Time-frequency analysis of the dynamics of different vorticity structures generated from a finite-length triangular prism

Transcription

1 Time-frequency analysis of the dynamics of different vorticity structures generated from a finite-length triangular prism Giacomo Valerio Iungo a, Edoardo Lombardi b a Dipartimento di Ingegneria Aerospaziale, Università di Pisa b Optimad engineering s.r.l. G. Collegno, , Torino SOMMARIO In questo lavoro è stata eseguita l analisi tempo-frequenza delle dinamiche di differenti strutture vorticose di scia mediante una nuova tecnica basata sulla Proper Orthogonal Decomposition. A tal fine sono state analizzate misure di anemometria a filo caldo eseguite in prossimità della scia prodotta da un prisma triangolare, posto con uno spigolo verticale opposto alla direzione del vento. Da questi dati sperimentali sono state individuate ed estratte tre differenti componenti spettrali, connesse con le dinamiche di differenti strutture vorticose. La tecnica utilizzata per l analisi tempofrequenza è utilizzabile genericamente per qualsiasi tipo di segnale, anche non-stazionario, in quanto produce una base direttamente dipendente dai dati analizzati e non verte su alcuna ipotesi. Le componenti principali estratte dai segnali sono ordinate in funzione della loro energia fluttuante, e mediante l utilizzo della trasformata di Hilbert è possibile effettuare la caratterizzazione temporale delle loro rispettive ampiezze e frequenze. ABSTRACT In the present paper time-frequency analysis of the dynamics of different wake structures is carried out through a novel procedure based on Proper Orthogonal Decomposition. The analyzed data consist in hot-wire anemometer signals acquired in proximity of the wake generated from a triangular prism orientated with its apex edge against the incoming wind. From these experimental signals three different spectral components, connected to the dynamics of different vorticity structures, are detected and the time histories of their amplitudes and frequencies are characterized. The technique used for time-frequency analysis is adaptive, completely data-dependent with no a- priori assumption and applicable to non-stationary signals. The principal components are extracted from the signals and sorted by their fluctuating energy; moreover, the time-variation of their amplitude and frequency is characterized through the Hilbert transform. Keywords: Triangular prisms, Time-frequency analysis, Proper Orthogonal Decomposition. 1. INTRODUCTION Fluid dynamic signals are generally characterized by significant fluctuations that must be characterized in order to investigate on their physical origins. In several conditions, as for instance in wakes and jets, the flow fluctuations may show dominating spectral components, which can be singled out through the conventional Fourier transform. However, this technique gives only a timeinvariant amplitude and frequency for each spectral component and thus becomes highly inappropriate for non-stationary signals. By considering for instance fixed-point measurements carried out in proximity of a wake, shedding of vorticity structures with different intensities and a varying distance from the measurement point can produce amplitude and/or frequency modulations in the acquired signal (AM-FM signal). Therefore, time-analysis of the amplitude and frequency of each spectral component is needed in order to investigate on the corresponding fluid dynamic phenomena. This procedure is typically

2 denoted as demodulation and the most popular demodulation technique is certainly the Hilbert transform (HT in the following), see e.g. Bendat & Piersol, However, all demodulation techniques can only be used for time-frequency analysis of AM-FM mono-component time-series, i.e. a component consisting of only one carrier with its amplitude and frequency modulations. Consequently, the different AM-FM mono-components present in a multicomponent signal must be singled out and separated before performing their demodulation. For instance Sreenivasan, 1985 carried out the mono-component extraction through ad-hoc band-pass filters; however, this technique is non-adaptive requiring a filtering customization for each signal. Alternatively, the wavelet transform may be directly applied to multi-component