- Non Stationary Signals
- Final Report
- FrFT Bartlett
- 20150409054106156
- Automatic real-time electric guitar audio transcription .pdf
- 33-Prabhu T N - An Iterative Image-c
- An Adaptive Threshold Method for Spectrum Sensing in Multi-Channel Cognitive Radio Networks.pdf
- Chapter03_2007
- Russound Multiroom
- wavemeas
- Time Delay Estimate
- intro_2015.ppt
- Camp
- IJCSE12-04-07-098.pdf
- Performance Analysis of Downlink OFDMA Resource Allocation With Limited Feedback
- Assignment#2 (Khawaja Ali 15RB909)
- ICMAS96goto.pdf
- SmaartLive_Spectrum_Mode_Measurements
- Use of Carter's Coefficient With Narrow Teeth - 05248314 - 1967
- Fatigue Random
- Whitepaper_nCode_AHP_AcceleratedVibrationTestingBasedonFatigueDamageSpectra_v2-Halfpenny.pdf
- Uplink Interference Management for Hspa and 1xevdo Femtocells
- 02_stationarity.pdf
- Frequency Analysis Using the DFT
- Taslp Code
- Gregory
- Comparison of Vertical Handover Mechanisms Using Generic QOS Trigger for Next Generation Network
- EMC_EMI.pdf
- MohantyTCASII2006WatermarkingChip
- A
- VMD420_datasheet_NAE1032020
- afpsug11_ed09.pdf
- Breakers Segregations En
- 6 Spectral Kurtosis Theory
- ASM Handout Display
- ASM Handout Display
- Ml 080640184
- Relay Device Numbers
- Relay Device Numbers
- RelayDeviceNumbers.pdf
- pcf8574a
- 3VT93001SC00 - VT630 SHUNT TRIP UNIT.pdf
- 3vt93001sc00 - Vt630 Shunt Trip Unit
- 3VT93001SC00 - VT630 SHUNT TRIP UNIT.pdf

**Electrical and Electronics Engineering
**

Volume 15 Issue 6 Version 1.0 Year 2015

Type: Double Blind Peer Reviewed International Research Journal

Publisher: Global Journals Inc. (USA)

Online ISSN: 2249-4596 & Print ISSN: 0975-5861

**Spectral Kurtosis Theory: A Review through Simulations
**

By Venkata Krishna Rao M

Vidya Jyothi Institute of Technology, Hyderabad, India

Abstract- Kurtosis of a time signal has been a popular tool for detecting nongaussianity.

Recently, kurtosis as a function frequency defined in spectral domain has been successfully

used in the fault detection of induction motors, machine bearings. A link between the

nongaussianity and nonstationaity has been established through Wold-Cramer’s decomposition

of a nonstationary signal, and the properties of the so-designated conditional nonstationary

(CNS) process have been analytically obtained. As the nonstationary signals are abundantly

found in music, the spectral kurtosis could find applications in audio processing e.g. music

instrument classification and music-speech classification. In this paper, the theory of spectral

kurtosis is briefly reviewed from the first principles and the spectral kurtosis properties of some

popular stationary signals, nonstationary signals and mixed processes are analytically obtained.

Extensive Monte Carlo simulations are carried out to support the theory.

Keywords: spectral kurtosis, stft, random amplitude sinusoid, chirp, harmonic sinusoid, higher

order statistics, wold-cramer’s decomposition, mixing processes.

GJRE-F Classification : FOR Code: 090699

SpectralKurtosisTheoryAReviewthroughSimulations

Strictly as per the compliance and regulations of :

© 2015. Venkata Krishna Rao M. This is a research/review paper, distributed under the terms of the Creative Commons

Attribution-Noncommercial 3.0 Unported License http://creativecommons.org/licenses/by-nc/3.0/), permitting all non commercial

use, distribution, and reproduction in any medium, provided the original work is properly cited.

wold-cramer’s decomposition. Spectral Kurtosis Theory: A Review through Simulations Venkata Krishna Rao M Abstract. stft. Hyderabad. nonstationary signals and mixed processes are analytically obtained. Short time fourier transform perceptual coding of speech [5] and voice activity (STFT) for dynamically estimating the magnitude detection [6]. 𝑡𝑡 − 𝜏𝜏)𝑋𝑋(𝜏𝜏)𝑑𝑑𝜏𝜏 (1) −∞ Author: Vidya Jyothi Institute of Technology. signals or processes at different signal-to-noise ratios (SNRs) The spectral kurtosis properties of both stationary and is estimated and the results are in perfect match with the nonstationary signals as well as the stochastic mixtures previous analytical findings. Extensive Monte Carlo Keywords: spectral kurtosis. C haracterization of a given signal as noise like or The mathematical basics of spectral kurtosis are tone like finds several applications in music. Multi Band Excitation (MBE) model based described in section III. kurtosis as a function was defined on the real part of the STFT filter bank frequency defined in spectral domain has been successfully output to overcome the deficiency of the power spectral used in the fault detection of induction motors.9].g.e. Spectral Kurtosis nongaussianity detection [10]. Antoni [19. diagnostics [16. the important aspects of original kurtosis of the simulated stationary. In Section V the details of gaussian/nongaussian. In this paper.20] formulated SK differently by using Wold-Cramér Year the so-designated conditional nonstationary (CNS) process have been analytically obtained. Finally the review summary and future communication signal passes through a fading wireless work is given in Section VI. He also 49 applications in audio processing e. the signals may be spectrum and the expression for estimation of SK from stationary/nonstationary/transient signal or STFT is given in section IV. The results are in perfect match with the previous analytical findings. mixing at different signal-to-noise ratios (SNRs) is estimated. bearings. simulations are carried out to support the theory. are analytically obtained. introduced in the section II. and the simulation results reflected by a fluctuating target or clutter [7] or when a are presented. Recently. processes. nonminimum phase Wold-Cramer’s decomposition uniquely system identification [12]. Introduction The paper is organized section wise as follows. The spectral In this paper. The fourth-order cumulant based kurtosis of the time signal was traditionally used for II. harmonic sinusoid. nonstationay and mixed spectral kurtosis theory is reviewed from fundamentals.s) means the linear causal impulse response of e-mail: mvk_rao@hotmail. and the properties of proposed an unbiased estimator of SK.15] and vibratory surveillance and diagnostics of rotating machines [16].Kurtosis of a time signal has been a popular tool for (SK) was originally introduced by Dwyer [17] where it detecting nongaussianity. music instrument practically used it for machine surveillance and classification and music-speech classification. denoising [23] and subterranean popular stationary signals. Harmonic Retrieval from a) Background nongaussian processes[11]. As the nonstationary signals decomposition with a theoretical basis for the SK are abundantly found in music. The Nongaussianity and/or Monte Carlo simulations of the Test Signal Set and the nonstationary signal also occur when a radar signal is derived mixture processes. Here h(t. higher order reviewed and the spectral kurtosis of several processes statistics. Spectral Kurtosis 𝑡𝑡 𝑌𝑌(𝑡𝑡) = � ℎ(𝑡𝑡. perceptual audio comprising several popular signals are analytically coding [4].14. Within each category. Recently the frequency describes a non-stationary signal y(t) as the response of dependent kurtosis defined in spectral domain was a causal linear system with time varying impulse proposed and successfully used in bearing fault response h(t. s) excited by a signal X(t) i. India. I. the spectral kurtosis could find estimation of non-stationary processes. channel [8. Other applications of spectral Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I the theory of spectral kurtosis is briefly reviewed from the first kurtosis reported in the literature include SNR estimation principles and the spectral kurtosis properties of some in speech signals [22].com the system at time instant t when excited by an impulse © 20 15 Global Journals Inc.2. (US) . Extensive Monte Carlo termite detection [24]. random amplitude simulations are carried out to support the theory sinusoid. A link between the nongaussianity and Vrabie [18] justified the theoretical definition of SK and nonstationaity has been established through Wold-Cramer’s 2015 decomposition of a nonstationary signal. machine density to detect and characterize the signal transients.3].21]. chirp. detection [13. The Test Signal Set speech classification [1.

