SIGNAL PROCESSING IN 3807 VIBRATION ANALYSIS R.B. Randall School of Mechanical and Manufacturing The University of New South Wales Sydney 2052, Australia gineering 1.“ INTRODUCTION Signal processing in vibration analysis can be divided single channel or “signal analys jon signals rest jon of the characteristics of the forcing function and the structural fe, Where one forcing function dominates, in so- called SIMO (single input, multiple output) as in a very large number of cases where stresses are below y jon and response are convolved in the time domain, mult win and additive on taking logs of the frequency fu ) of the forcing frequency are generated, very easily related to the fundamental frequency. A further reason for the widespread use of frequen can now ce the introd ng Signal analysis is of in cases where the forcing functions ‘cannot be measured, but changes in the response signals are assumed to be dominated either by the forcing function or the structural response, based on other criteria. An example is the co ing or diagnostics of operating machines, Even though frequency analysis using the FFT is the most commonly applied technique, other signal analysis procedures have advantages in certain circumstances as discussed below. For short signals, better frequency resolution can be obtained using parametric analysis techniques such as the maximum. entropy method [1.2]. Since basic Fourier analysis assumes infinitely long time records, information is lost about time localisation of events. Analysis in both time288 and frequency can be achieved by moving a finite time window along a signal and calculating the frequency spectrum for each position (so-called short time Fourier transform - STFT). Because of the property of the DFT that the frequency resolution is the reciprocal of the length of record transformed, the better the time mn and vice versa. Techniques have been developed (o improve the combined based on the Wigner-Ville distribution technique which has become octave because to vibrations, where standards recommend analysis in 1/3-octave bands. Before the advent of the FFT algorithm, most frequency analysis was done with analogue iters, measuring the power of time signals transmitted by a series of band-pass Iters each covering part of the frequency range to be analysed. Although most narrow band analysis is now done digitally using the FFT, digital filters working in the time domain in a similar manner (0 their analogue counterparts have advantages in certain situations, in particular for constant percentage bandwidth (Uin-octave) analysis over a wide frequency range on a logarithmic scale, ‘Some effects are more easily interpreted in the time domain rather than the frequency domain, and there are a number of techniques to enhance such effects. An example is autocorrelation analysis which gives a measure of how well a signal correlates with delayed versions of itself. This can be useful for enhancing periodicity and echoes, for example. A related time domain funct cepstrum, which in general is better for detecting echoes, and which hi ‘main applications in vibration analysis: . of equally spaced components in a spectrum such as harmonics and sidebands * Helping to separate forcing and transfer functions in response sigt they are addi ccepstrum. ‘A very important time domain technique is synchronous averaging, where signal sections are averaged toge example a once-per-rev tacho signa rate. The best known example is the pair of gears in mesh can be separated from each other (and background noise) by averaging synchronously with each gear in turn, In signals from rotating machines synchronous averaging must often be combined with so-called “order analysis” to remove the effects of speed variation by sampling the signals synchronously with the speed of the machine. This is now most efficiently done by digital interpolation (resampling) techniques. Apart from eliminating the effects of small 289 speed fluctuations, this process can also be used to study how the vibrations at the various harmonic orders vary over a wide speed range such as during a machine run-up or cosst-down, Synchronous averaging is most useful when signals are exactly periodic (or synchronous); however, if they are only approximately periodic but their autocorrelation function is then their eyclostationarity and spectral correlation properties may be of interest. Other effects in vibration signals are often due to modul ide or phase modulation, so that the signals obtained by demo interest than the example where amy of ‘signals from Ise responses due to imy ransform techniques, which have a close relationship with Fourier analysis, In multiple channel analysis, one or more forces are applied (and measured) by the user, and when combined with the corresponding response measurements from all over a structure can be used to determine the dynamic properties of the structure. Once again the most commonly applied techniques are based on the FFT process to extract frequency response functions (FRFs) from which a complete scaled modal model can be determined. FRFs are only strictly valid for linear systems and a measure of the degree of linearity of the relationship between the forcing function and response is given by the coherence. For calculation of FRFS and coherence itis necessary 10 measure cross spectra between pairs of signals, the ‘cross spectrum being the Fourier transform of the eross correlation function, similar to the autocorrelation furction but relating one signal to a delayed version of the other. The inverse Fourier transform of an FRF is the corresponding impulse response function, which represents the system dynamic properties in the time domain, In some situations, such as when the vibration signals represent “free all forcing functions have been removed, then otter techniques can be applied to determine the modal properties of a structure, such as the “Ibrahim i (ITD) method [1.6] and Prony method. The complex cepstrum and cepstrum have also been found useful for determining modal of structures based on response measurements only, because of the ly mentioned property of the cepstrum that forcing and transfer function ighted techniques are explained in a discussion of some basic theoretical concepts. channel signals for modal analysis is only the subject of other specialist papers. However, treated sketchily as290 2. BASIC THEORY 2.1 The Fourier Transform 2.1.1. Fourier Series The basic concept of Fourier analysis is to express signals as a summation of sinusoidal components, and with few exceptions virwally all signals can be ‘decomposed in this way. Fourier's original analysis (now called Fourier Series) was applied to finite length signals (or periodic signals, as the resulting solution is periodic wit ith as period). In vibration analysis itis used almost lusively for periodic signals, as produced by a machine rotating at constant speed. Thus, for any periodic signal g(t) of period. T for which (1) = g(t +T) itcan be shown that a, cos(kast)+ 5b, sin( kent) an ‘where @y_ is the fundamental angular frequency in rad/s (= 220/T ) and where the coefficients of the cosine and sine terms can be ol with g(1), as follows: _2_n dined by correlating them Cr 22) 2 pra =F], 8Osinkayndt 23) ‘Thus, the total component at frequency @,(= Key) is given by’ 4, cos(@.t}+B, sin(@,1) which can alternatively be written as C, c0s(a,t +9.) (2.4) and, =tan""(b, /a,) 291 Expression (2.4) can again be represented as Figure 2.1. Representation of a sinusoid as a sum of two rotating vectors 23) Cer oti —q lexpli(eor + O)) + exp jou +O) which can be interpreted as two rotating vectors, each of lengthC, /2, one rotating at angular frequency @, with initial phase, and the other rotating at angular frequency Ay Cos(k@gt + 9) with inFig2.1 Using this interpretation of Fourier analysis as representing g(t) asa sum of rotating vectors leads to the alternative version of Equ.(2.1) as: git)= SAcexp(jay, 1) eo phase — 9 , a illustrated where the coefficients A, are now complex and incorporate the phase shift form the = Lexpt jd) en ‘The equation for calculating the coefficients A, (equivalent to Equ.(2.2,23)) now becomes Lp Az Tian (1) exp(- j a0) dt 28)292 ‘This has the simple physical interpretation that multiplication by exp(— j @,¢) subtracts angular frequency @, from each component, meaning that the one originally rotating at @, is stopped in the position it had being extracted by the integrs ime zero (¢his then all other components still rotate at some other multiple of @, (either positive or negative) and thus integrate to zero over the periodic time. Thus each frequency component A, represents the position (and value) of the rotating vector at time zero, so th 0 obtain its position at any other time f it is necessary to cause it (0 rotate at angular frequency @, by ing by exp( jag!) Summing then over al frequency components gives Equ.