signals in order to characterize the time-variation of the components contributing to the signal fluctuations present in a certain frequency range. Component extraction may be performed by using, for instance, the socalled wavelet ridges, as reported in Carmona et al., 1998, or the technique proposed in Buresti et al., A more sophisticated wavelet decomposition is proposed by Olhede & Walden (2004), which enables spectral components to be extracted with ad-hoc filter banks, but with band-edge imperfections in correspondence of the boundary of contiguous frequency sub-bands for components spanning at least two sub-bands, and by almost losing the nonlinear characteristics of the signal. Summarizing, the wavelet based decomposition techniques are generally non adaptive, because the filter bank must be selected properly for the considered signal by detecting the dominant spectral frequencies and choosing filter amplitudes by-eye or through a statistical control. Doubtless the technique for component detection and extraction from multi-component signals that has earned more interest in the last decade is the Empirical Mode Decomposition (EMD) proposed by Huang et al.,1998. EMD is an empirical technique producing a completely data dependent (aposteriori) basis, in which each component is defined as an intrinsic mode function (IMF) that satisfies two conditions: the number of extrema and zero crossings of the IMF must be equal or differ at most by one and the IMF is zero mean locally. The primary limitation of the application of EMD for fluid dynamic time-series is its fixed frequency resolution, which depends only on the sampling frequency and on the number of samples of the analyzed signal. In fact, EMD is basically a bank of dyadic filters. A dyadic filter bank is a collection of band-pass filters that have an analogous shape but with neighboring filters covering half or double of the frequency range of any single filter in the bank. In the present paper a novel procedure for time-frequency analysis of fluid dynamic time-series is presented, which is based on the Proper Orthogonal Decomposition (POD in the following). The source-signal is first manipulated in order to generate an ad-hoc data set consisting of time-portions of the signal, each one composed of the same number of samples; this data set represents the input for the POD procedure. The result obtained from the POD is an orthogonal basis whose elements, denoted as POD modes, are sorted by their energy. The POD modes corresponding to higher energy represent the most significant signal fluctuations, and their Fourier analysis already provides spectral information related to each POD mode. However, and more importantly, the AM-FM mono-components associated with each POD mode can be obtained through a procedure based on the convolution of the source signal with the relevant POD mode, suitably manipulated. The POD procedure is extremely adaptive and the only significant input is the chosen frequency resolution. The calculated basis is completely a-posteriori and data dependent, orthogonal, and enables a complete and unique decomposition of the signal. Furthermore, the main advantage of this new technique is represented by its optimality in the detection and sorting of the principal components present in a fluid dynamic signal as a function of their energy; therefore, an iterative procedure can be implemented for extracting spectral components, starting from the most energetic one and then analysing the residual signal, i.e. the source-signal from which the extracted spectral component is subtracted. In this way the flow fluctuations can be considered adequately characterized when a sufficient fluctuating energy is extracted (e.g. a certain percentage of the total fluctuating energy) or when all the dominant spectral components are captured. Furthermore, the different extracted components can be grouped as a function of their mean IF, in order to produce AM-FM mono-components. The present paper is organized as follows: the POD procedure for time-frequency analysis of fluid dynamic time-series is described in Sect. 2, while it is then applied to hot-wire anemometer signals acquired in proximity of the wake generated from a triangular prism (Sect. 3). Finally, conclusions are reported in Sect. 4.