e. a) Real Sinusoids If Y(t) is a purely stationary process. this connectivity between the CNS and nongaussianity makes the kurtosis. which can be interpreted as the complex The spectral kurtosis of this mixture is given by [19] envelope of signal Y(t) at frequency f and dX(f ) is an ortho-normal spectral process associated with input driving process X(t).(6) becomes down that the simplest way to convert a nonstationary process to a CNS process is the time datum К𝑌𝑌 (𝑓𝑓) randomization. If the noise power becomes larger 𝑆𝑆4𝑌𝑌 (𝑓𝑓) and larger. the spectral kurtosis becomes 𝑆𝑆2𝑌𝑌 (𝑓𝑓) 𝐸𝐸{|𝐻𝐻(𝑛𝑛. and 𝜅𝜅𝑋𝑋 is the time kurtosis of the input process. the process Z(t) becomes purely К𝑌𝑌 (𝑓𝑓) = −2 the mixing noise process. a When the process Y(t) is gaussian. originally defined К𝑧𝑧 (𝑓𝑓) = К𝑌𝑌 (𝑓𝑓) 𝑓𝑓 ≠ 0 (9) on time processes to characterize the nongaussianity. −∞ 𝑍𝑍(𝑡𝑡) = 𝑌𝑌(𝑡𝑡) + 𝑁𝑁(𝑡𝑡) (5) where H(t. H(t. The spectral kurtosis of some other signals not (stochastic) frequency response of the time varying filter considered earlier is also obtained analytically. The frequency counterpart of eq.e. 𝑆𝑆2𝑌𝑌 (𝑓𝑓0 ) 𝐸𝐸{|𝐴𝐴|2 }2 © 2015 Global Journals Inc. some commonly found signals cumulants comes from the fact that dX(f) is a circular are considered for analytically computing the spectral random variable. (US) .e. f) depends on the outcome of the random the frequency f given by Year variable 𝜉𝜉. eq. 𝜅𝜅𝑋𝑋 = 0). f) be conditioned to 𝜉𝜉 i. stochastic and independent of the spectral 𝜌𝜌(𝑓𝑓) = (7) 𝑆𝑆2𝑌𝑌 (𝑓𝑓) 50 process dX(f). Thus eq. 𝑓𝑓)|4 } К𝑧𝑧 (𝑓𝑓) = К𝑁𝑁 (𝑓𝑓) 𝑓𝑓 ≠ 0 (11) = (2 + 𝜅𝜅𝑋𝑋 ) − 2 ≥ 𝜅𝜅𝑋𝑋 𝐸𝐸{|𝐻𝐻(𝑛𝑛. 𝑓𝑓. In particular. 𝜋𝜋). 𝜉𝜉) where 𝜉𝜉 is a [1 + 𝜌𝜌(𝑓𝑓)]2 [1 + 𝜌𝜌(𝑓𝑓)]2 representative random parameter of filter’s time varying transfer function. 𝜉𝜉) can be assumed to be time 𝑆𝑆2𝑁𝑁 (𝑓𝑓) stationary. Test Signal Set where the factor 2 in place of 3 as in usual definition of In what follows.(1) is b) Mixing Processes given by Let Z(t) be the mixture of two processes (i) a ∞ non-stationary process Y(t) and (ii). signal is clean. It may be noted frequencies except at 𝑓𝑓 ≠ 0. the spectral kurtosis of a stationary 𝑆𝑆4𝑌𝑌 (𝑓𝑓0 ) 𝐸𝐸{|𝐴𝐴|4 } К𝑌𝑌 (𝑓𝑓0 ) = −2= − 2 = −1 (13) gaussian process is zero (i. Any CNS process driven by a white К𝑧𝑧 (𝑓𝑓) = 𝑓𝑓 ≠ 0 (8) [1 + 𝜌𝜌(𝑓𝑓)]2 process X(t) of order p ≥ 4 is likely to be leptokurtic i. In many cases. f. H(t. its probability density function having tails flatter than If the noise power is zero i. 𝑓𝑓)|2 }2 = 𝛾𝛾4𝐻𝐻 (𝑓𝑓) (2 + 𝜅𝜅𝑋𝑋 ) − 2 ≥ 𝜅𝜅𝑋𝑋 𝑓𝑓 ≠ 0 (3) III. then К𝑁𝑁 (𝑓𝑓) vanishes at all nonstationary (CNS) process in [19]. which is a constant and independent of the frequency. then constant frequency is given by spectral kurtosis of Y(t) is given by 𝑦𝑦(𝑡𝑡) = 𝐴𝐴 cos(2𝜋𝜋𝑓𝑓0 𝑡𝑡 + 𝜑𝜑) (12) К𝑌𝑌 (𝑓𝑓) = 𝜅𝜅𝑋𝑋 (4) where 𝜑𝜑 is a constant initial phase from 𝑈𝑈(−𝜋𝜋.e. Thus the signal X(t) is stationary in general but non-stationary for a particular outcome 𝜉𝜉. 𝜌𝜌(𝑓𝑓) → ∞.(6) very useful in analyzing the nonstationary processes becomes through kurtosis defined in spectral domain. 𝑓𝑓) exp(𝑗𝑗2𝜋𝜋𝑓𝑓𝑡𝑡) 𝑑𝑑𝑋𝑋(𝑓𝑓) (2) noise N(t) . If the mixing process N(t) is stationary (white or Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Such a process was designated as conditionally colored) gaussian process. f) is the time varying transfer function of the system. those of its generating gaussian process and hence then 𝜌𝜌(𝑓𝑓) = 0 and hence non-gaussian. the 2015 where 𝜌𝜌(𝑓𝑓) is the local Noise-to-Signal power ratio at shape of H(t. Let H(t. a stationary additive 𝑌𝑌(𝑡𝑡) = � 𝐻𝐻(𝑡𝑡. In fact. the spectral kurtosis of the nonstationary [1 + 𝜌𝜌(𝑓𝑓)]2 signal Y(t) is defined as the normalized fourth-order spectral cumulant [19] as where 𝜌𝜌(𝑓𝑓) is finite. For the stationary white driving process X(t) of 𝜌𝜌(𝑓𝑓)2 К𝑁𝑁 (𝑓𝑓) К𝑧𝑧 (𝑓𝑓) = 𝑓𝑓 ≠ 0 (10) order p ≥ 2n. f) is stochastic К𝑌𝑌 (𝑓𝑓) 𝜌𝜌(𝑓𝑓)2 К𝑁𝑁 (𝑓𝑓) К𝑧𝑧 (𝑓𝑓) = + 𝑓𝑓 ≠ 0 (6) and can be represented with 𝐻𝐻(𝑡𝑡. Spectral Kurtosis Theory: A Review through Simulations at time instant t-s. then A real sinusoid of constant amplitude and 𝛾𝛾4𝐻𝐻 (𝑓𝑓) is independent of frequency and is unity. 𝛾𝛾4𝐻𝐻 (𝑓𝑓) is the kurtosis of the kurtosis.

