Professional Documents
Culture Documents
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
a r t i c l e i n f o abstract
Article history: The modulated oscillation model provides physically meaningful representations of time-
Received 27 December 2013 varying harmonic processes, and has been instrumental in the development of modern
Received in revised form time-frequency algorithms, such as the synchrosqueezing transform. We here extend this
10 July 2014
concept to multivariate signals, in order to identify oscillations common to multiple data
Accepted 5 August 2014
channels. This is achieved by introducing a multivariate extension of the synchrosqueez-
Available online 13 August 2014
ing transform, and using the concept of joint instantaneous frequency multivariate data.
Keywords: For rigor, an error bound which assesses the accuracy of the multivariate instantaneous
Amplitude and frequency modulated signal frequency estimate is also provided. Simulations on both synthetic and real world data
Synchrosqueezing transform
illustrate the advantages of the proposed algorithm.
Analytic signal
& 2014 Elsevier B.V. All rights reserved.
Multivariate modulated oscillations
Multivariate signal
Time-frequency analysis
Instantaneous frequency
http://dx.doi.org/10.1016/j.sigpro.2014.08.010
0165-1684/& 2014 Elsevier B.V. All rights reserved.
332 A. Ahrabian et al. / Signal Processing 106 (2015) 331–341
The synchrosqueezing transform (SST) [16,17] belongs to this where an(t) and ϕn ðtÞ represent the instantaneous ampli-
class, it is a post-processing technique based on the contin- tude and phase for each channel n ¼ 1; …; N. The work in
uous wavelet transform that generates highly localized time- [9] proposed the joint instantaneous frequency (power
frequency representations of nonlinear and nonstationary weighted average of the instantaneous frequencies of all
signals. Synchrosqueezing provides a solution that mitigates the channels) of multivariate data in the form:
the limitations of linear projection based time-frequency
algorithms, such as the short-time Fourier transform (STFT) d
I xHþ ðt Þ x þ ðt Þ 0
n ¼ 1 an ðtÞϕn ðtÞ
∑N 2
dt
and continuous wavelet transforms (CWT). The synchros- ωx ðt Þ ¼ ¼ ð4Þ
queezing transform reassigns the energies of these trans- J x þ ðtÞ J 2 ∑Nn ¼ 1 an ðtÞ
2
joint analytic spectrum around the scale factor as ¼ ωψ =ωs , where ωψ is the center
1 frequency of a wavelet, while the energy of the original
SðωÞ ¼ J X þ ðωÞ J 2 ; ð7Þ frequency ωs is spread across as. Thus, the estimated
E
frequency present in those scales is equal to the original
where X þ ðωÞ is the Fourier transform of x þ ðtÞ and E is the frequency ωs. Consequently, the instantaneous frequency
total energy of the Fourier coefficients given by ωx ða; bÞ can be estimated as
Z 1
1
E¼ J X þ ðωÞ J 2 dω; ð8Þ ∂Wða; bÞ
2π 0 ωx ða; bÞ ¼ ı_Wða; bÞ 1 ð15Þ
∂b
this makes possible to express the joint global mean for each scale-time pair (a,b). The resulting wavelet coeffi-
frequency expressed as the first moment of the joint cients that contain the same instantaneous frequencies
analytic spectrum as can then be combined using a procedure referred to as
Z 1
1 synchrosqueezing (SST). For a set of wavelet coefficients
ω¼ ωSðωÞ dω: ð9Þ
2π 0 Wða; bÞ, the synchrosqueezing transform1 Tðωl ; bÞ is given
by
The joint global second central moment (multivariate
bandwidth squared) measures the spectral deviation of Tðωl ; bÞ ¼ ∑ Wðak ; bÞa 3=2 Δak ð16Þ
the joint analytic spectrum from the joint global mean ak : ∣ωx ðak ;bÞ ωl ∣ r Δω=2
frequency, and is given by where ωl are the frequency bins with a resolution2 of Δω.
