SIGNAL/DATA ANALYSIS & PROCESSING

1

The Fourier Decomposition Method for nonlinear
and nonstationary time series analysis

arXiv:1503.06675v2 [stat.ME] 31 Aug 2015

Pushpendra Singh∗,1,3, Shiv Dutt Joshi1, Rakesh Kumar Patney1, and Kaushik Saha2

Abstract—Since many decades, there is a general perception
in literature that the Fourier methods are not suitable for
the analysis of nonlinear and nonstationary data. In this paper, we propose a Fourier Decomposition Method (FDM) and
demonstrate its efficacy for the analysis of nonlinear (i.e. data
generated by nonlinear systems) and nonstationary time series.
The proposed FDM decomposes any data into a small number
of ‘Fourier intrinsic band functions’ (FIBFs). The FDM presents
a generalized Fourier expansion with variable amplitudes and
frequencies of a time series by the Fourier method itself. We
propose an idea of zero-phase filter bank based multivariate FDM
(MFDM) algorithm, for the analysis of multivariate nonlinear
and nonstationary time series, from the FDM. We also present an
algorithm to obtain cutoff frequencies for MFDM. The MFDM algorithm is generating finite number of band limited multivariate
FIBFs (MFIBFs). The MFDM preserves some intrinsic physical
properties of the multivariate data, such as scale alignment, trend
and instantaneous frequency. The proposed methods produce the
results in a time-frequency-energy distribution that reveal the
intrinsic structures of a data. Simulations have been carried out
and comparison is made with the Empirical Mode Decomposition
(EMD) methods in the analysis of various simulated as well
as real life time series, and results show that the proposed
methods are powerful tools for analyzing and obtaining the timefrequency-energy representation of any data.
Index Terms—The Fourier decomposition method (FDM);
Fourier intrinsic band functions (FIBFs) and analytic FIBFs
(AFIBFs); zero-phase filter bank based multivariate FDM
(MFDM); Empirical Mode Decomposition (EMD).

I. I NTRODUCTION

T

HE time-frequency representation (TFR) is a well established powerful tool for the analysis of time series
signals. There exist many types of time-frequency (TF) analysis methods, e.g. linear (the short-time Fourier transform),
quadratic (the Wigner-Ville distribution) and Wavelet transforms. The TFR is achieved by formulation often referred
as time-frequency distribution (TFD) and provides insight
into the complex structure of a signal consisting of several
components. These approaches have many useful applications,
however the analysis of nonstationary signals are not well
presented by these methods.
Recently developed Empirical Mode Decomposition
(EMD) [1] has provided a general method for examining
the TFD, and has been applied to all kinds of data. The
EMD is an adaptive signal analysis algorithm for the
analysis of nonstationary and nonlinear signals (i.e. signals
∗ Corresponding author’s e-mail address: spushp@gmail.com; pushpendra.singh@ee.iitd.ernet.in
1 Indian Institute of Technology Delhi, Delhi-110 016, India.
2 Samsung R & D Institute India - Delhi, India.
3 Jaypee Institute of Information Technology - Noida, India

generated from nonlinear systems). The EMD has become
an established method for the signal analysis in various
applications, e.g. medical studies [2]–[5], meteorology [1],
geophysical studies [6] and image analysis [7]. The EMD
decomposes any given data into a finite number of narrow
band intrinsic mode functions (IMFs) which are derived
directly from the data, whereas other signal decomposition
techniques (like the Fourier, Wavelets, etc.) incorporate
predefined fixed basis for signal modeling and analysis. The
Ensemble EMD (EEMD) is a noise-assisted data analysis
method developed in [8] to overcome the timescale separation
problem of EMD. The Multivariate EMD (MEMD) developed
in [9] is a generalization of the EMD for multichannel data
analysis. The Compact EMD (CEMD) algorithm is proposed
in [10] to reduce mode mixing, end effect, and detrend
uncertainty present in EMD and to reduce computation
complexity of EEMD as well. The IMFs, generated by EMD,
are dependent on distribution of local extrema of signal
and the type of spline used for upper and lower envelope
interpolation. The traditional EMD uses cubic spline for
upper and lower envelope interpolation. The EMD algorithm,
proposed in [11] to reduce mode mixing and detrend
uncertainty, uses nonpolynomial cubic spline interpolation to
obtain upper and lower envelopes, and have shown in [12]
that it improves orthogonality among IMFs.
The energy preserving property is important for any kind
of transformation, and it is obtained by the orthogonal decomposition of signal in various transforms like the Fourier,
Wavelet, Fourier-Bessel, etc. The energy preserving property
is especially important for the accurate and faithful analysis
of three dimensional time-frequency-energy distribution of a
signal. The EMD algorithms, proposed in [13], ensure orthogonality or energy preserving property or both in decomposition
of signal into IMFs and refereed to as energy preserving EMD
(EPEMD).
In spite of considerable success of EMD, all of the EMD
algorithms are based on empirical, heuristic and ad hoc procedure that make them hard to analyze mathematically, and EMD
may suffer from mode mixing, detrend uncertainty, aliasing
and end effect artefacts [14]. There is a lack of mathematical
understanding of the EMD algorithms, e.g. IMFs dependence
on the number of sifting and the stopping criteria, convergence
property and stability to noise perturbation. In spite of all
these limitations, EMD is the widely used nonstationary data
analysis method. Therefore, in this paper, EMD is used as a
reference to establish the validity, reliability and calibration of
the proposed method.
Since many decades, there is an understanding in the litera-