2 decaying real envelope and 𝜑𝜑 is the constant initial phase from 𝑈𝑈(−𝜋𝜋. (US) . the spectral kurtosis can be obtained as 2015 b) Random Amplitude Sinusoid When a radar transmitted carrier signal is К𝑌𝑌 (𝑓𝑓) = 0 (21) Year reflected by a fluctuating target. inharmonic sinusoids appear frequently in music produced by several instruments. 𝑓𝑓𝑚𝑚 ≠ 𝑚𝑚𝑓𝑓0 .(17) and eq. η(f). then the spectral kurtosis from eq. The focus of this paper is Similarly.(18). density function. In case fundamental frequency 𝑓𝑓0 and its finite number of of deep fluctuations the constant part becomes zero harmonics 𝑓𝑓𝑚𝑚 = 𝑚𝑚𝑓𝑓0 as denoted by and then 𝐴𝐴(𝑡𝑡) = 𝐴𝐴𝑟𝑟 (𝑡𝑡). the carrier amplitude of Similarly the spectral kurtosis for other density the echo signal undergoes random fluctuations [7]. 𝑒𝑒 −𝑘𝑘𝑡𝑡 is the 𝑀𝑀2𝑘𝑘 = 2𝑘𝑘 𝑏𝑏 2𝑘𝑘 𝑘𝑘! 𝑘𝑘 = 1. Both harmonic and expressions [26. Then from eq.2 (20) If the amplitude decreases exponentially. then the where 𝑀𝑀𝑘𝑘 are the k-th standard moments. where 𝜅𝜅𝐴𝐴𝑟𝑟 is the coefficient of kurtosis or time kurtosis of If the frequencies are independent and no harmonic the random amplitude. In [25] the spectral kurtosis of such a where M is the number of harmonics. The coefficient of kurtosis 𝜅𝜅𝐴𝐴𝑟𝑟 relation among them i.(18). functions can be obtained. the simulations are carried out as described in eq. The spectral kurtosis of such a 𝑦𝑦(𝑡𝑡) = ∑𝑀𝑀−1 𝑚𝑚 =0 𝐴𝐴𝑚𝑚 cos(2𝜋𝜋𝑚𝑚𝑓𝑓0 𝑡𝑡 + 𝜑𝜑) (22) process is totally dependent on the probability density function of 𝐴𝐴𝑟𝑟 (𝑡𝑡). 𝐶𝐶𝑘𝑘 are the k. then the sinusoids based on cumulants can be obtained from the following are called inharmonic sinusoids. The amplitudes of random amplitude sinusoid was shown to be the harmonics decay at different rates depending on the К𝑌𝑌 (𝑓𝑓) = 𝜅𝜅𝐴𝐴𝑟𝑟 + 1 (17) instrument or the note played.(17) and constant frequency is given by eq. Spectral Kurtosis Theory: A Review through Simulations A real sinusoid of time varying amplitude and b). noise is called colored noise and is characterized by th central moments and 𝛾𝛾𝑘𝑘 are the cumulants. a) If the amplitude of the random sinusoid in eq. 𝐴𝐴(𝑡𝑡) = 𝐴𝐴𝑐𝑐 + 𝐴𝐴𝑟𝑟 (𝑡𝑡) (16) c) Harmonic and Inharmonic Sinusoids where 𝐴𝐴𝑐𝑐 is the constant part and the random part 𝐴𝐴𝑟𝑟 (𝑡𝑡) A harmonic sinusoid comprises a sine wave of which is also called as the multiplicative noise. a communication signal passing through a 51 not to derive such expressions.(17) and eq. 27]. 𝐶𝐶2 = 𝑀𝑀2 − 𝑀𝑀1 2 d) Additive Gaussian Noise 𝐶𝐶4 = 𝑀𝑀4 − 4𝑀𝑀1 𝑀𝑀3 + 6𝑀𝑀1 2 𝑀𝑀2 − 3𝑀𝑀1 4 Additive White Gaussian Noise (AWGN) signal 𝛾𝛾2 = 𝐶𝐶2 is a noise commonly found in communication channel which adds to the transmitted signal.(14) can be represented by the section V. The spectrum of 𝛾𝛾4 = 𝐶𝐶4 − 3𝛾𝛾2 2 = 𝐶𝐶4 − 3𝐶𝐶2 2 this signal is flat with a constant one-sided power 𝛾𝛾4 𝐶𝐶4 spectral density η within the channel bandwidth. 𝜋𝜋).(16) is e) Modulated Signals totally random with a gaussian density function. the amplitude of the signal at is a positive constant depending on the probability Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I the receiver input undergoes random fluctuations [8.(18) is given by 𝑦𝑦(𝑡𝑡) = 𝐴𝐴(𝑡𝑡) cos(2𝜋𝜋𝑓𝑓0 𝑡𝑡 + 𝜑𝜑) (14) (𝑏𝑏 − 𝑎𝑎)4� К𝑌𝑌 (𝑓𝑓) = 80 − 3 + 1 = 144 = −0.9]. the spectral kurtosis is frequency fc modulated by a sine wave of a low given by frequency fm and is given by 3𝜎𝜎 4 К𝑌𝑌 (𝑓𝑓) = 𝜅𝜅𝐴𝐴𝑟𝑟 + 1 = −3+1 =1 (19) 𝑦𝑦(𝑡𝑡) = 𝐴𝐴𝑐𝑐 (1 + 𝑚𝑚𝐴𝐴 𝑚𝑚𝑒𝑒𝑚𝑚(t) cos(2𝜋𝜋𝑓𝑓𝑐𝑐 𝑡𝑡 + 𝜑𝜑) (23𝑎𝑎) (𝜎𝜎 2 )2 where 𝐴𝐴𝑐𝑐 is the amplitude of the carrier and 𝑚𝑚𝐴𝐴 is the b) If the amplitude of varies uniformly in the range (a. 𝑚𝑚𝑒𝑒𝑚𝑚(𝑡𝑡) is the low frequency message © 20 15 Global Journals Inc. the (𝑏𝑏 2 2 80 signal is called a damped sinusoid and is given by � − 𝑎𝑎) �12� 𝑦𝑦(𝑡𝑡) = 𝐴𝐴 𝑒𝑒 −𝑘𝑘𝑡𝑡 cos(2𝜋𝜋𝑓𝑓0 𝑡𝑡 + 𝜑𝜑) (15) c) For a rayleigh distributed random amplitude. A typical profile of the harmonic amplitudes is given by 𝐴𝐴𝑚𝑚 = 1/𝑚𝑚. then An amplitude modulated signal with a carrier from eq. the standard moments are given by where k is the damping (bandwidth) factor. 𝜅𝜅𝐴𝐴𝑟𝑟 = = 2− 3 (18) If power spectral density of the noise within the 𝛾𝛾2 𝐶𝐶2 channel bandwidth is function of frequency. but to show that К𝑌𝑌 (𝑓𝑓) fading wireless channel. However for some more density In these cases. the time varying amplitude 𝐴𝐴(𝑡𝑡) in functions.e. modulation index.

