope.
double that of the message signal x(t).
Nevertheless, the analytic envelope of s(t) is
h i1/2
Equation (25) shows that the upper envelope eu (t) only con- |m(t)| = s2 (t) + ŝ2 (t) = Ac A |cos(2π ft)| (30)
tains some new components.
The analytic envelope of x(t) shown in (20) is which also does not contain the message signal x(t). In addi-
h i1/2 tion, it has some discontinuous points at ωt = nπ + π/2
|m(t)| = x 2 (t) + x̂ 2 (t) where ω = 2π f and n = 0, 1, 2, · · · . Hence, the envelope
detector cannot be used to demodulate DSB-SC signals.
= {a21 + a22 + 2a1 a2 cos[2π (f1 − f2 )t]}1/2 (26)
An example is presented in Fig. 5 where fc = 120 and
where m(t) is the analytic signal. It is obvious from (26) f = 40 to compare the two types envelopes for DSB-SC
that the analytic envelope does not contain the signals signals. It is clear from this Figure that the main component
x1 (t) and x2 (t). of the two types of envelopes is the component of 80 Hz
A simple example is shown in Fig. 4 to compare the inter- instead of 40 Hz. However, the analytic envelope has some
polative envelope with the analytic envelope and to testify the high frequency components. Although they cannot be used
above theoretical analysis. The two tones of the signal used by to detect the message signal x(t), the upper envelope is more
Fig. 4 are x1 (t) = cos(2π × 100t) and x2 (t) = cos(2π × 20t). reasonable than the analytic envelope for the DSB-SC signal.

5626 VOLUME 5, 2017

Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals

FIGURE 6. Illustration of the upper envelope of a conventional AM signal.

(a) The original signal and the upper envelope. (b) The spectrum of the
FIGURE 5. Example of envelopes of a DSB-SC signal. The meaning of each
conventional AM signal. (c) The spectrum of the upper envelope.
sub-panel is the same as in Fig. 4.

where Ŝ(f ) denotes the spectrum of ŝ(t). By taking the inverse

C. ENVELOPES OF CONVENTIONAL AM SIGNALS
Fourier transform of Ŝ(f ), we have
Considering a conventional AM signal in which a carrier
signal of frequency fc is modulated by the message signal x(t), ŝ(t) = Ac [1 + x(t)] sin(2π fc t) (34)
the amplitude modulated signal is [24]
Hence, the analytic envelope of s(t) can be written as
s(t) = Ac [1 + x(t)] cos(2π fc t) h i1/2
|m(t)| = s2 (t) + ŝ2 (t)
where |x(t)| ≤ 1. Suppose that the bandwidth of x(t) is w1 , = Ac |1 + x(t)| . (35)
where we are using fc >> w1 and fe = fc for the signal s(t).
For simplicity, assume that extrema are equally spaced. As a consequence, the analytic envelope contains the message
By using (18), the upper envelope of s(t) can be expressed as signal x(t). Hence, it can be used to demodulate conventional
∞
AM signals.
X t − tu
eu (t) = Ac A0 [1 + x(kTe + tu )] sin c( − k) (31)
Te
k=−∞

where A0 = cos(2π fc tu ) is a constant. Then, we can get

eu (t) = Ac A0 [1 + x(t)] (32)

Equation (32) shows that the upper envelope of the conven-

tional AM signal s(t) contains the messages signals x(t) and
a constant. Therefore, the envelope detector can be used to
demodulate.
A simple example is demonstrated in Fig. 6. The signal is
xc (t) = cos(2π × 100t) and x(t) = cos(2π × 20t). From this
figure, it is obvious that the main component of the upper
envelope is the component of 20 Hz, and we can be sure that
it is just the message signal.
The spectrum S(f ) is easy to obtain by taking the Fourier
transform of s(t). Then, we get
FIGURE 7. Example of envelopes of a AM signal. The meaning of each
Ŝ(f ) = S(f ) · [−jsgn(f )] sub-panel is the same as in Fig. 4.