3 2. POD PROCEDURE FOR TIME-FREQUENCY ANALYSIS OF TIME-SERIES Proper orthogonal decomposition, denoted by the acronym POD, is a method providing an orthogonal basis for the modal decomposition of an ensemble of data functions. Therefore, a zeromean random process, u(x, t), where x is a space-vector x=x 1,,x N and t is time, may be represented through a linear combination of deterministic functions, the POD modes, φ j (x): u M ( x t) = ( x) a ( t), ϕ (1) j= 1 j j where φ j (x) are the eigenfunctions of the covariance function of the analyzed process and represent the typical realizations of the process in a statistical sense, while a j (t) are uncorrelated coefficients denoted as principal components. It should be pointed out that POD provides a modal decomposition that is completely data dependent, a-posteriori, adaptive and does not assume any hypothesis on the process that generates the data. Therefore, it does not neglect the nonlinearities of the original dynamical system, even being a linear procedure. The POD basis is orthogonal and allows a unique and complete representation of the signal, as analytically reported in Eq. 1. The most peculiar feature of POD is optimality: among all linear decompositions it provides the most efficient detection, in a certain least squares optimal sense, of the dominant components and trends of an infinite-dimensional process. These dominant components represent the typical realizations of the most energetic phenomena of the physical process. Let us then consider a generic zero-mean time-series u(t), which can represent a measurement performed in a fixed point of the flow field with a sampling frequency F samp and a total number of samples equal to N samp. If the signal u(t) is a stationary process with infinite energy (i.e. it is not square-integrable since it does not vanish for t tending to infinite) POD modes are classic harmonic functions, as highlighted in Lumley (1970), and can be considered as generalized Fourier modes, but an advantage in using POD is still present since the modes are sorted by their energetic significance. However, u(t) may also be a non-stationary process and then the POD modes represent the most typical realizations of the process, sorted by their significance, and can be generic functions. In order to perform POD, a certain number of observations of the analysed process are required, the so-called snapshots; to this end, time-portions of the source-signal u(t) are generated, all with the same number of samples, N period. The source-signal is divided into adjacent time-portions in order to generate the snapshots, whose total number, N snap, depends on the total length of the source-signal, N samp, and on the time-length of each snapshot, N period. An adequately high number of snapshots is required to perform a satisfactory POD, i.e. in order to separate the typical realizations of the main phenomena from other random processes. Generally, the number of snapshots, N snap, is gradually increased through a sensitivity analysis, so that the convergence of the POD eigenvalues and of the most energetic POD modes is obtained. In case a non-sufficient number of snapshots are produced from adjacent time-portions of u(t), a higher number of snapshots can be generated through partially overlapped time-portions of the signal uniformly distributed along the time-length of the signal. Obviously this technique decreases the statistical independency of the process observations; however, this method turns out to be very useful to annihilate all random influences or disturbs present in a signal and better highlight all the typical realizations of the process. A crucial task of the procedure is represented by the choice of the time-length of each snapshot, i.e. the number of samples, N period, composing each snapshot, which is strictly related to the frequency resolution of the analysis; in other words, if Δf is the required frequency interval between two consecutive elements of the signal power spectrum, in order to discern different spectral components, then N period will be equal to the ratio between the sampling frequency and the frequency resolution, F samp /Δf. Therefore, the higher is the number of samples composing each snapshot, N period, the higher is the frequency resolution of the spectral analysis; however, with increasing frequency resolution the time-resolution of the POD tool is reduced as a consequence of the Heisenberg uncertainty principle.

4 Once the required number of samples for each snapshot, N period, is chosen and the total number of snapshots, N snap, is fixed, a matrix M is generated, whose rows are the time-series of the snapshots; so that the size of M is N snap x N period. Subsequently, the covariance matrix of M is evaluated as C = M T M. C is a square matrix of size N period and each element C ij represents the scalar product between the two respective columns of M, i.e. between the vectors composed of the values assumed by the samples with indexes i and j for the different snapshots. C is Hermitian symmetric and nonnegative definite, thus its eigenvalues are real and non-negative, the eigenvectors are orthogonal. Furthermore, the eigenvectors are normalized through the L 2 -norm so that the eigenvalues represent the energy associated to the respective eigenvectors. From a geometrical point of view, the eigenvectors φ j (x) represent the principal axes of C and are the so-called POD modes. Any snapshot, i.e. each row of the matrix M, analogously to what already reported in Eq. 1, is represented through a linear combination of the POD modes, φ( j): M i = N period j= 1 ( j) a ( j) ϕ (2) i where a i (j) are denoted as principal components, which are evaluated through the following projection: a i ( j) M, ϕ( j) = (3) i where < *,* > is the scalar product. Eq. 2 states the completeness and the uniqueness of the POD. The computer-generated signal used to describe the POD procedure for time-frequency analysis is a stationary time-series composed of three different spectral components ( f 1 = 40 Hz, f 2 = 60 Hz and f 3 = 70 Hz) and white noise with an energy equal to 23% of the total energy of the signal: ( 2 f t) + 2sin( 2πf t) + 4sin( πf t) WN y sin π + (4) 1 = The signal is sampled with a frequency of 1 khz and the POD procedure is applied by using for this test-case 10 4 snapshots, which represent adjacent time-portions of the signal, each comprising 501 samples, i.e. the used frequency resolution is about 2 Hz. The criteria for the evaluation of an adequately high number of snapshots will be discussed in the following. The POD eigenvalues reported in Fig. 1a, which represent the fluctuating energy of the respective POD modes, are reported as percentage of the total energy of the signal, and show that just the first six POD modes are characterized by a significant fluctuating energy and, thus, they are the only POD modes to be considered for the component extraction. The time-series of the first eight normalized POD modes are reported in Fig. 1b, while the respective Fourier power spectra are reported in Fig. 2. The most energetic POD modes, with the same energy, are POD modes 1 and 2, which are characterized by a dominant frequency of 70 Hz, as expected being the most energetic spectral component of y 1 characterized by this frequency. a) b) Figure 1. POD procedure applied to the computer-generated signal y 1 : a) POD eigenvalues; b) time-series of the first eight normalized POD modes.