we explore and provide algorithms to analyze nonlinear and nonstationary data by the Fourier method termed as the Fourier Decomposition Method (FDM). i. in this study. yi (t)+j yˆi (t) = ai (t) exp(jφi (t)). An analytic signal zi (t) can be represented by zi (t) = yi (t) + j yˆi (t) = ai (t) exp (jφi (t)). The necessary conditions [1]. i (t) = φ′i (t) > 0 ∀t) and well-behaved Hilbert with ai (t). The Fourier transform is valid under very general Dirichlet conditions (i. and finite number of finite discontinuities in any finite interval) thus include nonlinear and nonstationary signals as well. directly. Synchrosqueezed wavelet transforms [18]. Therefore. (3) The FIBFs provide analytic FIBFs (AFIBFs) with instantaneous frequency (IF) and amplitude always greater than zero. the WignerVille distribution. for a basis to represent a nonlinear and nonstationary time series.e. etc.e. This multivariate FDM (MFDM) algorithm generates matched multivariate FIBFs and residue through zero-phase filtering. [14]. The power of the Fourier transform can also be realized from the fact that the analytic representation and.e. T HE F OURIER D ECOMPOSITION M ETHOD We propose a class of functions. We propose the Fourier decomposition method (FDM) in section III. we propose the FDM and MFDM general adaptive data analysis methods that are inspired by the EMD methods and their filter bank properties [15]. with ai (t). Rb (2) The FIBFs are orthogonal functions. b]. using a zero-phase filter bank (ZPFB) approach to construct the multivariate FIBFs and residue components.g. is zero. evolutionary spectrum [17]. i. which generates small number of Fourier intrinsic band functions (FIBFs). ∀t.e. the average of the values of upper and lower envelopes.e. we also propose a method. belonging to C ∞ [a. yi (t) ∈ C ∞ [a. where ai (t) = [yi2 (t) + yˆi 2 (t)]1/2 and φi (t) = tan−1 [yˆi (t)/yi (t)] are instantaneous amplitude and phase of yi (t). signals which can be represented by the following model [24]: x(t) = M X yi (t) + n(t). and various reasons (e. dφdt where n(t) is a noise representing any residue (constant or trend) components and the yi (t) are M single component nonstationary signals. inherently. yˆi (t) = 1 ∞ yi (τ ) π −∞ t−τ dτ and the Hilbert transform emphasizes the local properties of yi (t). (1) i=1 where yi (t) is the ith IMF and rℓ (t) = yℓ+1 (t) is final residue. we show that the Fourier method is also a best tool for time-frequency analysis and processing of any signal. we propose an adaptive. yi (t) ≈ ai (t) cos(φi (t)). signal is absolutely integrable with finite number of maxima and minima. II. The IMFs admit amplitude-frequency modulated (AM-FM) representation [18] (i. present in the Fourier transform. defined above. the IF has meaning only for monocomponent signals. For the adaptive data analysis approach. i. obtained by the interpolation of local maxima and the local minima.e. in this study. which captures the features of the MEMD. The main objective of this study is to obtain unique representation of multicomponent signal as a sum of constant and monocomponent signals. hence.g. It is already well established that the Fourier method is best tool for spectrum analysis and. ZPFB based time-frequency analysis method. [24]) that the Fourier methods are. a yi (t)yj (t) dt = 0. The Bedrosian and Nuttall theorems [19]. This paper is organized as follows: In section II the EMD algorithm is briefly presented. data-driven. not suitable for nonlinear and nonstationary data analysis. signal has only one frequency or a narrow range of frequencies varying as a function of time [24]. its Hilbert transform yˆi (t) is Rdefined as convolution of yi (t) and 1/πt. the Hilbert transform of a signal are.e. a yi (t) dt = 0. Simulation results are presented in section V. the empirical orthogonal function expansion (EOF) (or principal component analysis or singular value decomposition). the AFIBFs are monocomponent signals and. III.e. i. . The instantaneous frequency (IF) of yi (t) is defined as: ωi (t) = φ′i (t) = yˆi ′ (t)yi (t)−yˆi (t)yi′ (t) . The EMD is a well established signal decomposition method that decomposes nonlinear and nonstationary data into a set of finite band-limited IMFs and residue through the sifting process. i. Definition 1: The Fourier intrinsic band functions (FIBFs). which in our proposed framework would be the FIBFs. The first condition ensure that IMFs are narrow band signals and the second condition is necessary to ensure that the IF does not have redundant fluctuations because of asymmetric waveforms [1]. for i 6= j. b]. B RIEF OVERVIEW OF THE EMD ALGORITHM There are various tools for nonstationary data processing such as the spectrogram. d i. Thus. All IMFs must satisfy two basic conditions: (1) In the complete range of time series. dt φi (t) ≥ 0. are functions that satisfy the following conditions: Rb (1) The FIBFs are zero mean functions. Thus. periodicity or stationarity) are provided to support it.. In this paper. Finally conclusions are presented in section VI.e. In this study. physically. the number of extrema (i. [20] further impose non-overlapping spectra constraints on the pair [ai (t). linearity. [1]. the EMD.2 SIGNAL/DATA ANALYSIS & PROCESSING ture (e. We propose the ZPFB based multivariate FDM algorithm in section IV. here with the following formal definition. (2) At any point of time in the complete range of time series. termed as the Fourier intrinsic band functions (FIBFs). Thus. the wavelet analysis. the FIBF is sum of zero mean sinusoidal functions of consecutive frequency band. [16]. The physical meaning of IF ωi (t) conyˆi 2 (t)+yi2 (t) strains that φi (t) must be a mono-component function of time. The decomposed signal x(t) is expressed as the sum of ℓ IMF components plus the final residue as x(t) = ℓ X i=1 yi (t) + rℓ (t) = ℓ+1 X transforms [1]. cos(φi (t))] of a signal yi (t) = ai (t) cos(φi (t)). maxima and minima) and the number of zero crossings are equal or differ at most by one. (2) i=1 yi (t). For any IMF yi (t). the most difficult challenge has been to establish a general adaptive decomposition method of analysis without a priori basis. are completeness.

et al. a = 0 T T t1 0 0 R t1 +T0 2 = xT0 (t) cos(kω0 t) dt and bk ak = T0 t1 R t1 +T0 2 xT0 (t) sin(kω0 t) dt. (4) 2 k=1 where ck = (ak − jbk ) and c∗k = (ak + jbk ). respectively. i. aM (t) exp (jφM (t)) = ∞ 0 k=(NM −1 +1) k ai (t) exp (jφi (t)) = Ni X ck exp (jkω0 t). The presence of energy at each frequency in MFHS h(f ) means that. (12) 2 Hence. follow all the necessary conditions by virtue of the decomposition. · · · . (8) would change to k = Ni to (Ni−1 − 1). nonstationary and generated from the systems that are generally nonlinear. we obtain the energy of signal x(t) (or power of signal xT0 (t)) P∞ by the Parseval’s theorem as Ex = a20 + 12 i=1 [a2n + b2n ] and from Eq. which can be used to measure the fluctuation of energy with time. derived from Hilbert spectrum. The frequency in either H(f. whereas. energy of zero mean signal is half of the energy of its analytic signal. it is clear that xT0 (t) = a0 + Re{zT0 (t)}. The FIBFs. The exact occurrence time of that oscillation is given in the full Fourier-Hilbert spectrum. t1 + T0 ] real valued signal which follows the Dirichlet P∞ conditions. adaptive and purely Fourier based. From Eq. For each FIBFs. In order to obtain minimum number of AFIBFs in low to high frequency scan (LTH-FS). ωi (t) = The marginal Fourier-Hilbert spectrum offers a measure of total amplitude (or energy) contribution from each value of frequency in a probabilistic sense. t) or h(f ) has a different meaning from the Fourier spectral analysis [1]. and adaptiveness. (7) by the Fourier method itself. here we define the marginal Fourier-Hilbert spectrum (MFHS) from the FHS as follows: Z T0 H(f. (3) k=1 R t1 +T0 1 where frequency (rad/s) ω0 = 2π xT0 (t) dt. (6) energy of the analytic signal (or power of signal P∞ zT0 (t)) as Ez = i=1 [a2n + b2n ]. for t1 ≤ t ≤ t1 + T0 and zero otherwise. Similarly. Moreover. Observe that (7) has precisely the form that in the literature [1] is termed as a generalized Fourier expansion. a1 (t) exp (jφ1 (t)) = k=1 ck exp (jkω0 t). (9) is satisfied for i = 1. it is worth noting that this representation is complete. in forward search (i. intrinsically. (10) h(f ) = 0 k=1 P ∗ is complex conjugate of z˜T0 (t) . . start with (Ni−1 + 1) and append more term till we reach the maximum value of Ni such that (Ni−1 + 1) ≤ Ni ≤ ∞ and dφi (t) ≥ 0. We write (3) as T0 t1 ∞ xT0 (t) = a0 + 1X [ck exp (jkω0 t) + c∗k exp (−jkω0 t)]. we define the three dimensional {t. orthogonal. we have obtained a variable amplitude and frequency representation. in (7). P c exp (jkω t). where w(t) = 1. k=(N1 +1) k exp (jkω0 t). Thus. t) df. Since in practice the continuous time signals are. t). The marginal Hilbert spectrum (MHS). in reverse search (i. local. The FIBFs are the real part of AFIBFs presented in Eq. the classical Fourier expansion provides the constant amplitude and fixed-frequency representation. M . we have obtained a generalized Fourier expansion of a time series in Eq. ∞ k=1 ck exp (−jkω0 t) and Re{zT0 (t)} denotes the real part of zT0 (t). The variable amplitude and the IF have improved the efficiency of the expansion by expanding the signal into finite number of analytic FIBFs. we ∞ ck exp (jkω0 t). We construct the periodic signal as xT0 (t) = s=−∞ x(t − sT0 ) such that x(t) = xT0 (t)w(t). (11) E(t) = 0 where fM is a maximum frequency of signal. therefore.: THE FOURIER DECOMPOSITION METHOD 3 orthogonality. in the total duration of the signal. form high to low frequency scan (HTL-FS))P of AFIBFs. The Fourier series expansion of xT0 (t) is given by xT0 (t) = a0 + ∞ X [ak cos(kω0 t) + bk sin(kω0 t)]. and enabled the expansion to accommodate nonstationary data. (5) where analytic function zT0 (t) . low to P high frequency scan) N1 of AFIBFs. k 0 t). decrease and select minimum value of Ni such that 1 ≤ Ni ≤ (Ni−1 − 1) and Eq.SINGH P. there is a higher likelihood for such a wave (FIBF) to have appeared locally. for each i. generally. (3).e. as Z fM H 2 (f. with N0 = ∞. ai (t)} time-frequency distribution of amplitude as the Fourier-Hilbert spectrum (FHS) H(f. The available data are usually of finite duration. locality. and k=1 the lower and upper limits of sum in Eq. relation between Ex and Ez is given by Ez . (8) k=Ni−1 +1 with N0 = 0 and NM = ∞. NM = 1.e. We write zT0 (t) as zT0 (t) = M X ai (t) exp (jφi (t)). PN2 a2 (t) exp (jφ2 (t)) = c ···. Similarly. ∀t. the amplitude ai (t) and IF fi (t) are functions of time. is defined in [1]. (9) dt It is easy to observe that such a decomposition is always possible. k=N2 ai (t).e. therefore. (8). so Ex = a20 + . Here. fi (t). Let x(t) be a time limited [t1 . t) dt. discretized for further processing by a computing device. We can also define the instantaneous energy density. obtain a1 (t) exp (jφ1 (t)) = P(N1 −1) k=N1 a2 (t) exp (jφ2 (t)) = c exp (jkω ···. From (4). Thus. (7) i=1 where. The frequency in the MFHS indicates only the likelihood that an oscillation with such a frequency exists. we start with (Ni−1 − 1). ∞ X ck exp (jkω0 t) (6) P(NM −1 −1) aM (t) exp (jφM (t)) = ck exp (jkω0 t).