Ac
= j [X (f + fc ) − X (f − fc )] The example to compare the interpolative envelope with
2
Ac the analytic envelope for the conventional AM signal is shown
+ j [δ(f + fc ) − δ(f − fc )] (33) in Fig. 7. The signal is the same as used in Fig. 6. From this
2
VOLUME 5, 2017 5627

figure, it is evident that the main component of the two types
of envelopes is the component of 20 Hz. Thus, we can be sure
that the two types of envelopes are almost the same, though
the upper envelope has a little high-frequency component.
D. ENVELOPES OF NON-STATIONARY SIGNALS
Non-stationary signals exist widely in practice, whose fre-
quency contents vary with time. Typically, most vibration
signals in a fault rotate machine are non-stationary. Here, we
first consider a non-stationary signal which is written as
x(t) = 10 cos 2π (80t + 500t 2 ) + 10 cos 2π (460t + 310t 2 )
= x1 (t) + x2 (t)
where x1 (t) and x2 (t) are real chirp signals. The signal and
the short-time Fourier transform (STFT) are shown in Fig. 8.
FIGURE 8. The non-stationary signal. (a) The signal. (b) The STFT.
FIGURE 9. Envelopes of the non-stationary signal shown in Fig. 8. (a) The
original signal and the analytic envelope. (b) The original signal and the
upper envelope. (c) The STFT of the analytic envelope. (d) The STFT of the
upper envelope.
The comparison result on the two type envelopes is shown
in Fig. 9. It is clear that the interpolative envelope only
5628
Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals
contains the lower chirp signal component x1 (t). However, the
analytic envelop does not contain one of the two chirp signals.
Thus, the interpolative envelope is also fit for demodulating
non-stationary signals.
We then consider a simulation faulty bearing vibration
signal. According to [25] and [26], a vibration signal arising
from a rolling element bearing can be simulated as
 q  
−ξ ωn t
x(t) = a0 e sin 2π fn 1 − ξ t + n(t) (36)
2
where ξ denotes the ratio of damping, fn represents the natural
frequency of bearings, and n(t) is the noise. In [1], the bearing
fault signals are added by some harmonic interferences which
is written as
3
X
h(t) = Ai sin 2π fi t
i=1
For a0 = 10, ξ = 0.08, and fn = 3000, a series bearing
fault impulses repeated period of 0.008s with some harmonic
interferences is shown in Fig. 10. The fault characteristic fre-
quency modulated by high frequency components is 125 Hz.
It is not easy from this Figure to find the fault characteristic
frequency component through the spectrum. The comparison
result on the two type envelopes is shown in Fig. 11, where
the values of some frequency points of envelopes are listed
in Table 1.
FIGURE 10. The simulated bearing fault signal. (a) A series of bearing
fault impulses. (b) The spectrum of (a). (c) The simulated fault signal
with some noise and harmonic interferences. (d) The spectrum
of (c).
In Fig. 11, we can find the fault characteristic frequency
and its harmonics from both the analytic and interpolative
envelopes. Nevertheless, the interpolative envelope contains
VOLUME 5, 2017
Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals
FIGURE 11. Envelopes of the signal shown in Fig. 10. (a) The signal and
its analytic envelope. (b) The signal and its interpolative envelope. (c) The
spectrum of (a). (d) The spectrum of (b).
TABLE 1. The value of some points of Envelopes shown in Fig. 10.
the harmonic interferences, but the analytic envelop only con-
tains part of the harmonic interferences whose frequencies are
100, 200, and 1000 Hz, respectively. Table 1 and Fig. 10 show
that the analytic envelop has some new components which are
marked out with box in this table.
In order to clearly extract the fault characteristic fre-
quency embedded within the time waveform, a high-pass
filter is used to remove the low frequency component of
the signal. Then, the filtered signal is processed by the
analytic and interpolative envelopes. The result is shown
in Fig. 12.
Comparing this Figure with Fig. 11, it is obvious that the
fault characteristic frequency and its harmonics become very
visible for both the two type of envelopes. Furthermore, this
Figure shows that the interpolative envelope has the same
performance with the analytic envelope for this filtered sig-
nal. Hence, the interpolative envelope can be used to fault
diagnosis of bearing.
VOLUME 5, 2017
FIGURE 12. Envelopes of the signal shown in Fig. 10 filtered by a
high-pass filter. The meaning of each sub-panel is the same as in Fig. 11.
V. CONCLUSIONS
We investigated the envelopes of real-valued signals from the
viewpoint of signal processing. We have shown that the upper
and lower envelopes of a real-valued signal can be described
as the signals reconstructed from the maxima and minima
samples, respectively. We illustrated that the upper and lower
envelopes of signals can be obtained by two operations,
extrema sampling and signal reconstruction.
We have proven that the extrema sampling is a sub-Nyquist
sampling. Generally speaking, we concluded that envelopes
of real-valued signal contain some low-frequency compo-
nents with some new components except certain special
cases. A typical case is that the envelope can be used to
demodulate conventional AM signals.
The proposed interpolative envelope, unlike the analytic
envelope, is not limited to narrow-band signals. The simu-
lated results show that the interpolative envelope has a good
performance both for stationary and non-stationary signals.
Although the interpolative envelope has already many appli-
cations in engineering, further studies are needed to investi-
gate envelope analysis through interpolation such as bearing
fault detection.
ACKNOWLEDGMENT
The authors would like to thank Prof. Ran Tao and
Jiahao Deng for helpful discussions. The authors also
acknowledge the anonymous reviewers for their valuable
comments, which helped us to improve the manuscript.
REFERENCES
[1] J. S. Luo, D. J. Yu, and M. Liang, ‘‘A kurtosis-guided adaptive demodu-
lation technique for bearing fault detection based on tunable-Q wavelet
transform,’’ Meas. Sci. Technol., vol. 24, no. 5, pp. 1–11, Apr. 2013,
doi: 10.1088/0957-0233/24/5/055009.
5629