Z 1
1 The SST has been shown to reconstruct univariate modu-
σ2 ¼ ðω ω Þ2 SðωÞ dω: ð10Þ
2π 0 lated oscillations of the form in (1), as follows [16]
" #
Accordingly, the global moments of the analytic spectrum
xðbÞ ¼ R Rψ 1 ∑Tðωl ; bÞΔω ð17Þ
can be related to the moments of joint instantaneous l
frequency and bandwidth as R 1 n
Z where Rψ ¼ 12 0 ψ^ ξ dξξ, is the normalization constant
1 1
ω¼ J x þ ðt Þ J 2 ωx ðt Þ dt ð11Þ and ψ^ ðξÞ is the Fourier transform of the mother wavelet
E 1
ψ ðtÞ.
Z 1
1
σ2 ¼ J x þ ðt Þ J 2 σ 2x ðt Þ dt; ð12Þ
4. Multivariate time-frequency representation using the
E 1
SST
where σ x ðtÞ
2
is the joint instantaneous second central
moment, given by
In order to extend the SST to the analysis of multi-
σ 2x ðtÞ ¼ υ2x ðtÞ þ ðωx ðtÞ ω Þ2 : ð13Þ variate signals, recall that if the modulated oscillatory
components are known for each channel as in (3), then
we can determine the joint instantaneous frequency,
Remark 3. Observe that the multivariate bandwidth provided that the frequencies of the modulated oscilla-
squared, σ 2 , depends on both the joint instantaneous tions are sufficiently close together. With that insight, we
bandwidth, σ 2x ðtÞ, and the deviations of the joint instanta- propose to first partition the time-frequency domain into
neous frequency from the joint global mean frequency, ω . K frequency bands fωk gk ¼ 1;…;K . This makes it possible to
identify, a set of matched monocomponent signals from a
3. Synchrosqueezing transform given multivariate signal. The instantaneous amplitudes
and frequencies present within those frequency bands can
The synchrosqueezing transform is a post-processing then be determined (yielding amplitude and frequency
technique applied to the continuous wavelet transform in modulated oscillations, similar to the intrinsic-mode func-
order to generate localized time-frequency representa- tions (IMFs) of the EMD algorithm [18]).
tions of non-stationary signals. The continuous wavelet
transform is a projection based algorithm that identifies 4.1. Partitioning of the time-frequency domain
oscillatory components of interest through a series of
time-frequency filters known as wavelets. A wavelet ψ ðtÞ We propose to partition the time-frequency domain via
is a finite oscillatory function, which when convolved with a multivariate extension of an adaptive frequency tiling
a signal x(t), in the form technique first proposed in [19]; the underlying concept is
Z to determine multivariate monocomponent signals based
t b
W ða; bÞ ¼ a 1=2 ψ xðt Þ dt ð14Þ on the multivariate bandwidth. The time-frequency plane
a
is first partitioned into 2l equal-width frequency bands, for
gives the wavelet coefficients Wða; bÞ, for each scale-time the frequency range, ωl;m ¼ ½m=2ðl þ 1Þ ; ðm þ 1Þ=2ðl þ 1Þ , where
pair (a,b). In this way, the wavelet coefficients in (14) and
can be seen as the outputs of a set of scaled bandpass
1
filters. The scale factor a shifts the bandpass filters in the A detailed implementation of the synchrosqueezing transform can
be found in [22].
frequency domain, and also changes the bandwidth of 2
We have used linear frequency scales therefore Δω is constant,
the bandpass filters. Therefore, the energy of the wavelet however for logarithmic frequency scales, Δω would vary with frequency
transform of a sinusoid at a frequency ωs will spread out [22].
334 A. Ahrabian et al. / Signal Processing 106 (2015) 331–341
3
B3.0 B3.1 B3.2 B3.3 B B B B
3.4 3.5 3.6 3.7 Remark 4. For modulated oscillations separated in fre-
0 1/16 2/16 3/16 4/16 5/16 6/16 7/16 8/16 quency, the proposed partitioning method provides a
Normalised Frequency robust method for separating monocomponent signals.