5 From Fig. 2 it is seen that these two POD modes are characterized by the same power spectrum, but from Fig. 1b it can be observed that they have a relative phase-shift of 90, being elements of an orthogonal basis. The fact that two POD modes are ascribed to the same spectral component is physically interpretable by considering that each snapshot can be characterized by different projections onto these two different POD modes depending on the initial phase of the snapshot. The POD modes 3 and 4 are related to the component at frequency f 2 = 60 Hz, while 5 and 6 to the component at f 3 = 40 Hz; as expected they are sorted by their energy. The remaining POD modes do not show any dominant spectral component and are due to white noise, but they are not analyzed because, as shown in Fig. 1a, their relative energy is negligible. The convergence of the results obtained from the POD procedure with increasing number of snapshots, N snap, can be checked from the trend of the POD eigenvalues and from the Fourier power spectra of the most energetic POD modes. As shown in Fig. 3a, the POD eigenvalues reach asymptotic values for a number of snapshots higher than 100. Furthermore, by analysing the Fourier power spectra of the POD mode 6 (Fig. 3b), i.e. the POD mode of interest with the lowest energy and, thus, easily affected by white noise, for N snap < 50 this POD mode is basically a random function, whereas for a higher number of snapshots the frequency of interests f 1 = 40 Hz appears, although influenced by the other spectral contributions. With increasing N snap the spectral component of interest is isolated and the convergence of the POD mode is reached. From the power spectra of the POD modes the dominant frequencies of the signal are detected; however, their contribution along the time-length of the whole source-signal is not yet determined. The technique for principal component extraction from a source-signal is based on the convolution of the latter with the considered POD mode. However, a generic POD mode is not a symmetric filter due to its initial phase, as can be appreciated from Fig. 1b, and, thus, the convolution produces a certain phase-shift, Δφ, on the extracted principal component with respect to the source-signal. In the context of signal processing, it is common to avoid this phase-shift by using as filter the reversed function, denoted as POD flip-mode. This function is composed of the elements of the considered POD mode, but from the last one backwards to the first one. Consequently, the POD mode 1 POD mode 2 POD mode 3 POD mode 4 POD mode 5 POD mode 6 POD mode 7 POD mode 8 Figure 2. Fourier power spectra of the first eight normalized POD modes evaluated for the signal y 1. a) b) Figure 3. Convergence of the POD procedure applied to y 1 by increasing the number of snapshots, N snap : a) first eight POD eigenvalues; b) Fourier power spectrum of POD mode 6.