The following signals are simulated using In a quadratic chirp signal. frequency is given by 1. An un-biased estimator of the spectral kurtosis where 𝜑𝜑(𝑡𝑡) is the time-varying phase of the carrier. A matrix popularly known as a spectrogram is formed by arranging STFT coefficients f) Chirp Signal as columns as given by A chirp signal is a frequency modulated signal 2015 𝑁𝑁−1 2 in which the frequency of the carrier is linearly or 1 2𝜋𝜋 𝑛𝑛𝑘𝑘 Year hyperbolically varied.(14). A Gaussian Amplitude Sinusoid signal c. the carrier frequency f is number of discrete frequencies. where α=df/dt is the chirp rate. This kind of chirp signal is 𝑆𝑆(𝑘𝑘. eq. A Random Amplitude Sinusoid signal a. A Rayleigh Amplitude Sinusoid signal frequency is given by e.(20) through eq. which 0 ≤ 𝑘𝑘 ≤ 𝐾𝐾 − 1. A Damped Sinusoid signal © 2015 Global Journals Inc. A Constant Amplitude Sinusoid signal 𝑓𝑓(𝑡𝑡) = 𝑓𝑓𝑐𝑐 + 𝛼𝛼𝑡𝑡 2 (27) 2. the instantaneous eq. the spectral kurtosis К𝑆𝑆 (𝑘𝑘0 ) at the 𝑡𝑡 𝑡𝑡 frequency index 𝑘𝑘0 can be computed from the 𝑘𝑘0 -th row 𝜑𝜑(𝑡𝑡) = 2𝜋𝜋 � 𝑓𝑓(𝑡𝑡)𝑑𝑑𝑡𝑡 = 2𝜋𝜋 � (𝑓𝑓𝑐𝑐 + 𝛼𝛼𝑡𝑡 )𝑑𝑑𝑡𝑡 0 0 of the spectrogram matrix 𝑆𝑆(𝑘𝑘0 . The discrete fourier transform (DFT) on a K- logarithmic. If the L number of varied as 𝑓𝑓𝑐𝑐 + 𝛼𝛼𝑡𝑡. .(12). A Uniform Amplitude Sinusoid signal where 𝛼𝛼 = (𝑓𝑓1 − 𝑓𝑓0 )/𝑡𝑡1 2 . A Chisquare Amplitude Sinusoid signal where 𝛼𝛼 = (𝑓𝑓1 /𝑓𝑓0 )1/𝑡𝑡1 . the Lower IV. b. 𝑙𝑙)| 4 Substituting eq. eq. The Double Side Band (DSB) signal. A Exponential Amplitude Sinusoid signal In a logarithmic chirp signal. Simulations and Results = (f1-f0)/t1. A Lognormal Amplitude Sinusoid signal f.(28). quadratic or signal. Then nonoverlapping frames used in the STFT analysis are time-varying phase of the carrier is given by considered as L number of the independent stochastic signal realizations. then the chirp rate is given by α V. (US) . sample signal computes the signal spectrum at K- In a linear chirp signal. Here the input signal y(n) is 𝑡𝑡 divided into overlapping or non overlapping frames each of size N. the instantaneous d. A Gamma Amplitude Sinusoid signal 𝑓𝑓(𝑡𝑡) = 𝑓𝑓𝑐𝑐 𝛼𝛼 𝑡𝑡 1 (28) g. 3. In this section a means of estimating the The Frequency Modulated (FM) signal can be spectral kurtosis from the short time fourier transform obtained as (STFT) is presented. multiplied by a window function w(k) like 𝑦𝑦(𝑡𝑡) = 𝐴𝐴𝑐𝑐 cos �2𝜋𝜋𝑓𝑓𝑐𝑐 𝑡𝑡 + 2𝜋𝜋 𝑘𝑘𝑓𝑓 � 𝑚𝑚𝑒𝑒𝑚𝑚(𝑡𝑡)𝑑𝑑𝑡𝑡� (23𝑏𝑏) a hamming window of same size and analyzed by using 0 the Fourier Transform. Spectral Kurtosis Theory: A Review through Simulations signal. K is the total number of frequency bins of one-sided STFT and L is the total 𝑦𝑦(𝑡𝑡) = 𝐴𝐴 cos 𝜑𝜑(𝑡𝑡) = 𝐴𝐴 cos 2𝜋𝜋𝑓𝑓(𝑡𝑡)𝑡𝑡 (24) number of frames contained in the signal. STFT Based Spectral Kurtosis Side Band (LSB) signal and the Upper Side Band (USB) Estimation signal can be formed by appropriate processing the AM signal. )The spectral kurtosis 1 = 2𝜋𝜋 � 𝑓𝑓𝑐𝑐 𝑡𝑡 + 2𝛼𝛼𝑡𝑡 2 � (25) at all frequency indices 0 ≤ 𝑘𝑘 ≤ 𝐾𝐾 − 1 is given by 𝐿𝐿 (L + 1) ∑𝐿𝐿−1 𝑙𝑙=0 ∙ |𝑆𝑆(𝑘𝑘. 𝑙𝑙)| } 2 2 1 𝑦𝑦(𝑡𝑡) = 𝐴𝐴 cos 2𝜋𝜋 � 𝑓𝑓𝑐𝑐 𝑡𝑡 + 2𝛼𝛼𝑡𝑡 2 � (26) 0 ≤ 𝑘𝑘 ≤ 𝐾𝐾 − 1 (30) If the start frequency at 𝑡𝑡 = 0 is fc and the end frequency at 𝑡𝑡 = 𝑡𝑡1 is f1. M is the hop size. we get К𝑆𝑆 (𝑘𝑘) = � − 2� 𝐿𝐿 − 1 {∑𝐿𝐿−1 𝑙𝑙=0 |𝑆𝑆(𝑘𝑘.(23) in eq. A Weibull Amplitude Sinusoid signal h.(15) and eq. The is proposed in [19] based on L realizations of K-sample frequency profile f(t) can be linear. 𝑙𝑙) = �� 𝑦𝑦 (𝑛𝑛 + 𝑙𝑙𝑀𝑀) 𝑤𝑤(𝑛𝑛)𝑒𝑒 −𝑗𝑗 𝑁𝑁 � 𝑀𝑀𝑊𝑊𝑛𝑛 𝑁𝑁 extensively used in pulse radars for achieving higher 𝑛𝑛=0 52 range resolution using a longer transmitted pulse. 0 ≤ 𝑙𝑙 ≤ 𝐿𝐿 − 1 (29) is otherwise possible with a shorter transmitted pulse where k is the frequency index. l is the time frame Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I [14].(22). A phase modulated signal is given by index.