However, for closely spaced monocomponent functions
Fig. 1. The partitioned frequency domain with the multivariate band-
that are separated in both time and frequency (i.e. two
width given by Bl;m , where l corresponds to the level of the frequency
band and m is the frequency band index. parallel chirp signals), the method cannot resolve the
separate monocomponent signals.
transform of the inverse of the SST coefficients that is, at each point in time the instantaneous frequencies
" ( )# are well separated. The second step is to estimate the
Φl;m ðωÞ ¼ F Rψ 1 ∑ T n ðω; bÞ ð18Þ multivariate instantaneous frequency by combining, for a
ω A ωl;m given frequency band k, the instantaneous frequencies across
n ¼ 1;…;N
the N channels, using the joint instantaneous frequency in
where F fg is the Fourier transform operator, Rψ the
(4). As a result, the multivariate instantaneous frequency
normalization constant [16] and Φl;m ðωÞ A RN a column
band Ωk ðbÞ is given by3
multi
vector. The multivariate bandwidth is then determined via
n ¼ 1 ðAk ðbÞÞ Ωk ðbÞ
Eqs. (7)–(10), as outlined in Section 2. n 2 n
∑N
The rationale behind the adaptive frequency scales is Ωmulti
k ðbÞ ¼ n 2
ð22Þ
∑N
n ¼ 1 ðAk ðbÞÞ
then as follows: if the initial multivariate bandwidth is
calculated for the entire signal at level l ¼0, then the while the instantaneous amplitude Amulti
k ðbÞ for each
bandwidth is split based on the following condition: frequency band becomes
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Bl þ 1;2m Λl þ 1;2m þ 1 þ Bl þ 1;2m þ 1 Λl þ 1;2m þ 1
Bl;m 4 ð19Þ N
Λl þ 1;2m þ 1 þ Λl þ 1;2m þ 1 Amulti
k ðbÞ ¼ ∑ ðAnk ðbÞÞ2 : ð23Þ
n¼1
where
Now that we have determined the joint instantaneous
T
Λl þ 1;2m ¼ ∑ ðAmulti 2 amplitude and frequency for each frequency band, it is
l þ 1;2m ðbÞÞ
b¼1 possible to generate the multivariate time-frequency
T
Λl þ 1;2m þ 1 ¼ ∑ ðAmulti
l þ 1;2m þ 1 ðbÞÞ
2 3
A bound on the error of the estimated multivariate instantaneous
b¼1 frequency is provided in the Appendix.
A. Ahrabian et al. / Signal Processing 106 (2015) 331–341 335
coefficients Tmulti
k ðω; bÞ for each oscillatory scale k, as4 of the desired signal. We first considered a bivariate
sinusoidal oscillation in varying levels of white Gaussian
Tmulti
k ð ω; bÞ ¼ Amulti
k δðω Ω
ðbÞ
multi
k ðbÞÞ ð24Þ
noise
where δðÞ is the Dirac delta function and the multivariate " # " #
time-frequency coefficients for each oscillatory scale are cos ð2π ftÞ n1 ðtÞ
ys ðtÞ ¼ þ
given by Tmulti ðω; bÞ ¼ Tmulti
k ðω; bÞjk ¼ 1;…;K . However, it should cos ð2π ðf þ δÞtÞ n2 ðtÞ
be noted that phase information has been lost through
calculating the instantaneous frequency, and so the original
multivariate signal xðtÞ cannot be reconstructed. A summary 1
10
of the proposed method is shown in Algorithm 1. Proposed Method
MPWD
0
Algorithm 1. Multivariate extension of the SST. 10
5. Simulations
1
10
5.1. Sinusoidal oscillation in noise Proposed Method
MPWD
0
The first set of simulations provides a quantitative 10
Localization Ratio (B)
where f ¼ ½10; 40; 100 Hz are the set of frequencies pre- proposed multivariate time-frequency algorithm had a
sent, and n1 ðtÞ and n2 ðtÞ are independent white Gaussian high localization ratio, as compared to the MPWD, parti-
noise realizations and δ ¼ 1 Hz corresponds to a frequency cularly when the SNR of the input signal is relatively high.
deviation between the channels. The resulting joint The localization ratio of the MPWD remained largely
instantaneous frequency between the channels is given unchanged for different sinusoids, while the localization
by f j ðtÞ ¼ ½10:5; 40:5; 100:5 Hz. The values of the localiza- ratio for the proposed method decreases as the frequency
tion ratio B are shown in Fig. 2. It can be seen that the of the sinusoid increases.