(2.6). Before leaving this interpretation of Fourier series anal rotating vectors each of amy as a sum of ide half that of the corresponding sinusoid (ie C,/2), but having 2 two-sided spectrum in that each postive frequency ‘component is accompanied by its complex conjugate at negative frequency, it can be seen that the same result can be achieved by retaining the positive frequency components only but doubling their length to Cand then taking the projection fon the imaginary axis is the Hilbert transform of to the vector sum of a number of components as Prgection enimeg Figure 2.2. Equivalence of vector sum and projection on the real axis ord 293 exponentials, in particular transients to whic applies. The Fourier transform can be derived from. the periodic time to tend to infinity and at the same time T because transients have finite energy rather than finite power. Equations (2.8) and (2.6) then become: Gf) =f 8 exp(—j2mfi) at 29) g(t) = [GUS exp 2m df respectively, where angular frequency @, in rad/s has been replaced by the ‘continuous frequency function f expressed in Hz, Equations (2.9) and (2.10) are known as the forward and inverse Fourier (inte ‘At this point it is convenient to highlight transform of Equ.(2.9) and the Laplace transform, defined by: oo=f g(t) exp(-st) dr where s is a complex variable which can be represented as G+ j@ in terms of y pars. Thus, f functions, which are they do not exist teansform, known aS the frequency response fu transform (transfer function) for the special case that 5 is restricted 10 the imaginary axis ($= j= j2nf ) Note the symmetry between the forward and inverse transforms of Equ.(2.9, 2.10), the only difference being the sign of the exponent. This means that, most ofien, results ‘which apply to the forward transform also apply to the inverse transform, for example the convolution theorem to be discussed below. This is particularly the case for real, even functions, for which it makes no difference if time of frequency run forwards or backw its real Sampled time signals ly have to be digitised or the Fourier transform means me signal is periodic. ‘The corresponding versions of the forward and inverse transforms are GA) = Sates j2af t,) 2)294 1 pun 8 ETT ia OP RP 2A te) af (2.13) where t, =nAr=n/ f, 2.1.4 The discrete Fourier transform (DFT) ‘The sampled time signals in 2.1.3 are in principle of infinite length, but when the record length is fini series in the spectrum is itly periodic. As seen in 2.3(b, ¢) so that both the frequency spectrum are discretely sampled and periodic, The Continuous infinite integrals of the Fourier transform become finite sums, usually expressed as: ve G(k) =1N Y gin)exp(—j2nkn/ N) (2.14) = x g(n) = ¥ Gk) exp j2nkn/ N) 2.15) ms This version corresponds most closely to the Fourier series in that the forward transform is divided by the length of record N to give correctly scaled Fourier series components. If the DFT is used with other types of signals, eg ‘tansients or stationary random signals, the scaling must be adjusted accordingly as discussed below. Note that with the very popular signal processing package division by Nis done in the inverse transform, which requires a even though the forward transform is then closer to the Fourier lied by the discrete equivalent of dt Note also that for convenience, the time and frequency zero positions have been shifted from the centre of the record to the beginning, but because of the Periodicity this just means thatthe second half of each record represents the negative axes (see Fig.2.3(d), ‘The forward DFT operation can be understood asthe matrix mi 1 Gx =a MinBa 295 (a) Ineoeal Teun expti29m dp = ato expan at _ on FREQUENCY Periodic in time domain Discrete in frequency domain FREQUENCY (©) Sampled TIME lh Gis) = E sindew-f2970/5,) FREQUENCY Discrete and periodic in both ‘Time and frequency domains i Nt Figure 2.3. Various forms of the Fourier transform [2.1(a) Fourier integral transform (b) Fourier series (c) Sampled functions (d) Discrete Fourier ‘transform296 where Gy represents the vector of NV’ frequency components, the G(k) of Equ(2.142.15), while’ g, represans the N time samples g(n). Wha representa square matrix of unit wctors exp(—j2mkn/N) with angular ovientatia depending on the frequemy index k (the rows) and time sample index 7 (the columns). This is ilustated graphically in Fig 2.4 G) [Aa AAAAAAT (e, G Aa pav Ker [s, G Aryesrye fe, Gl _1|s eer yr el fe, Gl 8iAYAVAW AY Ie, _ G, AK PRY Te 4) [g, ° G, he yr hey rl! fg, eel G, AxeK yar! [e, Figure 2.4. Matrix mpresentation of the DFT Far k=O the zero frequency value G(O) is simi the time amples g(r) as would beempested. For k=1 the unit vector rotates = UN th a revolution for each time ample inerement, resulting in one compl (negative)evolution after N samples. For higher values of the rotation speed is mportionally higher. For k=.N/2 (half the sampling frequency, the so-called fyquist frequency) the vecta turns througt but itis mrpossible to see in which dreetion vector tum through more than #2 (a the nega interpreteas having turned through kss than thus if thaime signal has been lowpas Should aleys be the case) the secoad hal’ of Gy frequencyomponents from the negate Nyaui y the mean value of 2 for each time sampl has turned. For &>N/2_ the irection) but is more easily form (thew-called radix 2 algorithm) the and factoies the matrix Wyq into log, N’ mawices each with the property that imulpicdpn by them only requires -W complex operations as compared wit 297 IN? operations required for direct m by Wg. This the total number of complex operations is reduced from N? to N log, N , asaving by a factor fof more than 100 for the typical case where N= 1024 (=2'°). Other savings ccan be made in special cases, and a similar but not so effective gain can be made for factorisation other than in powers of 2, but the main point is that the properties of the FFT are those of the DFT. 2.2 The Convolution Theorem tates that a Fourier transform in either direction jon and vice versa. One example already sn an input force and the impuls their respective represents a moving average or swept of one function by the other reversed, as expressed by the Duhamel integral x) =f g(DAC—2) de represented as 2) = 8) #A) simple and involves shi the delta function and im factor can be complex and thus what follows, considerable use will be made gray relationships 2.2.1 Fourier Series from the Fourier transform ‘As a first example it can be seen that any periodic signal can be generated by convolving the transient representing one period wi ta functions with 1/T., so that the Fou series is obtained by mul functions (thus s fundamental frequency) and scaling the samples by the same time this corrects the physical dimensions series components have the same units as the o Figure 2.5 shows an example of how the Fourier series of a rectified cosine can be obtained from the Fourier transform of a single half-cosine pulse. To obtain the Fot necessary t spectrum at intervals of 1/27 and mi ng the Fourier tr the harmonics of the298 RO FREQUENCY Oe Liiifjiiy = 8 — ti Figure 2.5. Determining a Fourier series from a Fourier transform 1) half-cosine pulse and its Fourier transform )) train of delta functions and its spectrum (c) full-wave rectified cosine and its Fourier series spectrum the time domain that of the response of a si 1¢ domain is an exponen degree-of freedom damped sinewave (SDOF) system, which in the represented by exp(-ot)sin(27yf) 2.18) “The spectrum of the sinewave can be represented by two delta functions, cach of amplitude 1/2 the one at + fo. having a phase angle of — 1/2 and the fone at ~ fy having a phase angle of + 1/2. The Fourier transform of the exponenti 2.19) o+ jmp which changes in phase from +%/2 at ~s> through zero at zero frequency to =m/2 at +e. Figure 2.6 shows how the mul in the time domain results in a convolution of their resp. and giving the well-known frequency resonance frequency (where the phase is tuned through ~1./2), zero phase at zero frequency and —T at +20 wy TIME FREQUENY fe (a) fw FA Figure 2.6. Spectrum of an exponentially med sinewave. In the spectrum of (c) the arrows depict thesonance peak 2.3 Hilbert transform relationships ‘The Hilbert transform can be said to be the relatiaaip between the imaginary parts of the Fourier transform of a one-sidafetion. For exam impulse response function is causal and thus one-sidafinte time domain, and ‘means that the real and imaginary parts of the correyaing frequency function {eg that shown in Fig2.6) are related by a Hilbert tral. That there should be a relationship becomes evident when itis considered causal function is made up of even and odd components which are identiedfr postive time, and thus ‘cancel for negative time. Thus: = x,(0+x,() 2.20) and x,(0= x,(Oxsgn) 221) where sgn(t) is the sign function. Since the even partdatime function transforms 1m, and the odd tthe imaginary part, by ccanleseen that real part of its Fourier tran jing the convolution theorem to Equ.(2 (f= Xy(f* sen} (222)30 300 ‘The Fourier transform of the sign funetion is the imaginary hyperbolic function so th ag transform (X g(f)=X,(f)) to the imaginary pan (X,(F)= Xf) i and writing out the convolution in full, is given by: ie X(N= OO ‘The equivalent equation for the Hi the final expression relating the real part of the Fourier refequency Negative frequency is compos int ‘com a 2.24) TT Posie frequency (Components doubled ‘Taking the Fourier transform of Equ.(2 Xf) = XP) sents) (2.25) Zero frequency Component changed ‘Anaitie Time Signal of| ‘which he real part itt the same at bow Negative which shows that a Hilbert transform can be achieved more simply by transforming freeney to the frequency domain, shifting the phase of positive frequency components by ee 1/2 and of negative frequency components by +R/2, and then tzansforming back tothe time domain A one-sided frequency spectrum can similarly be divided into conjugate even and conjugate odd components which transform by the inverse Fourier transform to real and imaginary time signals, respectively, which are related by a ttansform. The sum of these two components is known as an "analytic which can be formed fom a given real time signal by adding j times its Hilbert transform. ively, it can be obtained more simply by transforming the realtime signal into the frequency domain, obtaining the equivalent one-sided spectrum by multiplying by 2H(f), where H(f) is the Heaviside or unit function, and transforming back to the time domain. 