we write analytic signal z1 [n] as N 2 −1 X M X[k] exp( k=1 X j2πkn ai [n] exp(jφi [n]). k= N +1 X[k] exp( j2πkn N ) 2 and we can write (14) as x[n] = X[0] + 2Re{z1[n]} + X[ N ](−1)n . where hz [n] = h[n] ∗ h[−n] and Hz [k] = |H[k]|2 . we can write x[n] as x[n] = N −1 X X[k] exp( k=0 j2πkn ). we observe that the operation that generates the FIBFs is nothing but the Fourier based zerophase filtering (ZPF).e. a2 [n] exp (jφ2 [n]) = 2 P(NMk=N −1 −1) aM [n] exp (jφM [n]) = X[k] exp( j2πkn k=1 N ). Similarly.4 SIGNAL/DATA ANALYSIS & PROCESSING we present the FDM for discrete signal. M with N0 = 0 and NM = N2 . a [n] exp (jφ [n]) = 2 2 k=1 N PN2 j2πkn k=(N1 +1) X[k] exp( N ). aM [n] exp (jφM [n]) = P N2 −1 j2πkn k=(NM −1 +1) X[k] exp( N ). STEP 1. i. Thus. (19) Apply ZP-HPF with cutoff frequency fc2 to set of residue rp1 (t) and obtain second set of MFIBF yp2 (t).e. i. for each i. and propose a multivariate FDM (MFDM) algorithm to generate multivariate FIBFs (MFIBFs) and residue as follows: Apply zero-phase high pass filtering (ZP-HPF) with cutoff frequency fc1 to each of the components of the P-variate (Pchannel) time series {xp (t)}P p=1 and obtain first set of MFIBF yp1 (t). then X[0] and X[ N2 ] are real numbers. y[n] = x[n] ∗ (hz [n]) ⇒ Y [k] = X[k]Hz [k]. k=1 N ) PN −1 is complex conjugate of z2 [n] . )= N i=1 (16) where. obtain maximum value of Ni such that (Ni−1 +1) ≤ Ni ≤ ( N2 −1) and phase φi [n] is a monotonically increasing function. Set AF IBFi = N k=(Ni−1 +1) X[k] exp( N ) = ai [n] exp(jφi [n]). · · · . p = 1. respectively. · · · . (2) The Fourier method.e. i. The FDM with Fourier transform (FT) and discrete time Fourier transform (DTFT) is summarized in appendix. N (13) PN −1 −j2πkn ) is the DFT of where X[k] = N1 n=0 x[n] exp( N signal x[n]. The ZPF of a real valued signal x[n] by zero-phase filter (hz [n]) can be obtained by two methods: (1) Convolution method.e. X[k] exp( j2πkn ). to use the Fourier or other methods of zero-phase filtering to decompose any data into a set of FIBFs. low to high frequency and high to low frequency view. i. where 1 PN −1 X[k] = N k=0 x[n] exp(−j2πkn/N ). P. and we can write x[n] as N 2 x[n] = X[0] + −1 X X[k] exp( k=1 + j2πkn N ) + X[ ] exp(jπn) N 2 N −1 X X[k] exp( k= N 2 +1 j2πkn ). · · · . therefore. i. NM = 1. in addition to the FB properties of IMFs. · · · . Let N be an even number (we can proceed in the similar fashion when N is an odd number). Observe that such a decomposition always exists. ωi [n] = (φi [n + 1] − φi [n]) ≥ 0 or ωi [n] = ( φi [n + 1] − φi [n − 1] ) ≥ 0. of the signal and generate two set of time-frequency-energy distribution. for i = 1. 2. Here. In order to obtain minimum number of FIBFs in LTH-FS. · · · .e. we scan from (Ni−1 + 1) to ( N2 − 1). Let. 2 Algorithm B: The FDM algorithm (HTL-FS) to obtain AFIBFs. STEP 1. M with N0 = N2 and NM = 1.e. where h[n] is a real sequence. we obtain P N2 −1 a1 [n] exp (jφ1 [n]) = X[k] exp( j2πkn N ). Now. Algorithm A: The FDM algorithm (LTH-FS) to obtain AFIBFs. we scan from (Ni−1 − 1) to 1. ωi [n] = i [n−1] ) ≥ 0.e. Using the discrete Fourier transform (DFT). i. both view may be same or sometimes they reveal two different types of features of the signal. ai [n] exp(jφi [n] = Ni X k=Ni−1 +1 X[k] exp( j2πkn ). be a discrete signal of length N . i [n−1] ωi [n] = ( φi [n+1]−φ ) ≥ 0. x[n]. The first set of residue is obtained as follows: rp1 (t) = xp (t) − yp1 (t). in HTL-FS for FIBFs.e. low to high frequency scan) we obtain a1 [n] exp (jφ1 [n]) = PN1 of AFIBFs. for i = 1. N (17) with N0 = 0 and NM = ( N2 − 1). obtain the minimum value of Ni such that 1 ≤ Ni ≤ (Ni−1 − 1) and phase φi [n] is a monotonically increasing function. obtain output by the inverse N −1 DFT. N (14) P N2 −1 X[k] exp( j2πkn Since x[n] is real. P. ∀n. Set AF IBFi = k=Ni N ) = ai [n] exp(jφi [n]). i. ( φi [n+1]−φ 2 IV. Obtain X[k] = F F T {x[n]}. for each i. in forward search (i. The second set of residue is obtained as rp2 (t) = rp1 (t) − yp2 (t) p = 1. (17) will change to k = Ni to (Ni−1 − 1). P i j2πkn STEP 2. This is another motivation. 2 ∀n (18) and ai [n] ≥ 0 for i = 1. 2. The FDM is summarized in Algorithms A and B. Depending on the signal. (20) . z1 [n] . set Hz [k] = 1 at desired frequency band and Hz [k] = P 0 otherwise. FDM provides two views. obtain minimum value of Ni such that 1 ≤ Ni ≤ (Ni−1 − 1) and phase φi [n] of AF IBFi is a monotonically increasing function. We use ZPF that does not shift the essential features of the signal. · · · . and the lower and upper limits of sum in Eq. P(Ni−1 −1) X[k] exp( j2πkn STEP 2. 2 (15) where Re{z1 [n]} denote the real part of z1 [n]. with N0 = N2 . Obtain X[k] = F F T {x[n]}. M ULTIVARIATE F OURIER D ECOMPOSITION M ETHOD From (8) and (17). ∀n. obtain maximum value of Ni such that (Ni−1 + 1) ≤ Ni ≤ ( N2 − 1) and phase φi [n] of AF IBFi is a monotonically increasing function. PN1 −1 k=N1 j2πkn X[k] exp( N ). · · · . y[n] = k=0 X[k]Hz [k] exp(j2πkn/N ). M and ∀n.