30 30 900
2 800
25 25
700
20 1.5 20 600
Frequency (Hz)
Frequency (Hz)
500
15 15
1 400
10 10 300
0.5 200
5 5
100
0 0 0 0
200 400 600 800 1000 1200 1400 1600 1800 500 1000 1500 2000 2500
Samples Samples
30 30
900
25 25 800
2
700
20 20
Frequency (Hz)
Frequency (Hz)
1.5 600
500
15 15
1 400
10 10 300
0.5 200
5 5
100
0 0 0 0
200 400 600 800 1000 1200 1400 1600 1800 500 1000 1500 2000 2500
Samples Samples
30 30
900
25 2 25 800
700
20 20
Frequency (Hz)
Frequency (Hz)
1.5
600
15 15 500
1 400
10 10
300
0.5 200
5 5
100
0 0 0 0
200 400 600 800 1000 1200 1400 1600 1800 500 1000 1500 2000 2500
Samples Samples
Fig. 3. The time-frequency representations for both the proposed method (left panels) and the MWPD (right panels) for a bivariate AM/FM signal, with
input SNR of (a) 10 dB, (b) 5 dB and (c) 0 dB. The window length used for MPWD was 681 samples.
A. Ahrabian et al. / Signal Processing 106 (2015) 331–341 337
5.2. Amplitude and frequency modulated signal analysis Therefore, the components s1 ðtÞ and s2 ðtÞ were AM signals,
with an amplitude modulation index of 0.5. A frequency
We next considered a multicomponent bivariate AM/ deviation, δ ¼ 0:3 Hz, was introduced to the carrier of s2 ðtÞ
FM signal yðtÞ corrupted by noise (this is analogous to a frequency bias that may arise
" # " # between sensors during data acquisition). The information
s1 ðtÞ þ s2 ðtÞ n1 ðtÞ
yðtÞ ¼ þ bearing components s3 ðtÞ and s4 ðtÞ of the bivariate signal
s3 ðtÞ þ s4 ðtÞ n2 ðtÞ yðtÞ were sinusoidally modulated FM signals, while s4 ðtÞ
also had a frequency deviation of δ ¼ 0:3 Hz.
where n1 ðtÞ and n2 ðtÞ are independent white Gaussian
Fig. 3 shows the time-frequency representations using
noise realizations and the signal components s1 ðtÞ; s2 ðtÞ;
both the proposed method and the MPWD, in processing
s3 ðtÞ; s4 ðtÞ are given by
the bivariate AM/FM multicomponent signal yðtÞ, over a
s1 ðtÞ ¼ ð1 þ0:5 cos ð2π tÞÞ cos ð2π t20Þ range of input SNRs. Observe from Fig. 3(a) that for an
s2 ðtÞ ¼ ð1 þ0:5 cos ð2π tÞÞ cos ð2π tð20 þ δÞÞ input SNR of 10 dB, that the proposed method localizes
s3 ðtÞ ¼ cos ð2π ð10t þ 3:5 cos ðtÞÞÞ the energy of the oscillations along the instantaneous
frequency frequencies that correspond to the components
s4 ðtÞ ¼ cos ð2π ðð10 þ δÞt þ 3:5 cos ðtÞÞÞ:
of yðtÞ. However as the noise power increased, the perfor-
mance of the proposed method degraded as the joint
Table 1
Localization ratios, B, for both the proposed algorithm and the MPWD.
instantaneous frequency estimator is sensitive to noise.
On the other hand, the MPWD is less localized at higher
SNR (dB) Proposed method MPWD SNRs, while the performances of both methods converge
for lower SNRs. Table 1 shows the localization ratio B for
10 0.208 0.037
both techniques, illustrating that as the SNR decreases the
5 0.105 0.025
0 0.04 0.014 localization ratio B for the proposed algorithm converges
with that of the MPWD, implying that while localized the
Lattitude
0.06 Longitude
0.04
0.02
Velocity
0
−0.02
−0.04
−0.06
−0.08
0 200 400 600 800 1000
Samples
0.15 0.15
0.25
Normalized Frequency
0.03
Normalized Frequency
0.025 0.2
0.1 0.1
0.02
0.15
0.015
0.1
0.05 0.01 0.05
0.005
0.05
0 0 0 0
200 400 600 800 1000 200 400 600 800 1000
Samples
Samples
Fig. 4. Time-frequency analysis of real world float drift data. (a) The time domain waveforms of bivariate float velocity data. (b) The time-frequency
representation of float data using the proposed multivariate extension of the SST (left panel) and the MPWD algorithm (right panel). A window length 501
was used for the MPWD.
