1 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Domain Analysis 2 Signals – Frequency Domain Analysis V. Rouillard 2003 • Concerned with analysing the frequency (wavelength) content of a process • Application example: Electromagnetic Radiation: • Represented by a Frequency Spectrum: plot of intensity vs frequency 3 Signals – Frequency Domain Analysis V. Rouillard 2003 Solar Radiation Spectrum 0.25 Solar radiation spectrum (≈1000 W @ equator @noon) 0.1 0.05 Visible 700 nm red 0.15 440 nm violet 2 Irradiance [W/cm /µm] 0.2 Near IR Infrared 0 0 500 1000 1500 Wavelength [nm] 4 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Vibration on a transport vehicle 2000 5 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Whale sound 6 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Apache helicopter flyover 7 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Ocean surface level fluctuations (waves) 8 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Axle load of passenger vehicle on test track 9 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: Vibration of structure due to explosion load (pyrotechnic shock) 10 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: 11 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: 12 Signals – Frequency Domain Analysis V. Rouillard 2003 • Why frequency domain analysis? • Consider the following processes: 13 Signals – Frequency Domain Analysis V. Rouillard 2003 • All continuous signals can be shown to be comprised of a summation of individual harmonic (Fourier) components of various frequencies, amplitudes and phases. • Common methods to compute the frequency spectrum of measured data: • Filter analysis consists of a number of band-pass frequency filters (analogue or digital) • These filters allow only a (narrow) band of frequencies to pass • A number of filters with adjacent frequency bands are used to generate a frequency spectrum Filter (octave band) analysis 0 -10 -20 dB -30 -40 -50 -60 15 31 62 125 250 500 1k 2k 4k 8k 16k (fractional) Octave bands 14 Signals – Frequency Domain Analysis V. Rouillard 2003 Digital Fourier Transform (DFT) or Fast Fourier Transform (FFT) • Transforming a signal from the time to the frequency domain can be achieved via the Fourier Transform (also called Fourier Analysis). • Information is neither gained or lost when transforming signals from the time domain to the frequency domain via the FFT. • The Fourier transform is reversible – Inverse Fourier Transform (IFT) • Although Fourier theory applies strictly to periodic signals, the periodicity of sampled or measured signals is assumed resulting in a estimate of he frequency spectrum Magnitude Spectrum Signal (real) Complex FFT Phase Spectrum IFT 15 Signals – Frequency Domain Analysis V. Rouillard 2003 Random signals (Dadisp: Freqa) • Broad-band signals contain more sinusoid components than narrow-band signals • A sinusoid can be considered as a very-narrow-band random signal. • The phase of the sinusoids which make up random signals are values uniformly-distributed between 0 and 2π. • The phase of the sinusoids which make up pulse signals are equal. Effect of phase on frequency spectrum • 16 An infinite number of sinusoidal components with equal phase produces the Delta function (very sharp pulse). Signals – Frequency Domain Analysis V. Rouillard 2003 Signal bandwidth examples: Uniformly-distributed random phase. 17 Signals – Frequency Domain Analysis V. Rouillard 2003 Signal bandwidth examples: Zero bandwidth signals 18 Signals – Frequency Domain Analysis V. Rouillard 2003 Signal bandwidth examples: Constant phase. 19 Signals – Frequency Domain Analysis V. Rouillard 2003 Example of constant phase signal: Gaussian wave packet Elevation [mm] 40 20 0 -20 -40 0 2 4 6 8 10 12 14 16 Time [sec.] 3.5 Energy Spectrum [mm 2 .s] 80 Phase [rad] 60 40 20 0 0.0 -3.5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 0.0 0.2 0.4 Frequency [Hz] 20 0.6 0.8 1.0 1.2 1.4 1.6 Frequency [Hz] Signals – Frequency Domain Analysis V. Rouillard 2003 Digital Fourier Transform (DFT) or Fast Fourier Transform (FFT) Random signals (Dadisp: Apache_PSD & Spectral_averag) • Each observation is unique – sample of process – one physical realisation of the process • The frequency spectrum of each sample