0 a very ient way of performing a Hilbert transform. Note from Fig.2.2 thatthe Hilbert Figure 2.7. Manipulation of the positive frequency spectrum to obiain a real time signal processing package such as M: transforming back to the time domai is not necessary (0 adjust all negative frequency components in the same way (but complex conjugate) as the positive frequency component ‘much simpler to multiply the positive frequency components ty 2 (but not the zer0 frequency component), set the second half of the spectrum (the negative frequency components) to zer0, perform an inverse m to an analytic signal and simply take the real part Unis last operation is usually necessary even when working with 2-sided spectra, as the program does ‘now that the answer is supposed to be teal, and will usually calculate a very imaginary part). This process isilustrated in Fig 2.7302 3. PRACTICAL FFT ANALYSIS 31 Pitfalls ‘The so-called pitfalls of the FET are all properties of the DFT and result from the three stages in passing from the Fourier integral transform to the DFT. The first step is digitisation of the time signal which can give rise to step is truncation of the record toa finite length, which can give rise to leakage or window effects, while the third results from discretely sampling the spectrum, Which can give rise to the picket fence effect. Figure 3.1 shows these three steps graphically, using the convolution theorem [2 In Fig.3.1¢2c) the infinite continuous time signal is sampled as in Fig 2.3(€) producing a periodic spectrum witha period equal to the sampling frequency f, 1k can be seen that ifthe original signal contains any components outside the range E fy, where fy is the "Nyquist frequency” or half the sampling frequency, then these will overlap with the true components giving “aliasin frequencies represent removed, 50 the signal (rectangular) window. The speci which acts asa filter character ic, hence the term in Fig.3.1(Fg) the continuous spectrum is discretely sampled in the frequency domain, which corresponds in the time domain to convolution with a ain of delta functions of spacing T , making the time signal periodic, The spectrum is not necessarily sampled at peaks; hence the term "picket fence effect”; i is as though the spectrum is viewed through the slits in a picket fence. 303 FREQUENCY it e si) cw fHitittntttttt ! tek : sllea t GU *S() ai f “\Ep we fj YD T=Nat he GPS HX) f @) Kew Time samp Truncation -f/2 1 fy af == T at f Yo TeNat ‘Tractfltnatl f jure 3.1. Three steps in passing from the FT to the DFT (2.1] 8) Frequency sampling304 : Oy i i El E A © fol i bf Eve ure 3.2. (A) Time signal correctly lowpas filtered (B) Spectrum of (A) (C) ime signal without lowpass filtration (D) Spectrum of (C) (note aliasing) To avoid ali with a very steep of 120 dBioctave, used. Thus with a 1 transform, spectrum line number $12 is at the Nyquist frequency, and higher frequencies fold back towards the measurement range. Line number 624 folds back into the top of the desired measurement range ime signal, and vice vers. The effects of leakage are influenced by the final spectrum sampling, and sin(x) x sime(x) function) if the window contains an integer number of periods of a sinusoid even though each spectral ine is associated with a sine function, these are sampled atthe zeros and are thus not apparent. On the other hand, inthe worst ease ofa residual half period, the elfectve filter characteristic is Very poor. Use is made of this phenomenon synchronised with machine speed and it can be arranged harmonics of) which case a ‘window can be used. Otherwise, ly necessary to choose a data window other characters, This is discussed below. Fig.3.3 illustrates that for a rectangular window (whose FT is a achieve a 305 1) courmuous Here ‘SPECTRUM Figure 3.3. How frequency sampling affects the apparent filter characteristic of a window (A) Integer number of periods in record (B) Extra half period in record Because of the picket fence effect, spectral sampled at their peaks and the "picket fence he difference between the ‘tue value and the value of the maximum sp . For rectangular weighting this can be as much as 3.9 dB, and most other windows have a reduced value. are not necessarily 3.2 Data windows 3.2.1 Continuous signals For continuous signals. a major function of the window is to reduce the effect of the discontinuity which usually arises when a random section of signal is made periodic. Pra ‘means minimising the sidelobes in the filter characteristic, both the highest and the remaining ones (by maximising their rate of falloff). To improve enhancement of dis:rete frequency components with respect to broadband noise itis desirable to minimise the noise bandwidth of the characteristic, but on the other hand, Mn must also be paid to minimising the picket fence effect, Table 3.1 gives a comparison of the properties of the most common windows applied to stationary signals, and Figure 3.4 (from [3.{]}compares their worst case filler characteristics. The Hanning window, which can be considered as one period of a sine squared function, is a gocd general purpose window, with picket fence sd to 1.4 dB, noise bandwidth 1.5 (times Af306 Table 3.1. Comparison of window properties ~ ‘Noise Max. picket Window type | sidelobe (4B) bandwidth | fence eror, (x Af) (a3) Rectangular (133 100 Bs Hanning 32 —] Framing = Raiser-Bessel [69 Truncated Gaussian |-69 [Flatiop 35 <1 Figure 3.4. Worst case fer characteristics for some common window functions (a) Rectangular (b) Kaiser-Bessel (c) Hanning (d) Flautop 307 differing levels is probably the Kaiser-Bessel, but the same can usually be achieved by simply analysing with more resolution (200m analysis or a lauger transform), ‘The flattop window is specifically designed to minimise the picket fence effect, and is thus usually the best choice when calibrating measurements calibration signal whose frequency can fall anywhere between two analy can also be useful when analysing a signal dominated by one or more families of harmonics, since as long as they are resolved (keeping in mind that the noise bandwidth is 3.7 Af ) there is no need to compensate the indicated values of the various harmonics. Figure 3.5 shows how compensation can be made for both picket fence error and frequency error when using a Hanning window. Provided frequencies are stable along the record length, the difference in dB between the two highest samples around a frequency peak (AGB) determines the errors. As ‘mentioned above, the Hanning window has the desirable property thatthe effective dng in overlap averaging (see later) can be made completely uniform with an 0% overlap, the weighting varies by 2:1, but als. When finding the averaged spectrum of a long transient signal (longer than the transform size), @ uniform weighting is preferable. 3.2.2 Transient signals ‘Analysis of transient signals is common in impact measurements for modal analysis. The force signal is always short, and a rectangular window is suitable although this may be tailored to just more than the length of the force pulse in order to exclude noise. For the response signal it must be ensured that it has died away (ie by $0 to 60 dB) by the end of the record. This can be achieved by extending the record length (ie by zoom or a larger transform size), but where this is restricted by other constraints itis common to apply an exponential308 SPECTRUM LEVEL, os Boa é § | z 2 osf os 3 Baz os = 5 7 i é af. i: 2 on =< ose 9 + z 3 « 5 6° ace (d) 8 DIFFERENCE Figure 3.5 Compensation for the picket fence effect with a Hanning window. AdB = difference between two highest spectrum samples. AL= picket fence error. Af = frequency error (a), (b), (c) Minimum, maximum ‘and intermediate error cases (d) Error nomogram which attenuates the signal damping, which is known very precisely, and so can be subtracted from the resulting measurements, 3.2.3 Application in the frequency domain tion by a window function corresponds in the frequency domain to FT, this is sometimes the most efficient way to apply them, Examples of where this is advantageous are firstly where the FT of the window is very simple, such as with the Hanning function, secondly where only a part of the spectrum is required as with zoom spectra, and thirdly where several different windows can be applied to exactly the same FT. The basic principle can be explained using the Hanning window, which when repeated periodically (as icily with the DFT) can be represented as sin?@ or - a) happens 309 1 1 1 which has the convolution coefficients |-+ , + , 1] (keeping in mind [ 4°23 4] pine that the frequency corresponds to one period along the record length and thus t one line spacing. Convolution wi ients as are for a window with maximum value one, and are usually modified to scale the result (see below under "Scaling"). 