A Double Side Band (DSB) signal The mixture signal is generated for a duration of c. A Lower Side Band (LSB) signal 1. 0 ≤ 𝑘𝑘 ≤ 255 which is spectral kurtosis is estimated for each mixture process.1a gives In all simulations a sampling frequency of 44100Hz is the Averaged STFT Spectrum of Mixture-1 for different used. Spectral Kurtosis Theory: A Review through Simulations a. Thus the mixture comprises a total five signals: a. Four peaks of different amplitudes observed in generating the test signals. 0dB. 𝑙𝑙) .7 and b. A Harmonic Sinusoid signal is added to the composite signal to form the first mixture 5.0. An Amplitude Modulted signal and -10dB. the noise floor also increases. Each frame is multiplied by a hamming window 7. 10dB. A Quadratic Chirp (Concave) signal 𝑆𝑆(𝑘𝑘. A Logarithmic Chirp Signal each of 256 frequency bins of spectrogram matrix is 53 Different mixture processes are formed by averaged over 399 frames to obtain a mean STFT Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I adding two or more of the above simulated signals. 0 ≤ 𝑘𝑘 ≤ 255. As the SNR A composite signal if formed by summing four decreases. STFT is computed for frames or d. An Analog Modulated signals signal-to-noise ratios of 30dB. Fig.0 is formed. .25. exp(-kt2) envelope Sinusoid signal 2. The low amplitude second peak is submerged a) Mixture-1 in the noise floor at low SNRs below 0dB. Colored Gaussian Noise The variance of AWGN is adjusted so as to obtain 6. A Linear Chirp signal giving a spectrogram matrix of size 256 × 399 Year b. Averaged SFT spectrum of (b). The spectrum 𝑆𝑆𝑀𝑀 (𝑘𝑘) = 〈𝑆𝑆(𝑘𝑘. An Additive Gaussian Noise process. exp(-kt) envelope Sinusoid signal 18000Hz with respective amplitudes: 1. b. White Gaussian Noise four constant amplitude sinusoids and gaussian noise. 4000Hz.e. (US) . A Wideband Frequency Modulated Signal overlap. sinusoids of frequencies 1800Hz. Chirp signals 2015 of 256 samples and a 256-point FFT is computed thus a.1610 seconds. Estimated Spectral Kurtosis for SNRs 30dB to -10dB © 20 15 Global Journals Inc. for computing STFT and for the spectrum correspond to four sine waves in the estimating the spectral kurtosis. A Quadratic Chirp (Convex) signal c. 0. 0 ≤ 𝑙𝑙 ≤ 398 d. Customized Matlab code is developed for SNRs. mixture. )〉. -5dB a. 20dB. 128 samples) e. 1 : Sum of sinusoids (a). 9000Hz and Fig. 0. An additive white gaussian noise (AWGN) 4. A Upper Side Band (USB) signal window size of 256 samples with 50% (i. b. called here as Averaged STFT Spectrum.