6 convolution of the source-signal with the POD flip-mode produces a phase-shift equal to -Δφ; therefore, performing a double convolution of the source-signal, i.e. with the POD mode and with its POD flip-mode, a phase-correction of the extracted principal component might be performed. However, a more robust procedure and with lower computational effort is based on performing a single convolution of the source signal with a function obtained from the convolution of a POD mode with its respective POD flip-mode. This function, denoted as POD conv-mode, is a symmetric filter and, thus, no phase-shift is produced through the convolution of the source-signal with the POD conv-mode. It must be pointed out that each POD conv-mode is composed of 2xN period -1 samples, being obtained through the convolution of discrete functions consisting of N period samples. An important step regarding principal component extraction through the convolution procedure consists in avoiding any amplification or damping. The basic idea of the filtering is that the convolution of a certain POD conv-mode with itself must produce the POD conv-mode without any amplification or damping. In order to reach this goal, the result of the convolution must be normalized through the norm of the convolution of the POD conv-mode with itself, and then multiplied with the norm of the POD conv-mode in order to restore its initial energy. Therefore, the result of the convolution must be multiplied by the factor K: K conv mode = (5) convolution( conv mode, conv mode) where * represents the L 1 -norm, i.e. the sum of absolute values of the elements, consistently with the precision adopted for the convolution algorithm. As regards the computer-generated signal y 1, the three different spectral components are now extracted through the convolution of y 1 with the respective POD modes. The spectral component related to the frequency f 1 = 70 Hz can be evaluated through the POD mode 1 and its POD convmode is calculated as reported in Fig. 4a, while its Fourier power spectrum multiplied by the factor K is compared to the one of the POD mode 1 in Fig. 4b. The extracted spectral components are reported in Fig. 5 with their moduli obtained through the HT. Starting from the most energetic spectral component, i.e. the one related to f 1 = 70 Hz, its amplitude is equal to 4 (Eq. 4) and through the spectral component extraction its mean value is found to be with a standard deviation of (about 2% of the mean value), demonstrating that the spectral contribution of interest is completely captured by this spectral component. The instantaneous frequency of the spectral component is also evaluated through the HT and a mean value of Hz and a standard deviation of (about 0.02% of the mean value) are found. Since the spectral component related to the POD mode 1 is extracted, a residual signal can be evaluated by subtracting the extracted spectral component from the source-signal y 1. If the spectral component related to the POD mode 2 is then extracted, it corresponds basically to a null signal, which can be added to the one extracted with the POD mode 1 or just neglected. This assesses that the spectral component of interest is already completely extracted by only using POD mode 1 and POD mode 2 is considered only as its coupled POD mode. Subsequently, the spectral component related to the POD mode 3 is extracted, i.e. the one corresponding to f 2 = 60 Hz. The time-series of this spectral component is reported in Fig. 5b with a) b) Figure 4. POD mode 1 of the signal y 1 : a) time-series of the POD conv-mode 1; b) Fourier power spectra of POD mode 1 and POD conv-mode 1.

7 Figure 5. Spectral components extracted from the source-signal y 1 : a) spectral component related to f 1 = 70 Hz; b) spectral component related to f 2 = 60 Hz; c) spectral component related to f 3 = 40 Hz; d) reconstructed signal. its modulus evaluated through the HT (mean value of 2 and standard deviation of 0.087). From the HT of the extracted component a mean value of the IF equal to Hz is found with a standard deviation equal to The last component related to the frequency f 3 = 40 Hz, is extracted by using the POD conv-mode 5. The extracted component is reported in Fig. 5c (modulus with a mean value of and standard deviation of 0.091). The mean IF evaluated through HT is equal to Hz with a standard deviation of Finally, by adding the three extracted spectral components together, it is seen in Fig. 5d that the reconstructed signal well reproduces the source-signal y 1, except for the white noise, which is removed. The test case of the computer-generated signal y 1 well highlights the advantages of the POD procedure for time-frequency analysis of time-series: the only parameter to be chosen is the frequency resolution of the spectral analysis, and thus the number of samples for each snapshot, N period. At variance with other techniques, no frequency identifications of the main spectral components, preliminary spectral analysis or band-pass filtering are required. Once the POD modes are calculated, the principal components can be extracted in an automatic way starting from the most energetic one, and the characterization of the signal can be considered adequately performed when a certain percentage of the fluctuating energy of the source-signal is extracted or when the dominant spectral contributions are captured. 3. APPLICATION OF THE POD PROCEDURE TO HOT-WIRE MEASUREMENTS The procedure based on the POD for component detection and extraction from time-series was first assessed through different computer generated signals, as e.g. AM-FM multi-component nonstationary signals (see Iungo & Lombardi, 2010). Now the procedure is applied to the case of hotwire anemometer signals acquired in proximity of the wake generated from a triangular prism.