3.3 Zoom Analysis The basic DFT transform of Equ. extends in frequency from zero to the Nyquist frequency and has a resol if jided by the number of samples N . Sometimes it is desired to analyse in more in a limited part of the frequency range, in which case use can be made of so- det Increase the length of record NV. Some analysers include this op form of “non-destructive zoom", which make use of an algorithm to perf transform of size mXN by combining the results of m undersampled transforms of size N. This was useful when hardware restrictions size of transform which could be performed, but in modern analysers, signal processing packages such as Matlab® there is virally no restriction ‘on transform size, and so zoom can be achieved by performing a large transform and then viewing only part ofthe result Reduce the sampling frequency f, . This can be done if the centre of the desired zoom band is shifted to zero frequency so that the zoom band around the centre frequency can be i frequency is then reduced accordingly Fig.3.6. The lowpass filtering and resampling process is usually done in octave (2:1) steps, as a digital Iways remove the highest octave, relative to the sampling frequency, and halving the sampling frequency simply means discarding every second sample. This type of zoom is normally done in ret time by a specialised hardware processor, the advantage being that the sampling rate is reduced before signals have to be stored. Its discussed below that the zoom process is @ useful precursor to demodulation, even where the further processing is to be done in a computer. Nove thatthe time signal output the zoom processor is complex, as the corresponding spectrum is no conjugate even.310 Signal | re bo na] ADC Analogue Baseband LP fiter Zoom pent | pt i Frequency | shitt resampling a aan Repeated octave steps Figure 3.6. Schematic diagram of FFT zoom process 34 Spectrum Averaging ‘The need for averaging of FFT spectra is determined by whether the signal Contains random components or not. Averaging should always be done in terms of signal power (ic amplitude squared) as itis this which is conserved independent of phase. The DFT spectra of discrete frequency components always have the same tude, and therefore achieved by averaging the squared amplitudes, although this can be useful for clarifying which components are discrete frequency and which are random. Where analysis is being done for diagnostic purposes (eg a zoom spectrum to measure a single frequency or family of sidebands very is preferable not to average, as in practice machine speeds vary th time and averaging results in a smearing of frequencies and in icular disguises frequency spacings, When meaningful spectra are to be obtained from random signals, which in ‘mechanical signals are typically caused by fluid flow (turbulence, cavi road roughness etc, it is necessary to average a number of power spectrum estimates, ‘The number of averages required is determined by the desired accuracy, as the standard deviation of the result for gaussian signals) is given by where 1 is the number of independent averages, Thus, for m 1.4, meaning that 95% probability th within +3 averages ete. ‘With a rectangular window, “independent” means non-overlapping, but with ‘ther windows such as Hanning, advantage can be gained by overlapping, as 31 information is lost near the two ends where the weighting is near zero. In fact, very by overlapping $0%, and is recommended for as many effective averages can be obtained 8 is not there the successive Hanning windows ing is Y%; and thus their power weighting Y, he final weighting ofthat part of the signal inthe overlap average is }, and the total power weighting along the signal varies between Y, and ty. This gives no problem for s given length of record, in particular by a factor of at least 3 although -y signals, but to extract all information from a is non-stationary, itis advisable to overlap th typical FET record lengths in powers of 2 it is often simpler to overlap by 3 . In the latter case the effective number of rl 4 averages (o insert in Equ.(3.1) is half the actual number. The Matlab® function PSD performs overlap averaging (as does CSD for multiple signals) 38 Sealing 1 Discrete frequency signals ing value Ag at the pos the corresponding sinewave (Ck). To obtain the equivalent RMS (root mean square) value, |Ay | must be multiplied by V2 to take account of the power in the C , negative frequency component. The RMS value is also equal to “t I the 8 quency comp 4 AR signal contains discrete frequency components which do not have an integer number of periods along the record length, then a window function. such as Hanning will generally be used to reduce leakage. However, if the window is scaled to a maximum value of unity, the average “power” (ie mean square value) of the signal will obviously be reduced, and itis necessary to compensate for this From section 3.2.3 it will be seen thatthe convolution coefficients for a Hanning ed ue of 2 14,-4 th window scaled toa maximum value of 2 ate | —>, 1, —— | meaning that 2 2 * a single component in one line would be replaced by three components of which the central one would have the same value be sealed correctly), This frequency lieved by scaling the Note that because ofthe extraneous sidebands introduced, the total “power” in the spectrum has been increased as discussed below, 3.5.2 Parseval’s theores ' theorem which in broad terms power (or energy) in a signal can be obtained by integrating frequency, and in both domains is related to amplitude squared. nary signal with finite power, the frequency spectrum will either contain discrete frequency components whose amplitude squared directly epresents the power at each frequency, or for random signals the squared amplitude spectrum is continuously distributed over frequency and represents “power spectral density” (PSD) which has to be integrated over a finite bandwidth to give finite power. In both cases the equivalent “power” in the time domain is the ‘mean square value, obtained by integrating the instantaneous squared value (instantaneous power) over a sufficien egral of “power” over tim« le of its Fourier transform represents “energy spectral density” (ESD) Which when integrated over all frequency gives the same total energy as integrating the instantaneous power of the signal ove Henceforth, the terms power an energy by an impedance or lectrical power = I R = VR in terms of current I , ‘Thus for a signal U (where U N cic) power has units U*, energy has units U's, ESD has units U'yHz (= U's) stance) ut the power in each spectral line can be assumed to represent the integral of the PSD over the frequency band of width Af (= 1/7), and thus the average PSD is obtained by multiplying the squared amplitude by 7. The required averaging over a qumber of records does not change t ing. How well the average PSD represents the actual PSD depends on of peaks valleys) in the spectrum. The width of such peaks determined by the damping associated with a structural resonance e ie broadband random signal, and the 34B bandwidth is given by twice the value of (of Eq ‘expressed in Hz. The PSD will be sut accurate if the 24B bandwidth is a ‘minimum of five analy: Ifa window such as Hanning has been used to reduce leakage, and i scaled so as (0 read the peak value of discrete frequency components (as recommended in 3.5.1) the calculated PSD value will have to be divided by the “noise bandwidth” ind ‘Table | to compensate for the extra power given by This noise bandwidth is the sum of the squares of the for Hann is 05? + 1 + OS? = 1.5). When 1B Over several frequency lines (eg to convert a constant bandwidth 313 spectrum to constant percentage bandwidth) the total power in each integrated band must be divided by the noise bandwidth ofthe window because ofthe exira power associated with each lin: this isthe same as integrating the PSD over the requted bandwidth Note that diserete frequency components cannot be represented on a PSD scale as they are concentrated in an infinitely narow bandwidth and thus have infinite PSD. Note also that their power is independent of the analysis bandwidth AF used to analyse them, whereas the power of spec Varies directly wit the analysis bandwidth. Zoom analysis is sometimes used t make discrete frequency components stand out from random background noise. ines of random signals 354 1 goals Transient signals are also tested a6 being one period of aperiodic si only does the power in a spectral line have to be converted to an average spect density by dividing by AY, but also the average power must be converted eneegy per period by a further m than the transform length and thus a rectangular window will be used, and if the signal has decayed to near zero at the end of the record (eg following the recommendations of 3.2.2) the signal bandwidth will be sufficiently greater than the analysis bandwidth for the average ESD to represent the true ESD, The extra damping given by an exponential window will genuinely give a reduction in signal energy, 4. OTHER SPECTRAL ANALYSIS TECHNIQUES, 4.1 Parametric Spectral Analysis With Fourier analysis, as is evident from Fig.3.1(d), the spectral resolution is determined by the Fourier transform of the time window which limits temporal resolution. For a window of length T the spectral resolution is of the order of UT, ‘and thus the better the time localisation the poorer the frequency loc vice versa. This is one expression of Heisenberg's uncertainty pi because no assumption is made about the behaviour of the time funt the window (effectively itis set to 2er0, which is extremely improbabl With parametric spectral anal its behaviour