As the SNR decreases. The first peak at 2500Hz The third Mixture process is made up eight having the spectral kurtosis of -1 the constant amplitude sinusoids.(28) as explained in section IV. The Fig. amplitudes. The CGN is obtained by passing white gaussian the sinusoid. Estimated Spectral Kurtosis for SNRs 0dB Fig. 11. 0 ≤ 𝑘𝑘 ≤ 255 is sinusoids. colored the peaks have a value of -1 irrespective of the sinusoid uniform(i. Averaged SFT spectrum of (b). 2.5KHz. where the edges at 15KHz and 17KHz. the peaks values Global SNR is computed with respect to AWGN increase from -1 towards zero. which is computed based on the aggregate a 6-order butterworth band pass filter having passband 54 of all sinusoids. In fact these signals correspond to damping sinusoids are 1KHz. the spectral kurtosis tends to zero. Rayleigh. components: constant amplitude sinusoid(CAS). The other two components: CGN and CnGN These sinusoids are generated one by one and being stationary noise processes take a zero SK. relatively eq. the mixture becomes more and more also computed from the spectrogram matrix using gaussian. 10KHz. The peaks at 5000Hz and 8000Hz have the SK different probability density functions: gaussian. 2 : Mixture-2 (a). For higher SNR (30dB) factors. colored gaussian noise(CGN). The Year referred in the legend of the figures1(a) and (b) is the CnGN obtained by passing white uniform noise through global SNR. Local SNR computed at considering the other five components as composite peak locations vary depending upon the amplitude of signal. It may be noted down that the SNR having passband between 10KHz and 13KHz. 7KHz. 2 and 𝑒𝑒 −𝑘𝑘𝑡𝑡 .1b gives the spectral kurtosis (SK) of The second Mixture process comprises six Mixture-1 for different SNRs.5KHz. It may be noted that faster at 4000Hz where the local SNR is minimum. two The spectral kurtosis has four negative peaks damped sinusoids(DS1 and DS2) with different damping corresponding to four sinusoids. Spectral Kurtosis Theory: A Review through Simulations The spectral kurtosis К𝑆𝑆 (𝑘𝑘). separately.e. 4KHz. etc. nongaussian) noise(CnGN) and AWGN. Here it is maximum at 18000Hz and noise through a 6-order butterworth band pass filter 2015 minimum at 4000Hz. 13KHz. sinusoids which amplitude is changing with time as 𝑒𝑒 −𝑘𝑘𝑡𝑡 8. greater than zero indicate the nonstationary nature of exponential. К𝑆𝑆 (0) is to be ignored. The processed separately to obtain the SK estimate AWGN also takes a zero SK as expected. The frequencies of these the signals. For global SNRs below 0dB. each having a random amplitude following sinusoid. (US) .5KHz. uniform. 5.5KHz. AWGN power dominates the aggregate power of all Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig. Then the SK functions of all sinusoids are © 2015 Global Journals Inc. as the kurtosis is not defined at b) Mixture-2 𝑓𝑓 = 0. 14.2 gives the Averaged STFT Spectrum and c) Mixture-3 SK of Mixture-2 for 0dB SNR.5KHz and 16KHz.

2. 4 : Mixture-3 (a). Fig. As Fig. but the trend 2500Hz is +1. Estimated Spectral Kurtosis for SNRs 10dB © 20 15 Global Journals Inc.0. The 0. corresponding to uniformly distributed amplitude SK values of the sinusoids at 10dB compared to the variations (see eq. STFT spectrum correspond to eight sinusoids. The eight peaks in distributed amplitude variations (see eq. these random amplitude sinusoids for SNR=10dB. Averaged SFT spectrum of (b). 2015 Year 55 Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig. Estimated Spectral Kurtosis for SNRs 30dB Fig. Spectral Kurtosis Theory: A Review through Simulations overlapped and shown in single figure. corresponding to rayleigh amplitude sinusoids for SNR=30dB. (US) .(3) gives the amplitude variations (see eq. corresponding to gaussian distributed is same. The SK of the second sinusoid at respective SK values at 30dB are different. 3 : Mixture-3 (a).19). Averaged SFT spectrum of (b).20). the SK of first sinusoid at 1000Hz is .21).(4) gives the STFT spectrum and the SK of shown in Fig3(b). The SK of fourth STFT spectrum and the SK of these its random sinusoid at 5500Hz is 0.0.

over the chirp signals are considered. at the band edges. Spectral Kurtosis Theory: A Review through Simulations d) Mixture-4 shown in Figure 5(a) for SNRs of 20dB. Estimated Spectral Kurtosis for SNRs 20dB. Please note that each harmonics contaminated by an AWGN. However. the SK takes AM + AWGN. corresponding mixtures are are bandwidth. 10dB and 0dB e) Mixtures-5 SNR of 10dB.0 since the carrier is strong due to low modulation index and resembles a constant amplitude at carrier frequency of 12KHz.(6) through fig. However. 2015 Year 56 Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig. Averaged SFT spectrum of (b). exceptionally large values.(10) give the STFT spectra and SKs of these mixtures. The SK of the next two mixtures LSB and USB is over the signal bandwidth. The SK of AM signal is -1. 10dB and 0dB. AM-SC (DSB) + AWGN. the negative peaks constant with time. the SK takes exceptionally large values. but this amplitude remains higher SNRs. (US) . quadratic and logarithmic chirps generated using the eq. The STFT spectrum of this mixture is shown in Figure 11(a) for © 2015 Global Journals Inc. The STFT spectrum of this mixture is move from -1 towards zero. The fourth Mixture is basically a harmonic The ten negative peaks in SK plot of Figure 5(b) sinusoid with a fundamental at 800Hz and having 10 correspond to seven sinusoids. The SK of DSB is positive at carrier frequency of 12KHz showing its nonstationary nature. at the band edges..(24) through eq. f) Mixture-6 The sixth Mixture is made up of different chirp signals: linear. The amplitude of peak is -1 irrespective of the harmonic number for n-th harmonic is 1/n. 5 : Mixture-4 (a).(28). It may be noted down that the SK of a In this category basically five analog modulation chirp signal is nonzero positive. LSB+ AGWN. USB+ AGWN and FM+AGWN Fig. As SNR decreases. Each chirp is shown in different colors in Fig 11.