is an estimate of the frequency spectrum of the entire process • The estimate of the true frequency spectrum is improved by computing spectral averages. Important issues when computing the Fourier Transform: • Bandwidth: Frequency range to be analysed • Frequency Resolution [Hz] = 1/sub-record duration [secs] • Spectral estimate accuracy (random error): Std. Deviation of error = 1/√# averages. Spectral error is reduced by: • • Identifying sub-records within the measured record • Computing the spectrum of each sub-record • Computing the average spectrum Given a fixed record length, a compromise has to be reached with respect to frequency resolution and spectral error. 21 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of frequency resolution & spectral averaging. Example: Helicopter fly-by (sound) 22 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of frequency resolution & spectral averaging. Example: Heavy vehicle vibrations. 23 Signals – Frequency Domain Analysis V. Rouillard 2003 • Spectral averaging must be used carefully • When signals are strongly non-stationary (ie. Evident variations in vital characteristics such as RMS levels or frequencies) spectral averaging will conceal these non-stationary properties. 24 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of frequency content variation. Example: Whale cry. 25 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of frequency content variation. Example: Whale cry. 26 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of frequency content variation. Example: Whale cry. 27 Signals – Frequency Domain Analysis V. Rouillard 2003 Nyquist Frequency and Sampling Rate (Dadisp: Shannonsine) Leakage and the effects of widowing functions (Dadisp: leakage_sin & leakage_rnd) Overlapping Zero-padding 28 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of spectral leakage and windowing. Example: sinusoid. 29 Signals – Frequency Domain Analysis V. Rouillard 2003 Effects of spectral leakage and windowing. Example: Heavy vehicle vibrations. 30 Signals – Frequency Domain Analysis V. Rouillard 2003 Influence of signal clipping on frequency spectrum No clipping Clipped 31 Signals – Frequency Domain Analysis V. Rouillard 2003 Influence of broad-band and narrow-band (power line) noise on frequency spectrum Clean spectrum + Broad-band noise + Narrow-band noise 32 Signals – Frequency Domain Analysis V. Rouillard 2003 Influence of intermittent noise (switchgear interference) on frequency spectrum Clean spectrum + Intermittent noise (sharp pulses) 33 Signals – Frequency Domain Analysis V. Rouillard 2003 System Analysis (Excitation – Response Relationships) 34 Signals – Frequency Domain Analysis V. Rouillard 2003 • Frequency analysis is useful in determining the frequency characteristics of systems: relationship between output and input as a function of frequency. • Real systems are often assumed to approximate an ideal system. • Ideal systems: • • Have constant parameters (no variation in system characteristics wrt time) Are linear (ie. additive and homogeneous): • Additive: Response (output) to sum of excitations (inputs) = sum of responses due to each individual input: f ( x1 + x2 ) = f ( x1 ) + f ( x2 ) • Homogeneous: Response from excitation x constant = response x constant from excitation: f ( kx ) = kf ( x ) 35 Signals – Frequency Domain Analysis V. Rouillard 2003 Cross Spectrum A( f ) = Y • Excitation signal ∞ {a( t )} = ∫−∞ a( t )e− j2π ft dt Response signal H(f) B( f ) = Y System FRF ∞ {b( t )} = ∫−∞ b( t )e − j2π ft dt A( f ) = Re( f ) + i Im( f ) The cross spectrum of A wrt B is defined as: A = A e j( 2π ft +Φ A ) , A* ( f ) = Re( f ) − i Im( f ) B = B e j( 2π ft +Φ B ) S AB ( f ) = A* ( f ) ⋅ B( f ) = A e − j( 2π ft +Φ A ) ⋅ B e j( 2π ft +Φ B ) = A ⋅ B e j( Φ B −Φ A ) • Where A*(f) is the complex conjugate of the instantaneous spectrum of a(t) and B(f) is the instantaneous spectrum of b(t) • The amplitude of the cross spectrum is the product of the two amplitudes • The phase of the cross spectrum is the difference between the phase of B relative to A. • The cross spectrum SBA has the same amplitude but opposite phase. • Auto spectra and cross spectra are generally expressed in one-sided form: 36 Signals – Frequency Domain Analysis V. Rouillard 