inside the window. Jal signals. Wi ical system described by a of parameters when excited by a unit white n Thus the frequency response of the system represents the signal spect he improvement in spectral resolution is accompanied by a deterioration in amplitude accuracy.314 4.1.1 MA models Perhaps the easiest case to understand is w (finite impulse response) Which cas of the input signal with the finite length impulse response of the by the equation: u w= Deere 1) & where x, represents the input sig ly, represents the output sigr |, and the 1b, represent the convolution weights or samples of the impulse response convolution eq the term MA model. Applying @ is the equivalent of a Laplace transform for diserete time signal 3m becomes the product: Ve) = Sort X= Bie) X(2) 42) from which the uaofr function can be seen to be B= Sac" =f -2"s) (43) ‘which has no poles and is thus an This type of model is obviously most efficient when the effective length of the impulse response is short, meaning that it is highly damped and thus without sharp spectral peaks, 4.1.2 AR models AR or “autoregressive” models are more efficient where there are sharp spectral peaks, and thus the required transfer function has poles. infinite impulse response) filters where outputs are generat Previous outputs and the current input. maship between input and output signals can be expressed as: w= La ya te (44) a recursively from the rinciple N —> > but can be truncated when the terms become After Z-transformation this gives: ¥(z)A(z) = X(z) 5) 31s from which the transfer function can be seen to be: 46) which has no zeros and is an all-pole model. ‘There are a number of techniques which result in such an AR model, one of which ination of any biasing effect. used in speech analysis, and st “autoregressio Figure 4.1 shows the results of applying maximum enttopy an: very short record of envelope signal from a bearing with an inner race fault {4.3} The record length comprised only 1.29 revolutions of the shaft speed which determined the spacing of modulation sidebands in the envelope spectrum. The ‘maximum entropy spectrum of Fig.4.1(4) appears to give very good resolution of the sidebands, but Fig4.1(c) shows that Fourier analysis can give almost as much information provided a sufficient degree of spectrum interpolation is used. The spectrum interpolation was achieved by padding the data record with zeros to 7 times its original length. Nore that the maximum entropy spectrum had to be represented on a logarithmic amplitude scale because of the much wider range of amplitude values than forthe Fourier analysis cases316 Spectam 4) gees onbpt oHEEEE Figure 4.1 Envelope spectra of short time recording (1.29 revolutions) just spanning the first two sections where the inner race fault passes the loading zone. [43] (@) Envelope signal after bandpass filtration in frequency range 2.7 ~ 3.3 kHz, (b) Envelope spectrum without interpolation, (c) Envelope spectrum with interpolation, (d) Maximum entropy envelope spectrum AR models have been found useful for characterising a signal with a ‘minimum number of parameters, ¢g to use as a feature vector for input to a neural network to distinguish different cases. Where signals are nonstationary, the concept. 317 of “evolutive” AR models [4.4] may be useful. In this case the AR parameters are updated by moving a finite length window along the record. In [4.4] an evolutive ‘AR model gave a better separation of faults in reciprocating machine vibration signals 4.1.3 ARMA models Where the spectrum contains sharp peaks and valleys, it may be preferable to model it with an ARMA model which contains both poles and zeros. The relationship between input and output signals is then: ve Shea Saree an for which the transfer function is: B(z)/ A(z) (48) ‘with both poles and zeros. In general it is much more onerous in terms of computational effort to fit an ARMA model than either an AR or MA model, although this may be compensated by a reduction in the number of parameters required to obtain a good fit. Note that software to fit these models is available in signal processing packages such as the Matlab® System Identification Toolbox. 4.1.4. Other models For vibration measurements made when the structure isi excited, “free decay” after being ‘natural to fit a number of complex exponentials corresponding to the impulse responses of the various modes. One method which achieves this is the ‘another is the Prony method ITD method (1.6] referred to in the described in [4.1]. These sho Section 4.1 discussed one appl model a system wl noise source. Dig both infinite and fii produce a measured signal san also be used to frequency je by measuring the power ly done using a ered by repet 1e properties of318 1) By varying the filter coefficients (either FIR o 2) By varying the sampl iven coefficients the frequency (the sampling frequency. Thus for UB-octave anal required (ie three sets of coeffi Nyquist frequency. The signal is sampled at a r ‘octave (ie the upper cutoff frequency of the upp frequency) and is passed through the three 1 filter c iple, three bandpass filters are octave from 33% to 66% of the corresponding to the highest -octave at 66% of the Nyquist ave filters and a digital lowpass frequency. The sample rate of the y discarding every second sample) after which passage through the same di Il produce the filter outputs corresponding to the next lower octave in actual frequency. By repeating t Process, as many octaves can be covered as desired. Note that by being able to he filtering calculations twice as fast as required for the upper octave, itis to process any number of octaves in real-time, as each time the sample halved only half as many samples have (o be processed in a given time. If fen (0 process the upper octave is taken to be unity the time taken to rocess all octaves is 1 ry i 22 pociealocmesf ste ty ty fet be post-processed would have to be extremely long if the analysis is in Ea output would have to be averaged over 16 x 80 ~ 1280 + sample rate, corresponding 10. 640,000 samples ofthe samples of the lowes: original record, 4.3 Order Tracking In analysing ro ions itis often desired to have an_x-axis based “orders” of shaft speed. This can be to avoid smearing due to mS oF can be t0 see how the strength of the various harmonics changes over a greater speed range, for example as they pass through various i amplitude signal which is synchronous with the rotation sampled a fixed number of times per revolution, the ishable from those of a sinusoid, and thus give a line spectrum, whereas if normal temporal sampling is sed the spectrum spreads over 8 range corresponding to the variation in shaft speed. Thus, for order analysi necessary to generate a sampling signal from a tacho signal synchronous 1; 319 1S possible to use a shaft encoder mounted on the shaft in rovide a sampling signal, but more often the latter has to be generated eked loop to track the tacho signal and then generate a specified number of sampling pulses per period of the tracked frequency. However, an analogue phase-locked loop has a finite response time and cannot necessarily keep up with random speed fluctuations such as occur with an internal combustion engine from cycle to cycle, ‘The best method signal. This can be done in a number of ways, based on di way is simply to increase the sample rate of each section by say a factor of I then select the nearest sample to the theoreti the sample rate by an integer factor can be achieved by inserting the appropriate number of zeros in between each actual sample, and then applying a di lowpass filter. to limit the frequency range to the original maximum, smoothing the curve (it will also require rescaling proportional to the resa factor). Resampling by a factor of 4 is illustrated in Fig4.2. The same be achieved using the FFT by padding around the Nyquist frequency) and then inver sided) spectrum to the same increased number of time samples. Note that the length in seconds is the reciprocal ofthe frequency line spacing in Hz which for a quadrat interpolated positions. The accuracy of the interpolation can be judged by ring that the interpolation in the time domain corresponds to a 1 in the frequency domain by a filter characteristic which aliases back into the measurement range, For example, choosing the nearest sample value is the same as convolving the original samples with a rectangular function of width equal to the sample spacing, while a linear interpolation correspends to a convel with a triangular function of base width twice the sample spacing (the convol of the rectangular function with itself) In the former case the filter charac a sinc(x) function with zeros at multiples of the sampling frequency (so that al istic is320 Figure «4.2. Digital resampling with four times higher sampling frequency(a)Signal sampled at fy and its spectrum (b) Addition of zeros which changes sampling frequency tof, (c) Lowpass filtration and rescaling sidelobes fold back imo the measurement range), while in the later case square of the sne(x) function, which has much smaller sidelobes, In practice, cubic interpolation involving two samples on each side of a central one gives good results without excessive computational effort [45] Quite apart from errors intoduced by the interpolation, when resampling at as a machine speed reduce ensure thatthe signal is adequately lowpass fillered to pre filtering can be useful here a8 the cutoff frequency varies ing frequency, but the