338 A. Ahrabian et al. / Signal Processing 106 (2015) 331–341
0.2
High Gain
0.15 Low Gain
0.1
0.05
−0.05
−0.1
−0.15
−0.2
0 500 1000 1500 2000
Samples
10 10
0.18
6
0.16
8 8
5
0.14
Frequency (Hz)
Frequency (Hz)
0.12 4
6 6
0.1
3
4 0.08 4
0.06 2
2 0.04 2
1
0.02
0 0 0 0
500 1000 1500 2000 500 1000 1500 2000
Samples Samples
Fig. 5. Time-frequency analysis of Doppler radar data. (a) The time domain waveforms of both the high gain and low gain Doppler radar data. (b) The time-
frequency representation of Doppler radar data using the proposed multivariate extension of the SST (left panel) and the MPWD algorithm (right panel).
A window length 1063 was used for the MPWD.
proposed method is not accurately representing the com- using the proposed method, while the multivariate pseudo
ponents instantaneous frequencies. Wigner distribution had poorer localization.
is significantly higher than the output of the low gain the Fourier transform of the wavelet function ψ^ having a
radar. The multivariate time-frequency representations compact support in ½1 Δ; 1 þ Δ. The synchrosqueezing
using both the proposed method and the MPWD is shown transform with accuracy δ and threshold ϵ~ (where
in Fig. 5(b). Observe that the proposed method localizes ϵ~ ¼ ϵ 1=3 is the threshold for which jW f ða; bÞj 4 ϵ~ ) is then
the Doppler shift frequency more effectively, where it determined via [16]
should be noted that the speed of the robotic device Z
1 ω ωf ða; bÞ 3=2
between the samples 400–600, and the deceleration Sδf ;ϵ~ ðb; ωÞ ¼ W f ða; bÞ h a da
a:jW f ða;bÞj 4 ϵ~ δ δ
between the samples 1600–1800, can clearly be identified.
Finally the localization ratio for the proposed multivariate ð26Þ
time-frequency7 method is 0.38 while for the MPWD 0.11, where h(t) is a window function which satisfies,
implying that the proposed method has a higher energy R
hðtÞ dt ¼ 1. The following error bounds have been deter-
concentration around the Doppler shift frequency, fd. mined in [16]
6. Conclusion Given a scale band Z ¼ fða; bÞ: jaϕ0 ðbÞ 1jo Δg, then
for each scale-time pair ða; bÞ A Z and jW f ða; bÞj 4 ϵ~ , it
We have proposed a multivariate extension of the syn- follows that
chrosqueezing transform in order to identify oscillations 0
common to the data channels within a multivariate signal. jωf ða; bÞ ϕ ðbÞj r ϵ~ ; ð27Þ
For each channel, the instantaneous frequencies of the which implies that the SST coefficients are concen-
0
synchrosqueezed coefficients are determined for each oscilla- trated along the instantaneous frequency ϕ ðbÞ.
tory scale, and the resulting multivariate instantaneous fre- For all of b A R, there exists a constant C, such that as
quency is then found by calculating the joint instantaneous δ-0, the inverse of the synchrosqueezing transform
frequency of each oscillatory scale across the channels. The along the vicinity of the instantaneous frequency
0
performance of the proposed algorithm has been illustrated curve, ϕ ðbÞ, results in the following error bound:
both analytically, in terms of an error bound, and through
Z !