xj (t) = i∈[1. i. fc2 = fM fM 22 . 4). a 8Hz sinusoid is common to all channels. Thus. EMD [27] and FDM methods in data analysis. uniform and non-uniform CFs. From (23) we obtain fLi = [(2m − 1)/(2m + 1)]fHi . fcℓ corresponding to zero-phase high pass filters hz1 (t). Table I presents comparisons among Fourier. On the same machine. In MFDM. maxima. it can not be guaranteed simultaneously to all P-channel data. detrend uncertainty and end effect artefacts as extraction of FIBFs does not depend on distribution of local extrema across the range of signal. 2. Repeat step 2 to 3 for i = 1. hz2 (t). et al. The zerophase filtering of time series can be obtained through the finite impulse response (FIR) or infinite impulse response (IIR) filters. The MFDM algorithm. we observe that the ratios. CFs fci and BWs (fHi − fLi ) can be taken as a constant. · · · .5 1 1 rpℓ (t) = rp(ℓ−1) (t) − ypℓ (t) p = 1.4] sin(2πfi t) + nj (t) (for j = 1. and as [m → ∞. which is summation of sinusoids (with combination of frequencies f1 = 4Hz.SINGH P. and greater the value of m (or lesser the value of l) narrower the band.5 100 1 200 300 200 300 −1 0 1 −5 0 100 200 300 100 200 300 100 200 300 −1 0 1 100 200 300 0 100 200 300 −1 0 100 200 300 100 200 300 200 300 1 100 200 300 −1 0 1 100 200 300 −1 0 1 100 200 300 100 200 300 100 200 300 100 200 300 0 100 200 300 −1 0 1 0 100 200 300 0 100 0. we can use zero-phase low pass filtering (ZPLPF) in place of ZP-HPF to decompose signal in order of residue to first MFIBFs. Similar to the MEMD and noise-assisted MEMD (NAMEMD) [16]..e. 2.e. fc1 = f2M . f2 = 8Hz. ℓ. (8). (fHi+1 − fLi+1 ) (25) From (23). 2. Wavelet. The question is. This is similar to MEMD algorithm problem where in derivation of multivariate IMFs. MFDM is computationally more efficient. (23) where fHi is the highest frequency and fLi is the lowest frequency of ith band of a filter bank. If required. fc2 . second and third channels. A 32Hz sinusoid is present to first. · · · . P. we obtain l = (2m + 1)/(2m − 1) or m = (1/2)(l + 1)/(l − 1) with l > 1. to obtain MFIBFs. i. ypℓ (t). we define ratio of center frequency (fC ) to bandwidth (BW) as m = fCi /(fHi − fLi ).5 sec. f4 = 32Hz) and Gaussian white noise P of zero mean and standard deviation of 0. rpℓ (y). dyadic (i. 2. as in Eq. third and fourth channels. · · · . for P-variate time series as xp (t) = ℓ X ypi (t) + rpℓ (t). (24) and (25). We can take the Fourier transform of signal x(t) to obtain its spectrum details and make strategy to decide CFs. · · · . m > 1/2. 1: MFDM applied to a quadri-variate tone-noise mixture. we can vary the value of m (or l) for each band rather than taking the fixed value. where fM is the maximum frequency of a signal x(t) and for the sampled signal. i. (24) From (23) and (24).2. yp1 (t). EMD and EEMD [22] have been used in simulation results. e. Algorithm C: An algorithm to obtain cutoff frequencies fci . in this simulation.5 0 −1 0 1 −1 0 1 −1 0 300 0 −1 0 1 300 200 0 1 0 −1 0 1 0 −1 0 1 0 200 1 0 0 100 100 −1 0 1 0 0 0. 3. in all the four channels. Select suitable value of m and set fH1 = Fs /2. this MFDM algorithm produces the equal number of scale-aligned MFIBFs for all channels and preserving joint channel properties that make it suitable for direct multichannel modelling. P. generating perfectly aligned intrinsic bands. similar to (1). obviously. Set fHi+1 = fci . a 16Hz sinusoid is present to first. We 0 −1 0 1 100 200 300 100 200 300 5 5 0 0 0 −5 −5 0 1 0 −1 0 1 −1 0 1 0 0 −1 0 1 −1 0 1 100 200 300 0 −1 0 1 0. etc. A. p = 1. · · · .g. V.5 are fixed values. 4. S IMULATION RESULTS The online available MATLAB software of MEMD [21]. whereas conventional (non zero-phase) filtering shifts the features in the signal and hence cannot be used. how to obtain cutoff frequencies (CFs) fc1 .5 1 0 −1 0 1 0 0 100 x4(t) + n4(t) 5 0 0 −1 0 1 x3(t) + n3(t) x2(t) + n2(t) x1(t) + n1(t) 5 −1 0 1 0 −1 0 100 200 300 −1 0 Fig. l → 1]. second and fourth channels. minima. whereas. non-dyadic. · · · . (22) i=1 When we use the Fourier based zero-phase filtering. fcℓ = 2ℓ . computation time for MFDM is 0. Set fci = [(2m − 1)/(2m + 1)]fHi .) in the filtered time waveform exactly at the time where those features occur in the unfiltered waveform. use zero-phase filtering as it preserves salient features (e. fCi /fCi+1 = fci /fci+1 = (fHi − fLi ) = l. generating perfectly aligned intrinsic modes in all the four channels. for the consecutive ith and (i + 1)th bands.e. (21) Through the addition of (19). we here propose the compact and elegant way to decide CFs as summarized in Algorithm C. . maximum frequency is ( F2s ) half of the sampling frequency). [m → 1/2. and for MEMD is 69. · · · . The FDM does not suffer from mode mixing. 1.e. · · · . l → ∞]. of center frequencies fCi . a 4Hz sinusoid is present to first. first two conditions of FIBFs are fully satisfied and the third one is approximately satisfied (i. Here. fCi = (fHi + fLi )/2. as shown in Figure 1. we have liberty to select any suitable value of l or m. hzℓ (t)? There is lot of flexibility and are various ways to select CFs.e. For narrowband signal. satisfied in all practical sense). Multivariate data decomposition We used quadri-variate time series signal. similar to MEMD.: THE FOURIER DECOMPOSITION METHOD 5 We can repeat this ZP-HPF procedure ℓ times and obtain final set of MFIBF ypℓ (t) and residue (with cutoff frequency fcℓ ) 0 −5 0 0.45 sec.g. f3 = 16Hz. whereas in the case of dyadic FB l = 2 and m = 1. first condition of IMF is not imposed [9]. (20) and (21) we obtain expression.