[2] N. G. Nikolaou and I. A. Antoniadis, ‘‘Demodulation of vibration signals
generated by defects in rolling element bearings using complex shifted
morlet wavelets,’’ Mech. Syst. Signal Process., vol. 16, no. 4, pp. 677–694,
Jul. 2002, doi: 10.1006/mssp.2001.1459.
[3] J. Dugundji, ‘‘Envelopes and pre-envelopes of real waveforms,’’ IRE Trans.
Inf. Theory, vol. 4, no. 1, pp. 53–57, Mar. 1958, doi: 10.1109/TIT.1958.
1057435.
[4] R. Arens, ‘‘Complex processes for envelopes of normal noise,’’ IRE Trans.
Inf. Theory, vol. 3, no. 3, pp. 204–207, Sep. 1957, doi: 10.1109/TIT.1957.
1057417.
[5] S. O. Rice, ‘‘Envelopes of narrow-band signals,’’ Proc. IEEE, vol. 70, no. 7,
pp. 692–699, Jul. 1982, doi: 10.1109/PROC.1982.12376.
[6] E. Bedrosian, ‘‘The analytic signal representation of modulated wave-
forms,’’ Proc. IRE, vol. 50, no. 10, pp. 2071–2076, Oct. 1962.
[7] R. Kumaresan, ‘‘An inverse signal approach to computing the envelope of a
real valued signal,’’ IEEE Signal Process. Lett., vol. 5, no. 10, pp. 256–259,
Oct. 1998, doi: 10.1109/97.720558.
[8] D. A. Rice, V. Venkatachalam, and M. J. Wegmann, ‘‘A simple enve-
lope detector,’’ IEEE Trans. Instrum. Meas., vol. 37, no. 2, pp. 223–226,
Jun. 1988, doi: 10.1109/19.6056.
[9] N. E. Huang et al., ‘‘The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis,’’ Proc. Roy.
Soc. London A, Math. Phys. Sci., vol. 454, pp. 903–995, Mar. 1998.
[10] J. Brown, ‘‘Analytic signals and product theorems for Hilbert transforms,’’
IEEE Trans. Circuits Syst., vol. 21, no. 6, pp. 790–792, Nov. 1974,
doi: 10.1109/TCS.1974.1083928.
[11] H. B. Voelcker, ‘‘Toward a unified theory of modulation part I: Phase-
envelope relationships,’’ Proc. IEEE, vol. 54, no. 3, pp. 340–353,
Mar. 1966, doi: 10.1109/PROC.1966.4695.
[12] Y. L. Yang et al., ‘‘Nonuniform extrema resampling and empirical mode
decomposition,’’ Chin. J. Electron., vol. 18, no. 4, pp. 759–762, Oct. 2009.
[13] Y. L. Yang, ‘‘Theoretical analysis and application investigation of empiri-
cal mode decomposition,’’ Ph.D. dissertation, School Mech. Eng., Beijing
Inst. Technol., Beijing, China, 2010.
[14] Y. L. Yang, C. Miao, and J. Deng, ‘‘An analytical expression for
empirical mode decomposition based on B-spline interpolation,’’ Cir-
cuits, Syst., Signal Process., vol. 32, no. 6, pp. 2899–2914, Dec. 2013,
doi:10.1007/s00034-013-9592-5.
[15] Y. L. Yang, J. Deng, and D. Kang, ‘‘An improved empirical mode decom-
position by using dyadic masking signals,’’ Signal, Image Video Process.,
vol. 9, no. 6, pp. 1259–1263, Sep. 2015, doi:10.1007/s11760-013-0566-7.
[16] Y. L. Yang, ‘‘Empirical mode decomposition as a time-varying multi-
rate signal processing system,’’ Mech. Syst. Signal Process., vols. 76–77,