6 (a) : Averaged SFT spectrum of AM+AWGN (b).Estimated Spectral Kurtosis for SNR=30dB Fig. 7(a) : Averaged SFT spectrum of DSB+AWGN (b). (US) .Estimated Spectral Kurtosis for SNR=30dB © 20 15 Global Journals Inc. Spectral Kurtosis Theory: A Review through Simulations 2015 Year 57 Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig.

8(a) : Averaged SFT spectrum of LSB+AWGN (b).Estimated Spectral Kurtosis for SNR=30dB Fig. 9(a) : Averaged SFT spectrum of USB+AWGN (b). (US) .Estimated Spectral Kurtosis for SNR=30dB © 2015 Global Journals Inc. Spectral Kurtosis Theory: A Review through Simulations 2015 Year 58 Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig.

Averaged SFT spectrum of (b).Estimated Spectral Kurtosis for SNRs 10dB. (US) . © 20 15 Global Journals Inc. 11 : Mixture-6 (a). 10(a) : Averaged SFT spectrum of FM+AWGN (b). Spectral Kurtosis Theory: A Review through Simulations 2015 Year 59 Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I Fig.Estimated Spectral Kurtosis for SNR=30dB Fig.

R. second order statistics. Antoni and R. X. no. 2007. Systems and Signal Processing. August 2014. “An adaptive SK technique using spectral kurtosis at different SNRs. al. pp. Volume: 12. Pham Chau Khoa and Chng Eng Siong. 12. 2006. signal analysis”. ”Introduction to Radar Systems”.1099-1108. J. Conference. application to the vibratory surveillance and Fall 1996. CRC Press.39.1-5. Future work could be (i).8. 2. and Randall.37. July 1990. Randy Goldberg. vol. © 2015 Global Journals Inc. search. Speech and Signal Processing. Application”. 20. 2000. Vienna. 21. L. Y.. Giannakis and Jerry M.46-50. 2004.” Proc. 393-398. Some Properties. pp. review highlights the usage of spectral kurtosis for other No.200-203. Grado-Trieste.307. “Construction and frequency band persuit using a synthesized evaluation of a robust multifeature speech/ music criterion. Wheaton. pp. Speech and Signal Processing. Mendel.. no. detection and monitoring of electrical machines. 2. is a 10. Georgios B. March 1989. “Speech and Audio Processing in Adverse Environments”. radar signal and its application for fault detection of rolling classification using spectral kurtosis. N. Randall. 4. 1983. modulated) (ii). IEEE Region Systems and Signal Processing. Sawalhi.7. NSIP-2003. “The Spectral Kurtosis of Nonstationary local harmonicity feature for voice activity Signals: Formalisation. The Acoustics. Proceedings of the Third 17. 25. Blum. Obtaining closed-form expressions for spectral kurtosis Proceedings of Acoustics-2004. 12th European Signal Processing Language and Image Processing (ICALIP). 6. Jan. Slaney.. ”Bearing Fault Detection by resonant 1.1266-1276. and Lance Riek. 14-15 2. N. ”Identification and Estimation of Non- promising tool for analyzing nonstationary signals.” IEEE Multimedia Mag. 1997. Retrieval. No. Gold Coast. Chun Li et. “Fast computation of the kurtogram for gain diversity combining in a nongaussian multipath the detection of transient faults”.(iii).. 2013. pp. and J. Keislar.613 – 616. Austria. vol. and Liang.1 -37. pp. vol. al.3. Vol. Antoni. Wold. Monte Carlo simulations were carried out to support the pp. pp. pp. areas of signal processing like communications and 13. International Conference on Audio. Mechanical Systems and Signal References Références Referencias Processing. 2006. The spectral kurtosis of the simulated stationary. Mechanical fading environment”. 1750–1764. Prague. characterizing nonstationary signals”. 3. March 1991. Mendel. IEEE Trans on mixed processes are analytically given.179. France. 8. Valeriu Vrabie et. ”Exact Non-Gaussian Interference Model for Fading Channels”. pp. “Spectral 19. 124. Antoni. Makloi. pp.. Speech and Signal Processing. pp. Scheier and M. 2002. Extensive Signal Processing. element bearings”. Keh-Shin Lii. 2008. engineering and Mechatronics. “A Practical Definition To Application". al. J. T. pp. 94-100. Eberhard Hänsler and Gerhard Schmidt (editors).168 . IEEE Trans on complements the traditional power spectrum based on Acoustics. 2015 theory. evaluation of wireless communications using equal 21. on Mechnanical discriminator. IEEE Transactions on The cumulant based spectral kurtosis defined in Wireless Communications”. Ananthram Swami and Jerry M. on Acoustics. 607-610. E. Mechanical Second edition. 8. 6th IEEE International Handbook of Speech Coders”. pp. Conf. classification of communication signals 14. IEEE. spectral kurtosis is briefly reviewed from the 11.360-377.. 3-5 November of communication and radar signals (analog or digital 2004. Proc. and Rodet. Australia. “Content-based classification. Skolnik. G. 20. 108- 10 Conference. The properties of spectral kurtosis of Cumulant-Based Approach to the Harmonic popular stationary signals. Spectral Kurtosis Theory: A Review through Simulations VI. Wang. Conclusions and Future Work 9. Merrill I. spectral kurtosis to bearing diagnostics”. ”Detection of non-gaussian signals by International Conference on Music Information frequency domain kurtosis estimation”. J. frequency domain originally proposed for bearing fault Issue: 1. Vol.. Vol. Vol. E. conf. And detection”. TENCON 2004. McGraw-Hill. D.5. Italy. IEEE Trans on 60 good match with the previous analytical findings. 1980. pp. C. Yi Chen et. In this paper. pp. "Spectral Kurtosis: From 5.”proc 3rd Int. 2. and Chayawan. nonstationay and mixed processes at different signal-to- Year ”Identification of Non-minimum phase systems noise ratios (SNRs) is estimated and the results are in using Higher Order Statistics”. ”Toward Systems and Signal Processing. vol. Antoni. Workshop on Nonlinear Signal and Image pp. Processing. Paris. nonstationary signals and Retrieval and related Problems”. automatic music audio summary generation from 308 -331. “The application of Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I radar signal processing. 2004. M. and retrieval 16. ” fundamentals. Mechanical 3.38. Springer.37. “The spectral kurtosis: of audio. It Gaussian ARMA Processes”. Int. Burthe. pp. ”Performance 282 . Peeters. Vol. J. vol.. R. ICASSP-83. (US) . diagnostics of rotating machines”. “The spectral kurtosis: a useful tool for 7. 18. pp. Dwyer. Shanghai . A.ICASSP-97.407 – 413 20. 15. the theory of No. 27–36. 2012.