is a main cause of mode mixing [e. C. E.5 3 3.5 2 Time 2.5 3 3.5 2 Time 2. Time-Frequency-Energy Analysis 0 −1 0 EMD-Hilbert adaptive differentiation: local no frequencytime-energy yes yes no empirical 0.21 sec. our method well applies to these challenging cases with good accuracy.5 3 3.5 1 1. We demonstrate that.. unit variance.5 4 4 y −y 5 −0.6 SIGNAL/DATA ANALYSIS & PROCESSING TABLE I: Comparisons among the Fourier. Wavelet.2 cos(64πt)) + 1.5 1 1. These examples clearly demonstrate that the FDM can indeed analyze nonlinear signals and it is a nonlinear decomposition method..02 0 Time 0.02 y6−y7 0 1 y1 y2 1 0 −1 −2 −3 0 2 FDM a priori 7→ adaptive differentiation: local no frequencytime-energy yes yes yes 7→ no complete ting in EMD algorithm. Intrawave frequency modulation First. computation time for MFDM. −2 0 y4−y5 2 0 y6−y10 −2 0 0. 2: Decompositions of a sinusoid mixed with intermittent interference through algorithm of (b) FDM (c) EMD (d) EEMD.5 2 Time 2. There is a enhanced TFE tracking when using FDM with low to high frequency scan and other one (HTL-FS) is similar to plot obtained by EMD algorithm.5 1 1.5 4 0. 6 and 7. Fourier a priori convolution: global yes frequencyenergy no no yes complete Basis Frequency Uncertainty Presentation Nonlinear Nonstationary Harmonics Theoretical base Wavelet a priori convolution: regional yes frequencytime-energy no yes yes complete (a) x(t) 1 0 −1 0 0.94 dB signal-to-noise power ratio.5 1 1.02 0 −0. we decompose the following signal that has intrawave frequency modulation and it is considered challenging because the instantaneous frequency itself has very high frequency modulation [23] −0.5 3 3. 1024 samples and sampling frequency Fs = 100 Hz) obtained from the FDM and EMD .2 + cos(2πt) 1.5 3 3. obtained using the FDM and EMD. Analysis of a white Gaussian noise B.5 3 3.05 0 0.5 4 0.5 2 Time 2.01 y1 0 y2 0. signal x(t) in Figure 2(a)] and mode split- Figure 8 shows the TFE analysis of a white Gaussian noise (with zero mean.5 2 2. EMD and EEMD are 0.02 0 −0.5 2 Time (b) 2. we consider a model wave −0. 0.04 0 0.2 0.5.5 4 0.5 3 3.5 3 3.5 2 Time 2.5 1 1.05 0 y4−y5 0.01 0 x(t) = Time Fig.56 sec.5 1 1. in the time series. Intermittency and mode mixing The intermittency.18 sec.5 1 1.5 3 3.02 −0.04 0.5 4 0.5 2 2.5 4 y1−y3 0. The ensemble size for EEMD was N = 500 with the 16.g.5 4 0.5 2 Time (d) 2.5 4 0.5 3 3.5 4 0. Figure 5.5 3 3.5 1 1.5 4 0. On the same machine.5 3 3. The MFDM is able to localize the mono-component sinusoid within a single FIBF and outperforming EMD 2(c) and EEMD 2(d).05 0 −0. These issues are mitigated by EEMD and NA-MEMD. respectively. EMD-Hilbert and proposed FDM methods in data analysis. and 77.5 2 Time 2.5 4 0 D.5 1 1.5 2 2.5 2 2.1 0 −0.5 1 1.5 1 1.5 1 1.1 Time y6−y10 1 0 −1 0 1 cos(32πt + 0. x(t) = cos(ωt + ǫ sin ωt) (27) that satisfies the following highly nonlinear differential equation [1] p d2 x(t) +[ω+ǫω cos(ωt)]2 x(t)−[ǫω 2 sin(ωt)] 1 − x2 (t) = 0 2 dt (28) with ω = 1 and ǫ = 0.5 + sin(2πt) (26) Second. The decompositions of a signal x(t) through FDM algorithm is shown in Figure 2(b) without end effect artefacts.05 0.5 4 Time (c) 0 Figure 3 shows time-frequency-energy (TFE) estimates for a nonstationary signal mixture of a linear chirp and frequency modulated (FM) sinusoid.

2 0.4 0. There are enhanced TFE and PSD tracking when using FDM.1 0. . respectively. FIBF2 and highest frequency component by FDM. The DC n and highest frequency component correspond to term X[0] and X[ N of Eq.8 0.4 0. Sum of DC.9 1 0.1 0. 4: The signal x(t) (solid line).g.2 0.1 0.9 1 Frequency (Hz) 14 1000 10 8 6 EMD: Time−Frequncy−Energy estimate of IMFs 2000 4 1800 2 1600 Frequency (Hz) 12 1400 0 0.7 0. sum of DC and lowest frequency FIBF (dashed line) of (26).1 0. both LTH-FS and HTL-FS views of TFE is similar and complete data is decomposed in FIBFs.6 0.7 0.5 Time (s) 1200 EMD: Time−Frequncy−Energy estimate of IMFs 1000 0.6 0.3 0.5 Time (s) 0.8 0.8 0.7 0.5 Time (s) 0. FIBF1.9 1 Fig. FIBF2 and highest frequency component exactly synthesize signal x(t) given by Eq.8 0.3 0. it is mid frequency of the one band in other PSD (HTL-FS) view.5 Time (s) 0.7 0.1 0.3 0.SINGH P. 6: The TFE analysis of signal x(t) given by Eq. 2 ](−1) (15).9 1 Frequency (Hz) 600 20 15 10 Fig.5 Time (s) 0.9 1 0.9 1 0.4 0.9 1 FDM: Time−Frequncy−Energy estimate of FIBFs (HighToLowFrequencyScan) Amplitude Frequency (Hz) 2500 4 2 0 0 0.8 0.6 0.8 0. algorithm.4 0. 5 0 0 6 0. approximately 40 Hz is cutoff frequency for one of the band in PSD (LTH-FS).9 1 5 0 −5 0 2 0 −2 0 0.4 0.: THE FOURIER DECOMPOSITION METHOD 7 Amplitude FDM: Time−Frequncy−Energy estimate of FIBFs (LowToHighFrequencyScan) 3000 Amplitude 2000 Amplitude 1500 1000 500 0 0.8 0.6 0.2 0.7 5 Fig. e.6 0.1 0.4 0.6 0.3 0.7 0.9 1 30 800 25 400 200 0 0 0.1 0.3 0.7 0.1 −17 x 10 5 0 −5 0 0.4 0.1 0.3 0. Both LTH-FS and HTL-FS views of PSD looking similar but FIBFs have different frequency band.5 Time (s) 0. (26).5 Time (s) 0.8 0. FIBF1.5 Time (s) 0.5 Time (s) 0. whereas.2 0.7 0. 3000 Frequency (Hz) 2500 FDM: Time−Frequncy−Energy estimate of FIBFs (LowToHighFrequencyScan) 2000 18 16 1500 0 0.3 0. The FDM has dived the complete data in narrowband and orthogonal FIBFs. 3: The TFE analysis of nonstationary signals which is mixture of linear chirp and FM sinusoid: FDM (top and middle) and EMD (bottom).7 0. 5: The DC.2 0.6 0.2 0.3 0.6 0. Clearly.8 0. (26): FDM (top) and EMD (bottom).2 0.5 Time (s) 0.4 0.7 0.6 0.3 0.2 0.6 0.1 Fig.8 0. 4 Amplitude 3 2 1 0 −1 −2 0 0.4 0.2 0.2 0. Figure 9 shows the power spectral density (PSD) plot of same white Gaussian noise with the FDM and EMD algorithm.3 0.9 1 0.4 0. et al.

15 0.17 0.16 0.05 10 5 0 0 5 10 15 Time (s) 20 25 0 0 30 1 2 3 4 5 Time (s) 6 7 EMD: Time−Frequncy−Energy estimate of IMFs Fig.19 40 0.99 sec. Theoretically. its limitation could only be applied to such integral transforms. The uncertainty principle is a consequence of the Fourier transform (or any other type of integral transform) pair in the frequency definition. This clearly indicate that the TFE plot obtained by the FDM method is same as theoretical estimation. F4s Hz. Consequently. The IF is defined through differentiation of phase rather than integration and. (n0 = 199) and frequency f = 25 Hz. . hence. (27): FDM (top) and EMD (bottom). sampling frequency Fs = 100 Hz. Figure 10 shows the plots of real. a[n] = sin( π 2 (n−n0 )) | π (n−n |. we obtain the analytic representation of x[n] = δ[n − n0 ] ⇔ X(ω) = exp(−jωn0 ) as 30 25 20 15 10 (29) 0 )) where real part of z[n] is δ[n − n0 ] = sin(π(n−n π(n−n0 ) . overcome the restriction of the uncertainty principle. 45 40 F.e. ‘integrated out’ or smeared over the integral time limit. time-frequency analysis of signal would not be bounded by the uncertainty principle [26]. However.15 30 25 20 15 0. EEMD and continuous wavelet transform (CWT) methods. i. imaginary part and absolute value of z[n] with n0 = 199. Figure 12 shows the plots of FIBFs obtained by the FDM and Figure 11 shows the TFE analysis of unit sample sequence δ[n−n0 ] from the FDM. if we avoid an integral transform in the frequency computation. length N = 400. the cruder the resulting frequency resolution would be. whereas there is energy spread over a range of frequencies and lack of accuracy in TFE plot by the EEMD and CWT methods.14 10 0. in which time would be 5 0 0 1 2 3 4 5 Time (s) 6 7 Fig. 8: The TFE analysis of the Gaussian white noise (with zero mean and unit variance): FDM (top and middle) and EMD (bottom).18 Frequency (Hz) Frequency (Hz) 35 0.2 40 Frequency (Hz) Frequency (Hz) 35 0. Therefore. this example clearly demonstrate that the FHS by FDM is indeed not limited by uncertainty principle and signal can be highly concentrated in both time and frequency domain.13 0 5 5 10 15 Time (s) 20 25 0 0 30 1 2 EMD: Time−Frequncy−Energy estimate of IMFs 3 4 5 Time (s) 6 7 8 9 10 8 9 10 8 9 10 FDM: Time−Frequncy−Energy estimate of FIBFs (HTL−FS) 45 0. z[n] = Frequency (Hz) 35 The unit sample sequence is defined as δ[n − n0 ] = 1 at n R= n0 and zero otherwise. TFE Analysis of unit sample sequence sin(π(n − n0 )) + j[1 − cos(π(n − n0 ))] π(n − n0 ) z[n] = a[n] exp(jφ[n]). The uncertainty principle in signal analysis state that the finer the time resolution one wants.8 SIGNAL/DATA ANALYSIS & PROCESSING FDM: Time−Frequncy−Energy estimate of FIBFs (LowToHighFrequencyScan) FDM: Time−Frequncy−Energy estimate of FIBFs (LTH−FS) 45 0. this clearly indicate that most of the energy of signal δ[n − n0 ] is concentrated at time t = 1. One can obtain arbitrary precision on time and frequency resolution subject only to the sampling rate in case of discrete time signal.1 30 25 20 15 0. φ[n] = π2 (n − n0 ) and hence ω[n] = π2 which 0 )) 2 corresponds to half of the Nyquist frequency. 7: The TFE analysis of signal x(t) given by Eq. By using the relation z[n] = 1 π π 0 X(ω) exp(jωn) dω.