i components do not is problem, as from Fig.4.2c) it can be seen thatthe samy can be reduced by a large factor before overlap occurs. In fact, the factor is 5.9 in this case because a tracking digital filer cutting off at a particular shaft order useful band in terms of frequency. es the use of tracking to avoid smearing inthe spectrum bration signal from a gearbox in a variable speed mining shovel. The frequency components in the spectrum after tracking come mainly from gearselated components which were removed using synchronous averaging as described in the next section. isthe S. TIME DOMAIN ANALYSIS. S.1 Time Synchronous Averaging Synchronous averaging is wseful to extract that part of a signal which is periodic with the same period as a trigger signal, eg a once-per-rev tacho signal from a shaft in a rotating machine. In practice itis done by averaging together a series of signal segments each corresponding to one period of the synchronising signal. Thus: 321 1 yeO=UNY ye +nT) G3) = This can be modelled as the convolution of y(t) with a train of N delta functions displaced by integer multiples ofthe periodic time T , which corresponds in the frequency domain to a multiplication by the Fourier transform ofthis signal, which can be shown to be given by the expression [4.1 CCF) =1IN sin(NaTf)/sin(n Tf) 3.6) oor Before Tracking 500 +1000 oor ®) After Tracking 500 1000 Frequency (Hz) Figure 4.3. Use of tracking to avoid smegring of shaft speed related components ‘The filter characteristic corresponding to this expression is shown in Fig.(6.1) for the case where N= 8, and is seen to be a comb filter selecting the harmonics of the periodic frequency. The greater the value of N the more selective the filter, and the greater the rejection of non harmonic components. The noise bandwidth of the filter is 1/N’, meaning that the improvement in signal/noise ratio is 10 logioN dB for additive random noise. For masking by discrete frequency signals, it should be noted that the characteristic has zeros which move with the number of averages, so tis often possible to choose a number of averages which completely eliminates a particular masking frequency. The above characteristic is for an infinitely long time signal y(t), and in Ref. {5.1322 shown that for the practical situation of sampling frequenc; ossible to calculate an optimum number of averages to completely remove a discrete masking signal, in particular when the frequency is related by a rational fr the synchronous frequency. This is always the case for different shafts in gearboxes. inte length of signal Figure 5.1. Filter characteristic corresponding to 8 synchronous averages (from 5.) For good results the synchronising signals should correspond exactly with ‘samples of the signal to be averaged, as one sample spacing corresponds to 360° ‘of phase of the sampling frequency, and thus 10 144° of phase at 40% of it which yypical maximum jgnal using a sampling frequency derived from the synchronising (tacho) signal, as described in Section 4.3, solves both these problems and is always to be recommended. Figure 5.2 shows the results of using synchronous averaging on the data of Fig.4.3. The order tracked data was arranged to have an integer number of samples per period of the low speed gear, which allowed determination of the harmonics of this gear speed by synchronous averaging. The spectrum of this signal is shown in Fig.5. repetition of this signal was subt the data was resampled to have ar high speed gear, after which its (Fig.5.2(b)). Finally remaining signal was dominated by the effects of at Fig.5.2(0)) Signal from the data, the et race bearing eating 323, Frequency (Hz) Figure 5.2. Application of synchronous averaging to data of Fig.4.3. ; (a) Spectrum synchronous with low speed gear (b) Spectrum synchronous with high speed gear (c) Spectrum dominated by bearing fault after effects of two ‘gears removed 5.2 Autocorrelation Analysis ‘The autocorrelation function measures how well a signal correlates with delayed versions of itsell. The equation for calculating it for a transient is Ry (t) = xox oar (53) while fora stationary function (or which the above integral would be infinite) is 7 RD = tim L fx@xcedar = Elx)xe+0)] eT an (34) For t= 0, Equ.(5.3) gives the otal e the mean square value or power, and it is common to normal for a stationary random function. This th power spectrum replaced by energy spectrum) by comparing Equ(S.3) with (2.17) where it can be seen that the324 ‘autocorrelation represents a convolution with the same function reversed in time, 0 that its spectrum is the product of the original spectrum with its complex conjugate, and thus is the squared amplitude or energy spectrum, Autospect Autocorrelation Sinysoid sm Mf Lt Bandlimited Noise i Io | Figure 5.3, Autocorrelation vs autospectrum for three signals. Note that ‘spectrum of (c) isthe convolution of (a) and (b) @ f Figure 5.3 uses this relationship to derive the autocorrelation for a sinusoid and a bandlimited noise, and then a narrow-band noise as its spectrum can be generated as the convolution of the first two. This shows that the narrower the ‘bandwidth of a noise signal, whether or not itis shifted from zero frequency, the ‘more its autocorrelation is spread out in time. On the other hand, for a white noise, whose spectrum extends uniformly to infinity, its autocorrelation is concentrated at tocorrelation function can be used to enhance periodicity, as it separates, broadband noise, and converts all sinusoids to cosines. This ‘means that periodic func! ‘and thus apparent, since the function can ned spreading just reproduce the ion at a lag corresponding to the echo delay time, As seen below, the 325 ccepstrum is a better echo detector, as in principle the echo delay is given by a delta function, independent of the bandwidth. Figure 5.4 shows a case where masking noise was reduced considerably in an envelope spectrum for a bearing fault, by clipping the autocovariance function rear zero time lag. The autocovariance is the autocorrelation performed after removing the mean value, which gives considerable distortion in an all-positive envelope signal. A further squaring of the autospectrum, corresponding to a further autocorrelation operation on the autocorrelation function. gives even better rejection of the noise but at the same time enhances large discrete peaks at the expense of small ones spectrum of () Spectrum from o na! eva 27 ] | \ He \ HAN 09 0 OG at i Lu i Figure 5.4. Use of CAF (clipped autocovariance function) (c) and its squared spectrum (d) to reduce noise effects 5.3 Cepstrum Analysis ‘The cepstrum is usually calculated from the spectrum, and often used to extract eated here as a time domain technique related (0 the autocorrelation function just n of the cepstrum is as the inverse Fourier 65)326 where X(f)=S[x(NFA(Nexpljoc/)) 66) in terms of its amplitude and phase, so that: log(X(f))=In(A(/)) + JOP) 6 When Xf) is complex as in this case, the cepstrum of Equation (5.5) is known as the “complex cepstrum’, although since In(A(f)) is even and O(f) is odd, the complex cepstrum is real-valued. When the power spectrum is used to place the spectrum X ig cepstrum, known as the “power cepstrum” or C(O =F f2in(Acp) 6.8) and is thus a scaled version of the complex cepstrum where the phase of the spectrum is set to zero. It can also be seen from Equ(S.5) and (5.8) that the difference between the cepstrum and the autocorrelation function is given by taking the log of the amplitude squared spectrum. Since forcing function and transfer function effects are multiplied in both the complex and power spectra, the log converts this to an additive relationship, which remains in the cepstrum. Moreover, the two additive components are often better separated in the cepstrum than in the spectrum, ‘There is yet another type of cepstrum known as the “differential cepstrum”, which is defined as minus the inverse Z-transform of the derivative of the log spectrum, or: cxo=-2-[ Logo] (59) and can be directly calculated from the time signal as: -i[ Sf x(n} C, este (=3 [Se 6.10) yuri transforms are to be interpreted as evaluating the Z-transform on One advantage is that itis not then necessary to unwrap the phase of the spectrum to a continuous function as is the case with the complex cepstrum, a7 Where the frequency spectrum Xf) in Equ(S.5) is a frequency response function (FRF) which can be represented in the Z-plane by a gain factor K and the zeros and poles inside the u zeros and (where analytical formulae: C(n) = In(K) sn=0 .n>0 6. sn<0 in terms of the time index n (known as quefrency). When grouped in complex conjugate pairs, these terms represent exponentially damped cosines, futher weighted by the hyperbolic function Un . For the differential cepstrum, the differentiation in the frequency domain results in a multiplication by nin the quefrency domain, giving exponentially damped cosines which are mathematica similar to modal impulse responses. For minimum phase FRFS, thee are no poles for zeros outside the unit circle, and the cepstrum becomes causal (positive quefrency only), so thatthe real and imaginary pats ofits Fourier transform (the log amplitude and phase of the spectrum, respectively) are related by a Hilbert transform, This means that the phase does aot have to be measured or unwrapped, a the complex cepstrum can be obtained ftom the real cepstrum by multiplying by imes the Heaviside function Hi 1 cepstrum was originally proposed by Bogert