0 0 where
inf ϕ ðtÞ 4 0; sup ϕ ðtÞ o1
t AR t AR
Z
^
1 δ
0 0
jA ðtÞj; jϕ″ðtÞj r ϵjϕ ðtÞj; 8t AR A n ðbÞ ¼
lim S ðb; ωÞ dω
:
δ-0 Rψ ω:jω ωn ða;bÞj o ϵ~ f ;ϵ~ ;n
N 2
7
It should be noted that the localization ratio of the proposed ^
∑n ¼ 1 A^ n ðbÞωn ða; bÞ
jΩ ð bÞ Ω ð bÞj ¼
Ω ð bÞ
method and the univariate SST of the high gain channel are equal, as multi multi
2 multi
N
∑n ¼ 1 A^ n ðbÞðωn ða; bÞ Ωmulti ðbÞÞ
¼
: ð31Þ (34) as
2
∑N A^ ðbÞ
d h H τ τ i
n¼1 n
x t xþ tþ
jdτ þ 2 2 jτ ¼ 0
Using the following property, j∑N n ¼ 1 yn jr ∑n ¼ 1 jyn j,
N
〈ωðt Þ〉 ¼ τ τ
Eq. (31) can be written as follows: xþ t
H xþ tþ
2 2 jτ ¼ 0
¼ :
2j xHþ ðtÞx þ ðtÞ
^2
∑N n ¼ 1 A n ðbÞ
For the multivariate signal components xn ðtÞ ¼ an ðtÞeı_ϕn ðtÞ ,
the instantaneous frequency of a multivariate signal is
^2
N
A
o ∑
therefore of the form
n¼1
A 2
0
n ¼ 1 an ðtÞϕn ðtÞ
lower
∑N 2
〈ωðt Þ〉 ¼ ¼ ωx ðt Þ:
∑n ¼ 1 a2n ðtÞ
N
^ 2
N
A ðbÞ
¼ ∑
2n
: In a similar way, the instantaneous bandwidth (4)
n ¼ 1
A
A2
with
^ 2
N
A ðbÞ
Z
∑
2n
1 1
d h H τ τ i
2
n ¼ 1
A
A2
ω2 WDðω; t Þ dω ¼ x t xþ tþ :
lower lower
2π 1 dτ 2 þ 2 2 jτ ¼ 0
This analysis can be generalized to the Cohen class of
N
ðA ðbÞ þC ϵ ~ Þ2
n
o ∑
ð32Þ distributions and general time-scale representations,
n¼1
A2lower
A2lower
[14] F. Auger, P. Flandrin, Y.-T. Lin, S. McLaughlin, S. Meignen, T. Oberlin, [21] P. Loughlin, K. Davidson, Modified Cohen–Lee time-frequency dis-
H.-T. Wu, Time-frequency reassignment and synchrosqueezing: an tributions and instantaneous bandwidth of multicomponent signals,
overview, IEEE Signal Process. Mag. 30 (6) (2013) 32–41. IEEE Trans. Signal Process. 49 (6) (2001) 1153–1165.
[15] R. Carmona, W.-L. Hwang, B. Torrésani, Practical Time-Frequency [22] G. Thakur, E. Brevdo, N.S. Fuckar, H.-T. Wu, The synchrosqueezing
Analysis, Academic Press, 1998. algorithm for time-varying spectral analysis: robustness properties
[16] I. Daubechies, J. Lu, H.-T. Wu, Synchrosqueezed wavelet transforms: and new paleoclimate applications, Signal Process. 93 (5) (2013)
an empirical mode decomposition-like tool, Appl. Comput. Harmo- 1079–1094.
nic Anal. 30 (2) (2011) 243–261. [23] L. Stanković, M. Daković, T. Thayaparan, Time-Frequency Signal
[17] I. Daubechies, S. Maes, A nonlinear squeezing of the continuous Analysis with Applications, Artech House, 2013.
wavelet transform based on auditory nerve models, in: Wavelets in [24] I. Djurovic, L. Stankovic, Time-frequency representation based on
Medicine and Biology, CRC Press, 1996, pp. 527–546.
the reassigned s-method, Signal Process. 77 (1) (1999) 115–120.
[18] N. Huang, Z. Shen, S. Long, M. Wu, H. Shih, Q. Zheng, N. Yen, C. Tung,
[25] D. Mandic, N. Rehman, Z. Wu, N. Huang, Empirical mode
H. Liu, The empirical mode decomposition and the Hilbert spectrum
decomposition-based time-frequency analysis of multivariate sig-
for nonlinear and non-stationary time series analysis, Proc. R. Soc. A
nals: the power of adaptive data analysis, IEEE Signal Process. Mag.
454 (1998) 903–995.
[19] S. Olhede, A. Walden, The Hilbert spectrum via wavelet projections, 30 (6) (2013) 74–86.
Proc. R. Soc. A 460 (2004) 955–975. [26] V. Ivanovic, M. Dakovic, L. Stankovic, Performance of quadratic time-
[20] P. Loughlin, B. Tacer, Comments on the interpretation of instanta- frequency distributions as instantaneous frequency estimators, IEEE
neous frequency, IEEE Signal Process. Lett. 4 (5) (1997) 123–125. Trans. Signal Process. 51 (1) (2003) 77–89.