the fundamental frequency (inverse of glottal pulse or pitch period length) of the voice signal. Application to a speech signal analysis To illustrate the advantages of FDM in speech signal analysis. between modes three and four. the compression waves of small amplitude but higher frequency range of 10 to 20Hz. the EEMD’s is able to catch F0 in mode y5 . 50 0 Power/Frequency (dB/Hz) −50 −100 arrive from the epical center to the recording station. There is enhanced TFE tracking by FDM methods as it provide better details of how the different waves H. there is enhanced TFE tracking when using FDM. and all three methods indicate maximum energy concentration around 1. Besides this problem. The Elcentro Earthquake data (sampled at Fs = 50Hz) has been taken from [25] and is shown in Figure 13. 9: The PSD analysis of the Gaussian white noise (with zero mean and unit variance): FDM (top and middle) and EMD (bottom). n0 = 199. et al. sampling frequency Fs = 100 Hz. EMD and EEMD algorithms for same data (vowel ‘small-cap I’). C ONCLUSION In this paper. we decompose a quasiperiodic voice signal (small-cap I vowel). EMD and CWT methods are shown in Figure 15. the EMD is not able to catch F0 in any mode. length N = 400): Real part of z[n] (top). the major drawbacks in this case are huge computation complexity for this ensemble size and the reconstruction error that may be significant. The TFE distribution by the FDM. e. VI. The FDM is a generalized Fourier expansion with variable amplitudes and frequencies of a time series by the Fourier .5 0 0 −100 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 1 Imag{z[n]} Power/Frequency (dB/Hz) −50 −150 −200 0 −1 0 −250 1 |z[n]|=a[n] −300 −350 −400 0 5 10 15 20 25 30 Frequency (Hz) 35 40 45 50 FDM: Power spectral density of FIBFs (HTL−FS) 0. G. are similar in nature. Moreover. marginal spectrum by FDM and marginal spectrum by EMD in Figure 13 show that almost all the energy in this data is within 10Hz. and other body shear waves which are present over the full duration of the data span. Also. The instantaneous energy fluctuations by the FDM and EMD methods. The results in this application as well are in clear favor of the Fourier method proposed in this paper.SINGH P. Clearly. as shown in Figure 14. while y6 to y7 seem to express most of the signal information at high frequencies. The computation complexity can be reduced by selecting the smaller ensemble size but that increases the reconstruction error.: THE FOURIER DECOMPOSITION METHOD 9 FDM: Power spectral density of FIBFs (LTH−FS) 1 Re{z[n]} 50 0 0.g. −150 −200 −250 −300 −350 −400 0 5 10 15 20 25 30 Frequency (Hz) 35 40 45 50 EMD: Power spectral density of IMFs 50 Power/Frequency (dB/Hz) 0 −50 −100 −150 −200 −250 0 5 10 15 20 25 30 Frequency (Hz) 35 40 45 50 Fig. In contrast. which decomposes any data into a small number of ‘Fourier intrinsic band functions’ (FIBFs). the shear and surface waves of strongest amplitude and lower frequency range of below 5Hz which does most of the damage. although it can be observed that positive minima and negative maxima in the mode y4 and y7 . the FIBF y8 captures accurately F 0. The critical frequency range that matter in the structural design is less than 10Hz. and between modes four and five. EMD (Figure 17 (a)) presents mode mixing between modes two and three. In Figure 16. the FIBFs y2 to y5 seem to extract most of the noise.5 0 0 Fig. Figure 18 shows the plots of TFE obtained by the FDM. 10: The analytic representation of δ[n − n0 ] (with. Imaginary part of z[n] (middle) and absolute value of z[n] (bottom). In Figure 17 (b). we have proposed: (1) The Fourier Decomposition Method (FDM). for nonlinear and nonstationary time series analysis.7Hz and 2 second. and from the Fourier based power spectral density (PSD). there is clear violation to IMF condition (1). Application to a earthquake signal analysis Earthquake time series signal is nonlinear and nonstationary data.

02 Amplitude Frequency (Hz) 20 2 1 0 −1 0 −3 x 10 5 0 −5 0 Fig.5 2 Time (s) 2. length N = 400).5 1 1.5 3 3. n0 = 199. detrend uncertainty and end effect artefacts as extraction of FIBFs does not depend on distribution of local extrema across the range of data. (2) The zero-phase filter bank based multivariate FDM (MFDM) algorithm.. sampling frequency Fs = 100 Hz. then the Fourier transform (FT) of x(t) is defined as Z ∞ X(f ) = x(t) exp(−j2πf t) dt (30) 20 15 10 5 0 0 0. which is generating finite number of band limited multivariate FIBFs (MFIBFs).e.5 2 Time (s) 2.5 EEMD: Time−Frequncy−Energy estimate of IMFs 0. (30) X(−f ) = X ∗ (f ) and. As x(t) is a real function. (31) as Z ∞ [X(f ) exp(j2πf t) + X ∗ (f ) exp(−j2πf t)] df x(t) = 0 (32) In this Eq.5 1 1. (3) An algorithm to obtain cutoff frequencies required in MFDM algorithm for zero-phase high or low pass filtering of multivariate signals. FIBF1. From Eq. The proposed methods produce the final presentation of the results in a time-frequency-energy distribution that reveals the imbedded structures of a signal.5 4 Time−Frequncy−Energy estimate by CWT 50 45 physical phenomena and engineering systems.(33) as Z ∞ M X ai (t) exp(jφi (t)) (34) 2 X(f ) exp(j2πf t) df = 0 i=1 . We write Eq.5 1 1. The fundamental and conceptual contributions of the this study are the Fourier based decomposition method (i. FIBF2 and highest frequency component plots by the FDM of unit sample sequence δ[n − n0 ] (with.10 SIGNAL/DATA ANALYSIS & PROCESSING Amplitude FDM: Time−Frequncy−Energy estimate of FIBFs (LTH−FS) 30 Amplitude 25 10 0.5 3 3. second term is complex conjugate of first term. we rewrite Eq.5 1 1.5 4 0.5 4 0.5 3 3. Simulation results demonstrate the efficacy of the proposed methods. local and adaptive.5 3 3. 12: The DC.5 1 1. sampling frequency Fs = 100 Hz.5 3 Amplitude 5 3.5 4 0 −0. for the analysis of multivariate nonlinear and nonstationary time series. hence.5 1 1. The FIBFs form the basis of the decomposition that are complete. method itself. EEMD (middle) and CWT (bottom). we can write x(t) = Re{z(t)} (33) R∞ where analytic function z(t) = 2 0 X(f ) exp(j2πf t) df and Re{z(t)} denote real part of function z(t). the FDM and MFDM are mathematically well defined. FDM) and the introduction of the FIBFs.5 3 −∞ 3.5 2 Time (s) 2.5 2 Time (s) 2.02 0 15 0 0 −2 0 0.5 4 0. n0 = 199.5 2 Time (s) 2. length N = 400): FDM (top). The FDM for continuous time real function 25 Let x(t) be a non-periodic. real function of time and follow the Dirichlet conditions. Pseudo−Frequency (Hz) 40 A PPENDIX 35 30 A. 40 35 Frequency (Hz) 30 25 20 15 10 5 0 0 0.5 1 1. A time-frequency-energy distribution of a signal is used in various fields of science and engineering for analysis of and inverse Fourier transform is defined as Z ∞ x(t) = X(f ) exp(j2πf t) df (31) −∞ R0 It R ∞is easy to show that −∞ X(f ) exp(j2πf t) df = 0 X(−f ) exp(−j2πf t) df .5 Fig. The instantaneous frequencies of the FIBFs produce a time-frequency-energy distribution of any signal. supported by the well established theories of filter and the Fourier transforms.5 2 Time (s) 2. 11: The TFE analysis of unit sample sequence δ[n − n0 ] (with. orthogonal. Unlike the various EMD algorithms.5 2 Time (s) 2. The FDM and MFDM methods do not suffer from mode mixing.5 3 3.