2003 Cross Spectrum A( f ) = Y • Excitation signal ∞ {a( t )} = ∫−∞ a( t )e − j2π ft dt Response signal H(f) B( f ) = Y System FRF ∞ {b( t )} = ∫−∞ b( t )e − j2π ft dt The auto spectrum is obtained in the same way: S AA( f ) = A* ( f ) ⋅ A( f ) = A e − j( 2π ft +Φ A ) ⋅ A e j( 2π ft +Φ A ) = A 2 • The autospectrum is the power spectrum which has additive properties useful for averaging. • Auto spectra and cross spectra are generally expressed in one-sided form: G AA( f ) = 0 G AB ( f ) = 0 G AA( f ) = S AA( f ) G AB ( f ) = S AB ( f ) G AA( f ) = 2S AA( f ) G AB ( f ) = 2S AB ( f ) • f <0 f =0 f >0 As for the auto spectrum the cross spectrum of stationary random signals is best estimated by averaging over a number of records. 37 Signals – Frequency Domain Analysis V. Rouillard 2003 Coherence • The coherence is used to determine the level of linear dependence b/w two signals as a function of frequency • The coherence is defined as: γ 2( f ) = • 2 G AB ( f ) G AA( f ) ⋅ GBB ( f ) And can be viewed as a squared correlation coefficient which quantifies the linear relationship between two variables: σ xy σ x ⋅ σ y 2 2 ρ xy <1 2 ρ xy <1 2 ρ xy =0 Covariance ρ 2xy = • 2 ρ xy =1 Variance Note that for a single set of records (no averaging) the coherence is 1. 38 Signals – Frequency Domain Analysis V. Rouillard 2003 Coherence • In practical cases, reasons for obtaining coherences of less that unity are: 1. Contamination of either excitation or response signal with noise 2. The presence on nonlinearities in the relationship between the excitation and response 3. Spectral leakage due to insufficient resolution or unsuitable windowing function 4. Time delay between the excitation and response signals Signal:noise ratio • The coherence can be used to determine the signal:noise ratio of the measurement which is defined as: S:N = • γ2 1−γ 2 If noise contamination of the response signal is assumed to be the only factor influencing the coherence, then γ2 is proportional to the coherent power while (1- γ2) represents the non-coherent power which is due to the noise in the response signal. 39 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function • The system is excited with a signal of suitable bandwidth (Swept sine [not suited to FFT], band– limited random signal or impulse) and the system response is measured • The excitation and response signals are each transformed to the frequency domain (FFT) to obtain a complex, instantaneous spectrum which are averaged to produce a mean power spectrum (autospectrum). • Further frequency-domain functions such as the cross-spectrum are then computed and the (linear) relationship between the excitation and response signals as a function of frequency is established giving the Frequency Response Function (FRF). a( t ) 8Y A( f ) H(f) System FRF b( t ) = a( t ) ∗ h( t ) 8Y B( f ) = A( f ) ⋅ H( f ) Relationship between excitation and response for an ideal system without noise contamination • In the time domain the response is obtained by convolving the system impulse response function h(t) with the excitation. • In the frequency domain, the response spectrum is obtained by multiplying the system Frequency Response Function H(f) with the excitation spectrum. 40 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function a( t ) 8Y A( f ) H(f) System FRF b( t ) = a( t ) ∗ h( t ) 8Y B( f ) = A( f ) ⋅ H( f ) Relationship between excitation and response for an ideal system without noise contamination • For an ideal system without noise, the Frequency Response Function can be determined by: H( f ) = B( f ) A( f ) A = A e j( 2π ft +Φ A ) B = B e j( 2π ft +Φ B ) • H( f ) = B j( 2π ft +Φ B ) − j( 2π ft +Φ A ) e ⋅e A H( f ) = B j( Φ B −Φ A ) e A Which is a complex function in terms of the magnitude ratio and the phase difference. 