8 An experimental investigation on the near-wake flow field generated from a prism with equilateral triangular cross-section, aspect ratio h/w = 3, where h is the height and w the base edge of the model, and orientated with its apex edge against the incoming wind was presented in Buresti & Iungo, A sketch of the experimental setup is reported in Fig. 6, where the used frame of reference is also reported. The tests were carried out at a Reynolds number, based on w, of 1.5x10 5. For this configuration flow fluctuations at three prevailing frequencies were singled out, with different relative intensities depending on the wake regions. In particular, the frequency connected with alternate vortex shedding from the vertical edges of the prism was found to dominate in the regions just outside the lateral boundary of the wake at a Strouhal number of about St = f w/u 0.16, where U is the freestream velocity. On the other hand, a lower frequency, at St 0.05, was found to prevail in the velocity fluctuations on the whole upper wake. Simultaneous measurements carried out over the wake of the prism at symmetrical locations with respect to the symmetry plane showed that these fluctuations correspond to a vertical, in-phase, oscillation of two counter-rotating axial vortices detaching from the front edges of the free-end. This finding was confirmed by the results of a LES simulation of the same flow configuration, described in Camarri et al., In Buresti & Iungo, 2009 wake velocity fluctuations were also observed at an intermediate frequency St 0.09, and were found to prevail in the symmetry plane. By using the evidence provided by the abovementioned LES simulation, by flow visualizations and by pressure measurements over the prism surface, it was suggested that they may be caused by a flag-like oscillation of the sheet of transversal vorticity shed from the rear edge of the body free-end, and approximately lying along the downstream boundary of the recirculation region in the central part of the near wake. The three above described frequencies may also be found together for signals acquired aside the wake. Therefore, even if the fluctuations at the various frequencies are produced by the dynamics of different vorticity structures, they all contribute to the global oscillation of the wake. We consider a velocity signal characterized by the presence of a single dominant spectral component, for instance the one connected to alternate vortex shedding at St This is the case of the signal acquired aside the wake in correspondence to the point x/w=4, y/w=1.5, z/h=0.3. The Fourier power spectrum of this signal is reported in Fig. 7a. These velocity signals are acquired with a sampling frequency of 2 khz and they consist of 2 16 samples. Consecutively to the detection of a dominant frequency for the considered signal, a first attempt to carry out a time-frequency analysis consists in applying the wavelet transform. The modulus of the wavelet coefficients, reported in Fig. 7c, confirms the presence of a dominant spectral component in proximity of St 0.16 and also enables to highlight its strong irregularity both in amplitude and frequency. The deriving wavelet spectrum is reported in Fig. 7b, which shows an advantage of the wavelet transform in providing smoother and more clear spectra than the ones obtained from the Fourier transform. Let us now apply the POD procedure. From the Fourier power spectrum of the considered signal only one dominant spectral component has been observed and, thus, a high frequency resolution is not required for its time-frequency analysis. Consequently, the POD procedure is applied with a high time resolution, i.e. by using a frequency resolution of about 20 Hz (each snapshot is composed of 101 samples). From the signal 10 4 snapshots are generated, which is a sufficiently high number guaranteeing the convergence of the POD eigenvalues, but still requiring a small time for the calculation. The POD eigenvalues, which represent the energy associated with the respective Figure 6. Sketch of the experimental setup: a) model orientation; b) test layout.