3 0 1. for each i. fM = 0. North−South Component E(t) by FDM 0.7 0.4 0. M .4 0.2 0. (36) ai (t). start with fi−1 . the lower and upper limits of integration in (35) will change to fi to fi−1 .1 0 0.: THE FOURIER DECOMPOSITION METHOD 11 Elcentro Earthquake May 18.2 0.6 0. (36) is satisfied.4 0.4 Ez . (d) Marginal spectrum by EMD. fM = ∞) Z fi X(f ) exp(j2πf t) df.6 Energy Density Spectral Density 0.8 0.3 0. The GD measures the average time arrival of the frequency f .5 0. (b) Fourier based power spectral density (PSD).8 0. with f0 = ∞.2 0.1 0 0 5 10 15 20 25 Frequency (Hz) Marginal spectrum by FDM 0 0 10 20 30 Time (s) Fig. the local time behavior as a function of frequency is characterized by the group delay (GD) : ti (f ) = 1 dφi (f ) − 2π df .35 Energy Density Acceleration (G) 0.3 0.SINGH P. 13: Plot of (top to bottom): (a) Elcentro Earthquake May 18.9 0. for causal signal x(t). · · · . The IF characterizes a local frequency behavior as a function of time.2 0. and we can obtain minimum number of AFIBFs by selecting the minimum value of fi such that 0 ≤ fi ≤ fi−1 and Eq. it is easy to obtain relationship between the energy of original signal x(t) and energy of its analytic signal z(t) as 5 10 15 20 25 Frequency (Hz) Fig.2 0.2 x 10 10 20 30 Time (s) 40 50 0 0 60 10 20 Power spectral density −4 30 Time (s) 40 50 60 40 50 60 E(t) by EMD 0. fi (t) = 2π dt Similarly.05 −0. (38) ai (f ) exp(−jφi (f )) = Ex = 0. the energy of analytic signal is twice of the energy of original signal. 1940 North-South Component. 1.2 0.4 0. (c) Marginal spectrum by FDM. 1940.2 0 0 5 10 15 20 25 Frequency (Hz) Marginal spectrum by EMD 0. 14: Plot of the instantaneous energy density E(t) (top to bottom) using the: (a) FDM and (b) EMD.15 −0. (32) and (33). where (with f0 = 0.4 0. From Eq.6 Energy Density 0.7 0. In general.25 0. (37) 2 that is. et al. To obtain minimum number of AFIBFs in low to high frequency scan. ai (t) exp(jφi (t)) = 2 fi−1 (35) ti−1 . Similar to the IF process. in high to low frequency scan. increase and select the maximum value of fi such that fi−1 ≤ fi ≤ ∞ and 1 dφi (t) ≥ 0.3 0. In a dual way.4 1.5 0. respectively.6 0.5 0.8 1 0.1 −0.1 0. we obtain Z ti x(t) exp(−j2πf t) dt.2 Spectral Density 1 0.1 0 0 for i = 1.3 0.45 0. ∀t. the IF and GD define two different curves in the time-frequency plane.

The fundamental frequency F0 is captured accurately in FIBF y8 .15 0.2 0.2 y4 −0.1 0. and the EEMD is able to capture F0 accurately in IMF y5 .02 0.1 0.1 0. 16: Vowel ‘small-cap I’ sound with sampling frequency Fs = 16000 Hz (top left) and its Fourier spectrum (top right).2 0. (39) X(−ω) = X (ω) and.05 0. hence.1 4 −0.5 0 6000 −0. The FDM generates highest frequency component (y1 ).02 y6 −0.008 Fig.01 y 25 0 0 −1 5.316417 −0. As x[n] is real function.01 0.02 0 0 x 10 −0.1 0. real function of time and follow the Dirichlet conditions.5 y2 y 0 0 3 −0.05 0. 1 dφi (f ) − 2π df y ≥ Let x[n] be a non-periodic.1 0.5 0. (40) as Z π Z π 1 [ X(ω) exp(jωn) dω+ X ∗ (ω) exp(−jωn) dω].05 0 −0.35 0 0 0. we rewrite Eq.05 2 −0.5 10 0 y Frequency (Hz) −2 Fig.. second term is complex conjugate of first term.46746 x 10 −5 12 1 15.2 standard deviation of the added noise of normal distribution and the ensemble size of 300) algorithms for vowel ‘small-cap I’ with Fs = 16000 Hz.2 0. −3 2 0 −5 with t0 = 0. tM = ∞ such that ai (f ).05 10 15 20 25 30 Time (s) 35 40 45 50 −0.5 5 3000 0 −0.1 15 20 25 30 Seconds 35 40 45 50 −5 5 y 9 Fig. The EMD is not able to catch F0 in any mode.1 10 2000 0.02 −5 0.05 (39) −∞ and inverse discrete time Fourier transform (IDTFT) is defined as Z π 1 X(ω) exp(jωn) dω (40) x[n] = 2π −π R0 It R π is easy to show that −π X(ω) exp(jωn) dω ∗ = 0 X(−ω) exp(−jωn) dω.02 −3 8 y7 5 x 10 y 0.1 |X(f)| FDM: Time−Frequncy−Energy estimate of FIBFs (HTL−FS) 0 I 25 1 0 y 0.58531 0 0 2. ti (f ) = 0.05 0.5 0. then the discrete time Fourier transform (DTFT) of x[n] is defined as ∞ X x[n] exp(−jωn) −5 x 10 −3 0 −2 0.5 5 Pseudo−Frequency (Hz) −4 0 y 20 11 15 x 10 2 10 y9 0 y x 10 0 y 10 8000 −0.01 −3 −3 1 10 0 y y 9 Scalogram Percentage of energy for each wavelet coefficient −5 11 y 9.1 0.82613 (b) EEMD 0.835538 −0.119827 0 0 y y 3 1. M .02 0.05 y2 y1 3.514178 0 −0.02 6 −0.2 0.02 0. for i = 1.05 0.012 −4 12 0 −3 0.1 −2 0 5000 0 y y 5 y y 7 Frequency (Hz) 0 2 4000 −0.05 15 1000 Frequency (Hz) Time (s) y 20 0.2 −0.5 0 −7 x 10 5 0. FIBFs (y2 to y11 ) and DC term (y12 ). 20 0 5 y 0 0 y 5 −0.25 0. The FDM for discrete time real function X(ω) = x 10 y 11 5 x 10 0 −5 x 10 10 10 5 0 y 5 0 −0.01 8 7 −0.02 5 0 y 0 y 5 −0. ∀f .194718 0.35775 0. x[n] = 2π 0 0 (41) In this Eq. B.3 0.12 SIGNAL/DATA ANALYSIS & PROCESSING 0.20634 0 y 0.4 3 35 40 45 4 y6 y −2 2 x 10 2 50 12 25 30 Time (s) 0 0 −2 EMD: Time−Frequncy−Energy estimate of IMFs 15 (a) EMD 1 0. 17: IMFs (y1 to y12 ) generated by (a) EMD and (b) EEMD (with zero mean and 0. · · · . we can write x[n] = Re{z[n]} (42) . From Eq.5 0.2 0.1 −3 −4 5 7000 0 8 0 −0.3846 x 10 0 −1 0.5 0. 15: Plot of TFE (top to bottom) using the: (a) FDM (b) EMD and (c) CWT.