et al {5.2} to detect ‘ean be shown that an echo manifests itself in both the log ampl ion and thus transforms to a series of functions (known as “rahm« the cepstrum. Figure 5.5 and shows how they over function, This leads to one set of applications for the cepstrum, P the ori328, saree ised A plod Ie Met “sana ‘area | Ein] [een Figure 5.5. Use of complex cepstrum to remove echos the logarithm of the spectrum often enhances the pattern of equal ‘components, such from gearboxes, where on a spaced harmonics and sidebands, in for example signals very largest would be visible, The cepstrum conden: (© a sinall number of rahmonics and makes for easier comparison [ hhas been shown that the two sets of rahmonies produced by two gears in mesh are complementary in dat the sum of the values of in each series is a constant equal to 0.5 in the absence of noise, less with noise [5.4]. If the condition of one gear deteriorates its ‘component will increase, but is accompanied by a corresponding decrease in the other. This leads to a very general and sensitive indicator of change which is only dependent on the signal/noise ratio. This research has also led to a technique to detect and locate spalled gear teeth [5.5] Because the forcing function and transfer function are additive in the response cepstrum, at least in SIMO sit possible to separate them in response measurements if the log spectrum of the excitation is smooth and flat, so 329 that the corresponding cepstrum is short. In [5.6] it was shown how the poles and. zeros of the FRF could be extracted from the response cepstrum, in a windowed range of queffency with negligible force effect. To use this information to reproduce the FRF requires an overali scaling factor, and an equalisation curve to compensate for the effects of unmeasured out-of-band modes. In [5.6] it was these could be determined from an initial measurement, or a reasonably fe element model, even when actual natural frequencies were different response cepstrum with a scaling and equal and another where ihis was further extended by tracking the changes du to mi (5.6) the measurements were made Figure 5.6. FRFs regenerated from response cepstra (full) compared with direct ‘measurement (dashed) for impact measurements on a beam. (a) Using an FEM ‘model for scaling and equalisation (b) After milling a sto ‘beam (note reduction in frequency of odd-numbered modes} 6. OTHER SIGNAL ANALYSIS TECHNIQUES al, cease itis known, lered as a frequency n isthe difference in instantaneous frequency (angular constant carrier frequency. Thus, frequency modulation is the derivative of phase modulation, A direct mechanical example of phase/frequency modulation is shaft torsional vibration, which when expressed in terms of shaft angle is a phase ‘modulation, and when in terms of shaft speed is a frequency modulation. There is ‘no modulation term for the angular acceleration obtained by further differentiation from the330 A mechanical example of amplitude modulation is the variation in vibration amplitude at the meshing frequency in a gearbox, as the increase in tooth deflection load gives an increasing departure from ideal involute profiies, and often tooth load varies periodically with the rotation of the gears, ‘The signals produced by faults in rolling element bearings are a series of hhigh frequency bursts as resonance frequencies are excited by near periodic impacts. The diagnostic information is contained in the repetition frequency, not in the resonance frequen J, but spectra obtained by direct Fourier analysis are dominated by the latter, and the important information is disguised by smearing of the high order harmonics. Such signals can be modelled as an amplitude ance frequency by a near periodic series Thus, a generally modulated signal can be represented by A, (1) cos(2mf,t +O, (1)) 1) where A,,(t) represents the amplitude modulation function and y(t) represents the phase modulation function in radians. The corresponding frequency modulating funct Expression (6.1) will be seen to be the real part ofthe rotating vector: Aq ()expLi(2nf.t+o,(0)} 62) 1 corresponding imaginary part so as to form the complex expression (6.2). Provided the fluctuating part of 6.2) An Dexplin (Oh (63) has a half bandwid be one-sided, and (6. imaginary partis the 2.3 cam be used. As the spectra of the cartier frequency f., the spectrum of (6.2) will mn this case the required and the methods of Section {wo parts are convolved, the total bands is less than the sum ofthe individual bandwidths. The bandwidth ofthe amy that of A,,(f), and even though that of exp{jd,,(‘)}is the maximum phase de ss than 1 radian, the inthe dynamic range) is less than twice that of ®,(#) sie cessor can be used directly both to ext ‘and to remove the carrier component unwrapping. san be achieved using FFT -s, Where phase demodulation is required, the centre of I have to be shifted to zero frequency (and negative the other end of the frequency record). However, the negative frequency side with zeros, thus maintaining an analytic signal, In either case there should be at least as many contiguous zeros in the spectrum as components, since the modulus is the square root of the amplitude squared, and the latter corresponds to the convolution of the spectrum with its complex conjugate reversed end-for-end. The zer0s prevent extraneous wrap-around errors. re 61 Black sift procedure for selecting freuency band for demodulation (ar Original specrum of onesied benpass econ () Prequeey shi By (ceanatefpaband)(e) Frequency ah by mount corespondng 10 lowe postband limit (half sce ) “Rrequeney hit by amour corresponding to lower pasiband inl ul sie transform)332 6.2 Cyclostationarity and Spectral Correlation Many machine generated signals are almost but not completely periodic, even after for small speed changes by order tracking. One example are the ls from rolling clement bearing faults, where even though impact frequencies can be calculated on the basis of pure rolling amount because of random slip. signal from an internal combustion engine, as even though some events such as piston motion and valve operation are completely periodic, the combustion pressure signal varies randomly from cycle to cycle. In such cases Synchronous averaging cannot be used, at least without losing the variable part of the signal. However, such signals are often cyclostationary [6.1] meaning that their second order statistics, such as autocorrelation functions, are periodic. M9 ofthe autocorre ion, varies with both ¢ and 't,and itis posible fe nak Fourier transforms with respect to each By the Wienot Rniseng with respect to. ives autospecta agate np respect t0 gives the spectral corlaton a ignals, there isa peak ony for. ~ 0: mt for will be peaks also for frequency shifts where cach others an example, here two different de modulated by the same artow-band noise Signal, be a petet conelacon for a fomucacy shit corespoudng te ie sepatation of the two carr frequencies, giving Very narrow Peak mie och frequency direction, eventhough the width of peaks im the formal eqsoney direction is detcemined by the bandwidth of the modulating nose fencing Ref {6.1 is an exelent dtcssion ofthe various apy wil [62] and [63] desert the application othe detestiono ex ol 6.3 Time-Frequeney Analysis In theory the Fourier transform requites integration over ware that the ear can detect changes in frequency wi hat 50 an analysis technique has been sought which matches the ear changing frequency patterns. A simple approach is to move a shot along the record and obtain the Fourier spectrum as a fur However, the uncertainty principle means 1h useful resolution. It is described by frequency wi form 333 S(f.0) = J x(a) w(r—a)exp(j2ne) dt 4) where w(t) is a window which is moved along the record. Norma vs frequency diagram, in low could be of finite such as a gaussian 2 ide squared |S(f;1)| is displayed on a which case itis sometimes known as a spectrogram. The length such as a Hanning window, or theoretically in ‘window, but in practice of course it must be truncated. 63.1 The tribution ‘The Wigner-Ville distibution (WVD) appears to violate the uncertainty principle in appearing jon than the STFT, but suffers from interference ‘components between the actual components. The original Wigner distribu was modified by Vi 1] who proposed the analysis of the corresponding. analytic signal so a terference between positive and negative led “Cok ns [6.4], most of which have been fay. Even the STFT falls into this class. Cohen’s class may be represented by the formula: C.F,0) = SRO} 65) yonent frequency co where R(t;1) is a weighted autocortelation-like function defined by: renin Fur dee $ ews 66) and (14,7) is a kernel function used to smooth the WYD (wi is ob signal from WVD and the smoothed pseudo- ‘a portion of a diesel engine cycle, and ast in this case the smoothing yn in both’ directions that is better than the STFT i suppressing the major interference components. In Ref {6.5] the made (0 use various smoothing techniques to locate the interference ‘and then remove them from the unsmoothed WVD to retain optimum resolution,334 Time ) Figure 6.2.Comparison of time-frequency distributions for a diesel engine vibration signal (a) STFT (b) Wigner-Ville distribution (c) Smoothed pseudo figner-Vi 6.3.2 Wavelet analysi Another approach to time-frequency analysis is to decompose the signal in terms of | 8 family of “wavelets” which have a fixed shape, but can be shifted and dilated in time. The formula for the wavelet transform is: 17 a maid wove (2a (6 where W(t) is the mother wavelet, translated by b and dilated by factor a. Since s @ convolution, the wavelets can be considered as a set of impulse responses ters, which because of the dilation factor have constant perc es. In principle, they are not very Prop. zero phase shi Ss have described their use for detect 2 local faul bearings (eg [6.6, 6.7) in gears and 6.3.3. Other time-frequency techniques ‘A number of other time-frequency techniques have been proposed, inch evolutive AR methods mentioned in Section 4.1.2, and time-frequency “worms” mn area where considerable development is going on, One problem 335 with time-frequency representations is the problem with interpreting the results. Even if a trained eye can see the desired effect, quite advanced image processing is, sometimes required to extract the information. 