9 POD modes, are reported in Fig. 8. From this figure it can be observed that with only the first two POD modes roughly the 35% of the fluctuating energy of the signal is captured, and with the first four the 50%, confirming the optimality of the POD to easily detect the core of the fluctuations of a signal without investigating on the remaining part. In Fig. 9a the Fourier power spectrum related to the component extracted through the POD mode 1 is shown, and from the one of the residual signal (Fig. 9b), i.e. the original signal from which the component is subtracted, it is seen that it represents the main part of the dominant spectral component. We would emphasize the very good result that is obtained with the POD procedure: even if a preliminary spectral analysis is not performed, the first information gained from this analysis is that a dominant spectral component with a frequency about St 0.16 and energy equal to 35% of the total energy of the signal is present. Furthermore, from the related Hilbert spectrum reported in Fig. 9c (which is a 3D plot where the x-axis represents time, y-axis the frequency and grey level the amplitude of the extracted components), a complete time-frequency analysis of this component can be performed. The time-frequency analysis of the hot-wire anemometer signal acquired at x/w=4, y/w=1 and z/h=0.9 is also performed. The Fourier power spectrum of this signal, reported in Fig. 10a, shows that a large amount of the energy of this signal is included in a frequency range between St 0.03 and St 0.2; several spectral components seem to be present but their detection is not sufficiently clear due to their comparable energy and limited spectral separation. From the analysis of the map of the modulus of the wavelet coefficients, calculated through a Morlet function with an enhanced time-resolution, no further information is obtained (Fig. 10c). However, by using an increased frequency resolution, the wavelet spectrum in Fig. 10b highlights the presence of three dominant spectral contributions at St 0.05, St 0.09 and St As observed from the wavelet map, the extraction of these components through band-pass filtering is very challenging because they are highly modulated and spectrally close. Before performing the spectral decomposition, the signal is filtered through a high-pass filter with a cut-off frequency St = 0.03, in order to remove the typical low frequency flow fluctuations present in the wind tunnel freestream. The POD procedure is applied by using a number of samples for each Figure 7. Signal acquired at x/w=4, y/w=1.5, z/h=0.3: a) Fourier power spectrum; b) wavelet spectrum; c) map of the modulus of the wavelet coefficients. Figure 8. POD eigenvalues evaluated for the signal acquired at x/w=4, y/w=1.5, z/h=0.3.

10 a) b) c) Figure 9. Component extraction with POD mode 1 from the signal acquired at x/w=4, y/w=1.5, z/h=0.3: a) component Fourier power spectrum; b) Fourier power spectrum of the residual signal; c) Hilbert spectrum of the component. a) b) c) Figure 10. Signal acquired at x/w=4, y/w=1, z/h=0.9: a) Fourier power spectrum; b) wavelet spectrum; c) map of the modulus of the wavelet coefficients. snapshot equal to 1001, i.e. a frequency resolution of about 2 Hz; a total number of snapshots equal to 10 4 is produced. The obtained first 50 POD eigenvalues are reported in Fig. 11. Starting with the extraction of the spectral contribution related to the most energetic POD mode, viz. POD mode 1, it is seen from its Fourier power spectrum, reported in Fig. 12, that it represents a narrow-band signal at St As suggested in Buresti & Iungo, 2009, this spectral contribution is connected to the dynamics of a couple of axial vortices detaching over the model free-end and, indeed, being this velocity signal acquired at a relative high position, this phenomenon turns out to be the most energetic one. The POD mode 2 is the one coupled to the POD mode 1, i.e. characterized by roughly the same power spectrum but with a phase shift of 90. The spectral contribution extracted with the POD mode 2 from the residual signal, obtained after the extraction of the contribution due to the POD mode 1, represents a residue of the contribution previously extracted with the POD mode 1, as shown in Fig. 12. Moving to the extraction of the spectral contribution connected to the POD mode 3, the corresponding power spectrum shows that it is clearly due to the alternate vortex shedding, being characterized by a mean Strouhal number of St 0.16, while its coupled POD mode is the 4. The following analyzed POD mode, POD mode 5, represents a further contribution to the spectral component due to alternate vortex shedding, i.e. its AM and/or FM, being characterized by a mean IF of St The POD mode 6 is its coupled one.