79. 171-174. al. 351-354. 23. Moreno. G.Unnikrishna Pillai. Fourth edition. and A. Valeriu Vrabie et. 4th IEEE Workshop on Intelligent 2015 Data Acquisition and Advanced Computing Systems: Technology and Applications IDAACS Year 2007. © 20 15 Global Journals Inc. C. “SNR estimation of speech signals using subbands and fourth-order statistics”. March 1989. 27. Spectral Kurtosis Theory: A Review through Simulations 22. IEEE Trans on Acoustics. 2007. 25. Puntonet. Speech and Signal Processing. and S.1998. proc IEEE. Vol. Goubran. Random Variables and Stochastic Processes”. IEEE-SP International Symposium on Time-Frequency and Time-Scale Analysis. O. pp. P. J. Jerry M. 26. Amblard. Vol. de la Rosa. 7. pp. (US) . 24. Lett. 6_9 Oct. ”Probability. No. vol. Proc. Mahmoud. 625-628.. 1999.. Nemer. Proc. 2002. Mendel. G. March 1991.146-148. no. “Subterranean termite detection using the spectral kurtosis”. pp. R. IEEE Signal Process. J. “Denoising using wavelet packets and the kurtosis: application to transient detection”. E.3. ”Tutorial on Higher Order Statistics(Spectra) in Signal Processing and System Theory: Theoretical Results and Some Applications”.37. Surveilance 5 CETIM Senlis 11-13 October Global Journal of Researches in Engineering ( F ) Volume XV Issue VI Version I 2004. Jul. Athanasios Papoulis and S. No. pp. Ravier and P. 6. Tata McGraw-Hill. pp.278-305.3. " Application of Spectral Kurtosis To Bearing Fault Detection in Induction 61 Motors".

- Non Stationary SignalsUploaded byJames Foot
- Final ReportUploaded byUtpal Das
- FrFT BartlettUploaded bytazishaam
- 20150409054106156Uploaded bysiva
- Automatic real-time electric guitar audio transcription .pdfUploaded byRizvi Ahmadi
- 33-Prabhu T N - An Iterative Image-cUploaded byDinesh Goyal
- An Adaptive Threshold Method for Spectrum Sensing in Multi-Channel Cognitive Radio Networks.pdfUploaded byKriti Chhabra
- Chapter03_2007Uploaded byFreddy Pesantez
- Russound MultiroomUploaded byCarles Casanovas
- wavemeasUploaded byEgwuatu Uchenna
- Time Delay EstimateUploaded byNgoacmieng
- intro_2015.pptUploaded byKristen Hernandez
- CampUploaded byleadstructure
- IJCSE12-04-07-098.pdfUploaded byBiswajit Mohanty
- Performance Analysis of Downlink OFDMA Resource Allocation With Limited FeedbackUploaded bykhajarasool_sk
- Assignment#2 (Khawaja Ali 15RB909)Uploaded bykhawaja Ali
- ICMAS96goto.pdfUploaded byfifster
- SmaartLive_Spectrum_Mode_MeasurementsUploaded byapi-3781771
- Use of Carter's Coefficient With Narrow Teeth - 05248314 - 1967Uploaded byCasualKilla
- Fatigue RandomUploaded byபிரபாகரன் ஆறுமுகம்
- Whitepaper_nCode_AHP_AcceleratedVibrationTestingBasedonFatigueDamageSpectra_v2-Halfpenny.pdfUploaded byTaleb Alrayyes
- Uplink Interference Management for Hspa and 1xevdo FemtocellsUploaded byKunwar Atul Singh
- 02_stationarity.pdfUploaded byFlorent Bersani
- Frequency Analysis Using the DFTUploaded byDasaka Brahmendra
- Taslp CodeUploaded byMd Moinuddin
- GregoryUploaded byGeorge Baciu
- Comparison of Vertical Handover Mechanisms Using Generic QOS Trigger for Next Generation NetworkUploaded byijngn
- EMC_EMI.pdfUploaded byElder Silva
- MohantyTCASII2006WatermarkingChipUploaded byAnagachinmey
- AUploaded bythinckollam

- VMD420_datasheet_NAE1032020Uploaded byCafea Tchibo
- afpsug11_ed09.pdfUploaded byCafea Tchibo
- Breakers Segregations EnUploaded byCafea Tchibo
- 6 Spectral Kurtosis TheoryUploaded byCafea Tchibo
- ASM Handout DisplayUploaded byCafea Tchibo
- ASM Handout DisplayUploaded byCafea Tchibo
- Ml 080640184Uploaded byCafea Tchibo
- Relay Device NumbersUploaded byCafea Tchibo
- Relay Device NumbersUploaded byCafea Tchibo
- RelayDeviceNumbers.pdfUploaded byCafea Tchibo
- pcf8574aUploaded byCafea Tchibo
- 3VT93001SC00 - VT630 SHUNT TRIP UNIT.pdfUploaded byCafea Tchibo
- 3vt93001sc00 - Vt630 Shunt Trip UnitUploaded byCafea Tchibo
- 3VT93001SC00 - VT630 SHUNT TRIP UNIT.pdfUploaded byCafea Tchibo