” Resuscitation. Joshi S. [18] Daubechies I.” 2013 International Conference on Signal Processing and Communication. ∀n. respectively. Lipsitz L.1 0.. ai [n]. (2013) November.. Liu Y.. A. Rehman N.” Proc. π ωi−1 (44) for i = 1. Chiang W.J. and Saha K... Priplata A.. 204–237. [13] Singh P. (2008) 46...L. and Peng C. Patney R..2 Time (s) 0. · · · . [3] Cummings D.IT]. Soc.S. Lu J...B. (2014) 244. ωM = 0.05 0.. (2004) 427. “The Hilbert spectrum and the Energy Preserving Empirical Mode Decomposition.K.15 0.” Applied and Computational Harmonic Analysis.P.25 0.25 0.04104v1 [cs.05 0. and Goncalves P.2 Time (s) 0. “Compact Empirical Mode Decomposition: An Algorithm to reduce mode mixing. Wu Z. Geophys.. (2010) 466. “Evolutionary spectra and non-stationary processes.15 0. [10] Chu P. 809–821.E.” Proc. Ko P.E.e. Rπ where analytic signal z[n] = π1 0 X(ω) exp(jωn) dω and Re{z[n]} denote real part of function z[n]. Chen W. (2009) 1.E.. Long S. (2012) 4 (3).H.SINGH P. “The multi-dimensional ensemble empirical mode decomposition method. Wang C. 2279– 2292.K.15 0. (2011) 30. 18: The TFE plot by FDM (top) EMD (middle) and (b) EEMD of same data (vowel ‘small-cap I’). “Altered phase interactions between spontaneous blood pressure and flow fluctuations in type 2 diabetes mellitus: nonlinear assessment of cerebral autoregulation. “Nonpolynomial Spline Based Empirical Mode Decomposition.... [14] Mandic D.C..L. and Chen X.W. increase and select the maximum value of ωi such that ωi−1 ≤ ωi ≤ π and phase φi [n] is a monotonically increasing [1] Huang N.A. [2] Costa M.... 68008.. B. 344–347. Data Anal. (45) is satisfied.A.” Europhys. Soc.. “Synchrosqueezed wavelet transforms: An empirical mode decomposition-like tool.” Rev. [17] Priestley M. and Mandic D. [9] Rehman N. Endy T. and Huang N..3 0. (2013) 435-440. A.C.” J. 339–372. 1291–1302... “Noise and poise: enhancement of postural complexity in the elderly with a stochastic-resonance-based therapy. Huang N.. ur and Mandic D. R.P. 112–114..” Adv. Adapt.A.: THE FOURIER DECOMPOSITION METHOD 13 FDM: Time−Frequncy−Energy estimate of FIBFs (HTL−FS) function.E. (2004) 11 (2). Lin J.H. (2007) EPL 77.. Lett. and Huang N.Y. end effect. Tung C. 1250017.. Joshi S... “A review on Hilbert–Huang transform: Method and its applications to geophysical studies. [4] Hu K. [8] Wu Z. [12] Singh P. and we can obtain minimum number of AFIBFs by selecting the minimum value of ωi such that 0 ≤ ωi ≤ ωi−1 and Eq.K.. and Ma M. 2421–2426. Wu Z. Huang N. with ω0 = π.” Advances in Adaptive Data Analysis. the lower and upper limits of integration in Eq. Wu M. Zheng Q. Rilling G. RG2006. Patney R.” arXiv:1504.. (1988) 454. Goldberger A.. Lett. Ungchusak K... R.. Cavallerano J. (2011) 59 (5).. Wu Z. Hsieh W.35 EEMD: Time−Frequncy−Energy estimate of IMFs 7000 6000 Frequency (Hz) 5000 4000 3000 2000 1000 0 0 0.1 0..D... 243–261. Srivastavay P. and Wu H. “Multivariate empirical mode decomposition.H. [11] Singh P..05 0. Lipsitz L.. start with ωi−1 .I..E. In order to obtain minimum number of AFIBFs in low to high frequency scan. et al. P.35 0. Nisalak A. “Empirical Mode Decomposition-Based Time-Frequency Analysis of Multivariate Signals...E..” Nature. “Empirical mode decomposition as a filter bank. [16] Rehman N. i. 903–995. (2013) 84 (11). “A new method to estimate the amplitude spectrum analysis of ventricular fibrillation during cardiopulmonary resuscitation. Huang N. [5] Lo M. (2009) 1 (1). R...K. Chang Y... and Burke D.” IEEE Transactions on Signal Processing. in high to low frequency scan. 74–86. ... “Filter Bank Property of Multivariate Empirical Mode Decomposition. ωM = π) Z 1 ωi ai [n] exp(jφi [n]) = X(ω) exp(jωn) dω. [15] Flandrin P. Young V.K.D.C.. and Saha K... Fan C. Yen N.” Applied Mathematics and Computation.D.” IEEE signal processing magazine.Y.. Patney R.... and Liu H.. Data Analys..E. 1–41..1 0. 6000 Similarly. 1505–1511.” IEEE Signal Process.. and Huang N. (1965) B27.25 Fig. P. Peng C.A. ωi [n] = (φi [n + 1] − φi [n]) ≥ 0. (44) will change to ωi to ωi−1 ..T. Lin L. “Ensemble Empirical Mode Decomposition: a noise-assisted data analysis method. “The empirical mode decomposition and Hilbert spectrum for non-linear and non-stationary time series analysis.. [6] Huang N... Shen Z. Adapt.. Soc. Irizarry R. Lin C. (42) as 1 π Z π X(ω) exp(jωn) dω = 0 M X ai [n] exp(jφi [n]) (43) i=1 where (with ω0 = 0.. “Travelling waves in the occurrence of dengue haemorrhagic fever in Thailand..M.. M .. and detrend uncertainty. for each i. 5000 Frequency (Hz) (45) 4000 3000 2000 ACKNOWLEDGMENT The authors would like to thank JIIT Noida for permitting to carry out research at IIT Delhi.. Norden E. Joshi S. Statist.35 EMD: Time−Frequncy−Energy estimate of IMFs R EFERENCES 7000 6000 Frequency (Hz) 5000 4000 3000 2000 1000 0 0 0. 1000 0 0 0..3 0. We write Eq.. and Novaka V.” Adv.3 0. [7] Wu Z.H. and Wu Z.K. and Saha K. “Some studies on nonpolynomial interpolation and error analysis..” Physica A (2008) 387 (10).2 Time (s) 0. C.A. Shih H. There is enhanced TFE tracking when using FDM.

.com/elcentro.commsp. 1–23... “A product threom for Hilbert transform. (2008) 46 (02). of IEEE. Available: http://www.. “Estimating and interpreting the instantaneous frequency of a signal.. 63–93.” Advances in Adaptive Data Analysis. H..htm [23] Thomas Y. [24] Boashash B.” Proc. (1992) 80 (4).vibrationdata.edu. and Zuoqiang S. [20] Nuttall A. (2011) 63 (03).. “A review on Hilbert-Huang transform: Method and its applications to geophysical studies. [27] Huang N.tw/research1.” Reviews of Geophysics.. Adapt.” Adv. Data Anal. “On Hilbert Spectral Representation: a True Time-Frequency Representation for nonlinear and nonstationary Data. E. Chen X. and Wu Z.htm [22] Available: http://rcada.ac.htm [26] Huang N. and Wu Z.14 [19] Bedrosian E. and Bedrosian E. “On the quadrature approximation to the Hilbert transform of modulated signals.” Proc.ee.ncu.. 1458–1459. 520–538. IEEE.” Proc. “Adaptive data analysis via sparse timefrequency representation. Fundamentals. Lo M.. (2011) 3 (1 & 2). (1966) 54 (10). T. [21] Available: http://www. SIGNAL/DATA ANALYSIS & PROCESSING . [25] [Online]. . I. H. 868–869. 1–28.uk/∼mandic/research/emd.ic. E. of IEEE. (1963) 51 (5).