7. MULTIPLE CHANNEL ANALYSIS For system an normally necessary to measure both input and output signals at the , and process them in pairs to obtain transfer fun information, such as FRFs. With multiple inputs and/or output to process more than two si bby matrix methods, in other speci 7A The Cross Spectram The cross spectrum between two signals is obtained by mult fone by the complex conjugate of the spectrum considered as the input, since the phase of the rest where By the Wiener-Khinchin relationship, the cross spectrum inverse transforms tw the cross correlation function: gant) =F (12) which reverts to the autocorrel identical )) when the signals are 7.2 The Frequency Response Function (ERE) In the frequency domain the FRF is basically the ratio of output over input, o: Gah) Haa(f)= 13) aB(F) Ga . 73) the presence of nose a better estimate can be obtained as described in other papers. If noise is primarily located in the output signal (over which the experimentalist has less contol, the procedure is modified by multiplying spectrum, and numerator and denominator by the complex conjugate of the averaging, to give: FGoG4*) _ Gap HGaAGa*(N] Gaal) AY(f)= ay since averaging reduces the effect of noise on the cross spectrum. If noise is336 primarily located in the input signal, a better estimate is given by: EiGg(f)Gg*(f)| G, yt =H GaSe A] Gaps) Hea A G+] Goa) 75) ‘The system impulse response function can be obtained by inverse wansformation of the FRF. 7.3 Coherence ‘The coherence gives a measure of the degree of linear relationship between two signals as a function of frequency. It is calculated by the formula: Gaal p (f= TO” Ga Cos 6) ‘This has values between zero and one depending on the degree of linearity. For a single estimate, it will always be one, since |Gap|=|Gq|-|Gp| and in the absence of noise and nonlinearity each estimate of Gap will have the same amplitude and orientation so thatthe average value will not change. However, in the presence of noise and/or nonlinearity the various estimates will change in length and particularly phase, so thatthe average obtained by vector summation willbe less than if they were aligned. I there is no linea relationship the various estimates will have random orientation and the average will end to zero Ie is possible for signals to have partial coherence (7.1), but © determine this requires. muliple channel processing and is beyond the scope of this contribution 74 Separation of Sources This is an area of signal processing which is receiving considerable attention and will presumably result in far-reaching consequences in the next millennium. So far there has been little application to mechanical problems, but if the type of result Which has already Seen achieved in, for example, communications can be translated to structural dynamics, there isa strong likelihood that it will be possible to detennine modal properties of structures fom response measurements only. Such results obtained in service would give the values which actually apply in each situation. In Section 5.3 a case was mentioned where it was possible using singular value decomposition o separate out the response to the largest forcing func provided it was atleast four times larger than the next largest. Ths isa limitation given by "second order statistics”. However, in Refs,7.2-7.4] it is demonstrated that it should be possible to separate responses 10 different sources, even of comparable strength, if their higher order statistics are different. 337 References LI Cooley, SW & Tukey JW, (1965) “An Algorithm for the Machine Calculation of Complex Fourier Series”, Math, Of Comp., 1990), pp297- 301 1.2 Burg, JP, (1975), Maximum Entropy Spectral Analssis, PhD Dissertation, Stanford University, Stanford, CA, USA. 1.3. Wigner, EP, (1932) "On the Quantum Correction for Thermodynamic Equilibrium.” Physical Review, 40, pp749-759. 14 Ville, J, (1948) "Théorie et Applications de la Notion de Signal bles et Transmission, 2 A, pp61-74, 1.5 Newland, DE, (1993) An fnisaduction to Random Vibrations, Spectral and Wavelet Analysis. 3 Ed,, Longmans Sciemittc, Harlow, UK 1.6 brahim, SR (1984) “A Modal Identification Algorithm for Higher accuracy AIAA Proc, 25th Structures, Structural Dynamics & Materials Conf.Paper 84-0928, pp1 17-122. 21 Randall, RB (1987), Frequency Analysis, 3% ed, Bruel & Kjaer, Naerum, Denmark, 3.1 Randall, RB (1996) “Vibration Measurement Instrumer “Spectrum Analyzers and their Use”, Ch.ld in Shocl Handbook, (Ed. CM Harris), McGraw-Hill, NY, 4.1 Braun, S & Hammond JK (1986) “Parametric Methods” in Mechanical Signature Analysis, (Ed. 8. Braun), Academic Press, London, 42 Robinson, EA (1982) “A Historical Perspective of Spectrum Estimatior Proc.IEEE 70(9), 885-907. 43 Gao.¥, Randall RB & Ford, R (1998) “Estimation of Envelope Spectr using Maximum Entropy Spectral Analysis and Spectral Interpolation”, int J. of Comadem, 1(3), pp 44 Bardou.O & Sidahmed, M (1994) Exhaust and Discharge Sys th 13, and Vibration arly Detection of Leakages in the 3s of Reciprocating Machines by Vibration Analysis” Mech. Processing, 8(5), ppS51-570 4.5 McFadden, PD (1986) “Interpolation Techniques for the Time Domain Averaging of Vibration Data with Application to Helicopter Gearbox Monitoring”, DSTO-ARL Report No.AR-004-488, Comm. of Australia 5.1 McFadden, PD (1987), "A Revised Model for the Extraction of Periodic Waveforms by Time Domain Averaging”, Mech Systems & Signal Processing, 1(1), pp&3-95. 5.2 Bogert BP et al (1963) “The Quefrency Analysis of Time Series for Ec! Cepstrum, Pseudo-autocovariance, Cross-cepstrum and Saphe Cracking Proc. Symp. Time Series Analysis, (Ed. M Rosenblatt) Wiley, NY, pp209- 243, 5.3. Randall,RB (1982/3) “The Application of Cepstrum Analysis in Gear Diagnostics” Maintenance Management Int. 3. ppl338 34 58 56 3a 6 62 63 64 65 66 67 68 7 12 73 EI Badoui M, Guillet, F & Danitre, J (1998) “Energy Cepstrum for the Diagnostic of Complex Gears” (in French). 3rd Int. Conf. on Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques, CETIM, Senlis, France, pp291-300. Cahouet V, El Badaoui M. Velex P, imulations Numérigues et Détec d'Engrenages Cylindriques", 4" World Congress on Gears and Power . Paris La Défense, 16-18 March. RB (1996) “Determination of Frequency Response Functions from Response Measurements."Part I: “Extraction of Poles and Zeros from Response Measurements’ Mechanical Systems and Signal Processing, 10(3), May 1996, pp293-317.Part TI: "Regeneration of Frequency Response Functions ftom Poles and Zeros”, pp319-340 Randall, RB, Gao, Y & Swevers, J (1998) “Updating Modal Models from Response Measurements”. ISMA23 Conference, KUL, Leuven, Belgium, pp1153-1160. Gardner, WA (Ed), (1994) Cyclostationarity in Communications and Signal Processing, IEEE Press, NY. Capdessus. C, Lacoume, JL & Sidahmed, M (1994) “Cyclostationarty: a New Signal Processing Concept for Vibration Analysis and Diagnostic” 1994 Int. Gearing Conference, Newcastle, UK, Mechanical Engineering Publications, London. Dalpiaz. G, Rivola, A & Rubini, R (1998) “Gear Fault Monitoring Comparison of Vibration Analysis Techniques”, 3rd Int. Conf. on ‘Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques, CETIM, Senlis, France, pp623-637 Cohen, L (1995) Time-frequency Analysis, Prentice-Hall, NI 1 M_& Favre, B, (1993) “Engine Noise Characterisation with Time-Frequency Analysis", Mech. Systems & Signal Processing, (3), pp375-400. Lin, ST & McFadden, PD (1995) “Vibration Analysis of Gearboxes by the Linear Wavelet Transform.” Second Int. Conf. on Gearbox Noise, Vibration and Diagnostics IMechE, London Staszewski, WJ & Tomlinson, GR (1994) “Application of the Wavelet Transform to Fault Detection in a Spur Gear.” Mech. Systems & Signal Processing 8(3) pp289-307, May. Léonard, F (1998) “Object Transform and Time-frequency Worms” (in French), 3rd Int. Conf. on Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques, CETIM, Senl Bendat JS & Piersol, AG (1993) Engineering Applications of Correlation ‘and Spectral Analysis. 2nd Ed., Wiley, NY. Capdeville, C, Servigre, Ch & Lacoume, JL (1995) “Blind Separation of Sources in the Frequency Domain” Proc. ICASSP95, pp2080-2083. Capdeville. C, Serviére, Ch & Lacoume, JL (1996) “Blind Separation of Wide-Band Sources: Application to Rotating Machine Sources” EUSIPCO, p.2085-2088, 4 F & Daniére J (1999) varies sur les Dentures 14 339 Gaeta, M, Harroy, F & Lacoume, IL. (1992) “Utilisation de Statistiques 4Ordre Supérieur pour la Séparation de Sources Vibrantes, Ist int. Conf. on Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques, CETIM, Senlis, France, pp523-533.