41 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function • In practice it has been shown that when the response signal is contaminated by broad-band random noise (transducer noise), a better estimate of the FRF is obtained by normalising the cross spectrum by the auto spectrum of the input: H( f ) = • If the complex conjugate of the output spectrum is used the resulting FRF is improved when noise contamination is present in the excitation signal: H( f ) = • B( f ) A* ( f ) S AB ( f ) G AB ( f ) ⋅ = = = H 1( f ) A( f ) A* ( f ) S AA( f ) G AA( f ) B( f ) B* ( f ) S BB ( f ) GBB ( f ) ⋅ = = = H 2( f ) A( f ) B* ( f ) S BA( f ) GBA( f ) It is interesting to note that the ratio H1:H2 always gives the coherence: 2 G AB ( f ) H 1( f ) G AB ( f ) GBA( f ) G AB ( f ) G AB * ( f ) = ⋅ = ⋅ = H 2 ( f ) G AA( f ) GBB ( f ) G AA( f ) GBB ( f ) G AA( f ) ⋅ GBB ( f ) H 1( f ) = γ 2( f ) H 2( f ) • 42 Although the magnitude of H1 and H2 will be different for noise contaminated measurements, their phase is always the same. Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function – effects of noise • Consider an ideal system where the measured response signal b(t) is contaminated by extraneous uncorrelated noise n(t). • The noise signal may include some component generated by the system but not caused by the excitation signal a(t). n(t) Given sufficient averaging, the measured cross spectrum GAB will approximate the true cross v(t) h(t) spectrum GAV and the measured response auto a(t) b(t) Σ spectrum is the sum of the clean output and H(f) the uncorrelated noise: • G AB ( f ) = G AV ( f ) 2 GBB ( f ) = GVV ( f ) + GNN ( f ) = H( f ) G AA( f ) + GNN ( f ) G AB ( f ) = G AV ( f ) + G AN ( f ) = G AV ( f ) = H( f ) ⋅ G AA( f ) H 1( f ) = G AB ( f ) G AV ( f ) = = H( f ) G AA( f ) G AA( f ) ← Optimum FRF (effects of noise minimised) 43 Signals – Frequency Domain Analysis V. Rouillard 2003 n(t) Frequency Response Function – effects of noise v(t) h(t) a(t) H(f) Σ b(t) while 2 H 2( f ) = GBB ( f ) GVV ( f ) + GNN ( f ) H( f ) G AA( f ) + GNN ( f ) = = GBA( f ) GVA( f ) H * ( f ) ⋅ G AA( f ) 2 = H( f ) + H( f ) GNN ( f ) GNN ( f ) = H( f ) + H * ( f ) ⋅ G AA( f ) H * ( f ) ⋅ GVV ( f ) G ( f ) H 2 ( f ) = H( f ) 1 + NN GVV ( f ) ← FRF magnitude overestimated (phase OK) where Gvv is the coherent output power spectrum and 44 GNN ( f ) is the noise : signal ratio. GVV ( f ) Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function – effects of noise • When the measured excitation signal is contaminated by extraneous noise m(t) (measurement noise), it can be shown that: u(t) h(t) H(f) m(t) H 1( f ) = H( f ) GMM ( f ) 1 + G ( f ) UU Σ b(t) a(t) ← FRF magnitude underestimated (phase OK) while H 2 ( f ) = H( f ) • ← Optimum FRF (effects of noise minimised) When there is extraneous noise in both the measured excitation and response signals H1 and H2 give the lower and upper bound to the true FRF. 45 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Examples: 46 Excitation: Wave height System: Ship Excitation: Gusts System: Tower Excitation: Pavement topography System: Road vehicle Response: Vertical vibration Excitation: Engine vibrations System: Military submarine Response: Sound Excitation: Aerodynamic loads System: Aeroplane wing Response: Stresses or deflection Response: Pitch or roll Response: Sway or deflection Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Excitation: vertical vibrations System: Packaged product Response: Product vibration Dead Weight Guided platen Signal Analyser Response accel. signal Test sample Vibration Controller Input accel. signal (control) Vibration table Servo-hydraulic Actuator 47 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Response accelerometer Linear bearings housing Guided dead weight Fibreboard cushion sample Vibration table Guide rod Input (table) accelerometer Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function steady-state Random Nd = 25 10 Transmissiblity 48 5 0 0 5 10 15 Frequency [Hz] 20 25 30 49 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Accelerometer Charge Amplifier Dead Weights Accelerometer Packaged Unit Charge Amplifier Table Servo Amplifier PC with ADC and DAC Modules. 50 Function Generator with externally-controlled frequency & amplitude. Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Hydraulic Servoactuator 51 Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function Response acceleration Sensor Excitation acceleration Sensor Signals – Frequency Domain Analysis V. Rouillard 2003 Frequency Response Function 5.0 Before test After test Gain 4.0 3.0 2.0 1.0 0.0 00 Phase Difference [Deg] 52 5 10 15 Frequency [Hz] -30 20 25 Before test After test -60 -90 -120 -150 -180 0 5 10 15 Frequency [Hz] 20 25