pp. 759–770, Aug. 2016, doi: 10.1016/j.ymssp.2016.02.022.
[17] A. J. Jerri, ‘‘The Shannon sampling theorem—Its various extensions
and applications: A tutorial review,’’ Proc. IEEE, vol. 65, no. 11,
pp. 1565–1596, Nov. 1977, doi: 10.1109/PROC.1977.10771.
5630
View publication stats
Y. Yang: Signal Theoretic Approach for Envelope Analysis of Real-Valued Signals
[18] M. Unser, ‘‘Sampling-50 years after Shannon,’’ Proc. IEEE, vol. 88, no. 4,
pp. 569–587, Apr. 2000, doi: 10.1109/5.843002.
[19] M. Unser, A. Aldroubi, and M. Eden, ‘‘Polynomial spline signal approx-
imations: Filter design and asymptotic equivalence with Shannon’s sam-
pling theorem,’’ IEEE Trans. Inf. Theory, vol. 38, no. 1, pp. 95–103,
Jan. 1992, doi: 10.1109/18.108253.
[20] J. Li and R. M. Gray, ‘‘Context-based multiscale classification of docu-
ment images using wavelet coefficient distributions,’’ IEEE Trans. Image
Process., vol. 9, no. 9, pp. 1604–1616, Sep. 2000, doi: 10.1109/83.862641.
[21] Z. Xu, B. Huang, and S. Xu, ‘‘Exact location of extrema for empirical mode
decomposition,’’ Electron. Lett., vol. 44, no. 8, pp. 551–552, Apr. 2008,
doi: 10.1049/el:20083405.
[22] A. A. G. Requicha, ‘‘The zeros of entire functions: Theory and engineer-
ing applications,’’ Proc. IEEE, vol. 68, no. 3, pp. 308–328, Mar. 1980,
doi: 10.1109/PROC.1980.11644.
[23] G. Rilling and P. Flandrin, ‘‘One or two frequencies? the empirical mode
decomposition answers,’’ IEEE Trans. Signal Process., vol. 56, no. 1,
pp. 85–95, Jan. 2008, doi: 10.1109/TSP.2007.906771.
[24] J. G. Proakis and M. Salehi, Communication Systems Engineering, 2nd ed.
Beijing, China: Publishing House of Electronics Industry, 2004, pp. 70–81,
[25] D. Ho and R. B. Randall, ‘‘Optimisation of bearing diagnostic techniques
using simulated and actual bearing fault signals,’’ Mech. Syst. Signal Pro-
cess., vol. 14, no. 5, pp. 763–788, Sep. 2000, doi: 10.1006/mssp.2000.1304.
[26] R. B. Randall, J. Antoni, and S. Chobsaard, ‘‘The relationship between
spectral correlation and envelope analysis in the diagnostics of bearing
faults and other cyclostationary machine signals,’’ Mech. Syst. Signal
Process., vol. 15, no. 5, pp. 945–962, 2001, doi: 10.1006/mssp.2001.1415.
YANLI YANG received the master’s degree in
mechanical manufacture and automation from
Dalian Jiaotong University in 2006, and the Ph.D.
degree in mechatronic engineering from the Bei-
jing Institute of Technology in 2010. He is cur-
rently an Associate Professor with the School of
Electronics and Information Engineering, Tianjin
Polytechnic University. His current research inter-
ests include signal processing and equipment state
monitoring. He was a recipient of the second prize
from the Chinese Machinery Industry Science and Technology in 2009.
VOLUME 5, 2017