11 Interestingly, the spectral contribution related to the POD mode 7 is characterized by St 0.09, indicating that it is connected to the oscillations of the shear layer bounding the recirculation area located just behind the model. POD mode 8 is its coupled one. Subsequently, the extracted spectral contributions are grouped by their mean IF in order to obtain AM-FM mono-component signals. Therefore, the first mono-component is obtained by adding the contributions due to the POD mode 1 and 2, the second with the POD modes from 3 to 6 and the last one with the POD modes 7 and 8. Once these three AM-FM mono-components are evaluated, their related Hilbert spectrum, reported in Fig. 13, enables a simultaneously time-frequency analysis of the three different spectral components. Finally, in Fig. 14 the reconstructed signal obtained by adding the three AM-FM mono-components is compared to the source-signal showing that this simplified signal represents the skeleton of the source signal for which the effects due to turbulence or to instrument noise are removed. Figure 11. First 50 POD eigenvalues evaluated for the hot-wire anemometer signal acquired at x/w=4, y/w=1, z/h=0.9. POD mode 1 POD mode 2 POD mode 3 POD mode 4 POD mode 5 POD mode 6 POD mode 7 POD mode 8 Figure 12. Fourier power spectra of the components extracted with different POD modes for the signal acquired at x/w=4, y/w=1, z/h=0.9. Figure 13. Hilbert spectrum of the signal acquired at x/w=4, y/w=1, z/h=0.9.

12 Figure 14. Reconstruction of the signal acquired at x/w=4, y/w=1, z/h= CONCLUSIONS A procedure based on the Proper Orthogonal Decomposition (POD) for detection and extraction of components present in fluid dynamic time-series has been presented. The time-series is divided into different time-portions, denoted as snapshots, uniformly distributed along the signal time-length and composed of the same number of samples. POD is then applied to these snapshots producing an orthogonal basis composed of the so-called POD modes, for which the dominant ones represent the most significant realizations of the analyzed process. The obtained orthogonal basis is completely data-dependent, with no a-priori assumption and applicable to non-stationary signals. The POD modes are sorted by their significance and their contributions to the total energy of the signal is also evaluated. This optimality of the POD is very useful for real applications for which the components with significant energy may be analyzed, whereas the remaining ones can be simply neglected. The extraction of a component corresponding to a certain POD mode is performed through the convolution of the source-signal with the considered POD mode, suitably manipulated. The procedure based on the POD for component detection and extraction has been applied to hotwire anemometer signals acquired in proximity of the wake generated from a triangular prism placed with a vertical edge against the incoming flow. Flow fluctuations due to the dynamics of different vorticity structures can be easily extracted through the analysis of the most energetic POD modes and the time-frequency analysis is simultaneously carried out through the use of the Hilbert spectrum. Acknowledgements The authors would like to thank G. Buresti and L. Carassale for their invaluable suggestions and contributions to the paper writing. REFERENCES Bendat J.S., Piersol A.G. (1986). Random data: analysis and measurement procedures, 2nd edn. Wiley, New York Buresti G., Lombardi G., Bellazzini J. (2004). On the analysis of fluctuating velocity signals through methods based on the wavelet and Hilbert transforms. Chaos, Solitons and Fractals, Vol. 20, pp Buresti G., Iungo G.V. (2009). Experimental investigation on the connection between flow fluctuations and vorticity dynamics in the near wake of a triangular prism placed vertically on a plane. J. Wind Eng. Ind. Aerodyn. Vol. 98, pp Camarri S., Salvetti M.V., Buresti G. (2006). Large-eddy simulation of the flow around a triangular prism with moderate aspect-ratio. J. Wind Eng. Ind. Aerodyn. Vol. 94 (5), pp Carmona R., Hwang W.L., Torresani B. (1998). Practical time-frequency analysis. San Diego, CA: Academic Press Huang N.E., Shen Z., Long S.R.,Wu M.C., Shih H.H., Zheng Q., Yen N., Tung C.C., Liu H.H. (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc of the Royal society of London Series A- Math Phys Eng sciences, Vol. 454 (1971), pp Iungo G.V., Lombardi E. (2010). Time-frequency analysis of fluid dynamic time-series through a procedure based on proper orthogonal decomposition. Submitted to Experiments in Fluids. Lumley J.L. (1970). Stochastic Tools in Turbulence. Academic Press, New York Olhede S., Walden A.T. (2004). The Hilbert spectrum via wavelet projections. In: Proc. Roy. Soc. London A, Vol. 460, pp Sreenivasan K.R. (1985). On the finite-scale intermittency of turbulence. J. Fluid Mech., Vol. 151, pp