Spectral Analysis using Matlab

A typical measurement setup is shown in figure 1 where we measure a physical signal which is here a velocity \(v(t)\) using a geophone. The geophone has some dynamics that we represent with \(G_g(s)\), its output a voltage. The output of the geophone is then amplified by a voltage amplifier with a transfer function \(G_a(s)\).

Finally, the signal is discretised with an ADC and we obtain \(x(k\Delta t)\).

To obtain the real physical quantity \(v(t)\) as measured by the sensor from the obtained signal \(x(k\Delta t)\), one have to know the sensitivity of the sensors and electronics used (\(G_g(s)\) and \(G_a(s)\)).

Let’s say, we know that the sensitivity of the geophone used is \[ G_g(s) = G_0 \frac{\frac{s}{2\pi f_0}}{1 + \frac{s}{2\pi f_0}} \quad \left[\frac{V}{m/s}\right] \]

The parameters are defined below.

G0 = 88; % Sensitivity [V/(m/s)]
f0 = 2; % Cut-off frequency [Hz]

Gg = G0*(s/2/pi/f0)/(1+s/2/pi/f0);

And the dynamics of the amplifier in the bandwidth of interest is just a gain: \(G_a(s) = 1000\).

Ga = 1000;

Let’s here try to obtain the time domain signal \(v(t)\) from the measurement \(x\).

If \({G_a(s)}^{-1} {G_g(s)}^{-1}\) is proper, we can simulate this dynamical system to go from the voltage \(x\) to the velocity \(v\) as shown in figure 2.

If \({G_a(s)}^{-1} {G_g(s)}^{-1}\) is not proper, we add low pass filters at high frequency to make the system proper.

Let’s load the measured \(x\).

data = load('mat/data_028.mat', 'data'); data = data.data;

t = data(:, 3); % Time vector [s]
x = data(:, 1)-mean(data(:, 1)); % The offset if removed (coming from the voltage amplifier) [v]

dt = t(2)-t(1); % Sampling Time [s]
Fs = 1/dt; % Sampling Frequency [Hz]

We simulate this system with matlab using the lsim command.

v = lsim(inv(Gg*Ga), x, t);

And we plot the obtained velocity

From the Matlab documentation:

The goal of spectral estimation is to describe the distribution (over frequency) of the power contained in a signal, based on a finite set of data.

We now have the velocity \(v(t)\ [m/s]\) in the time domain.

The Power Spectral Density (PSD) \(S_v(f)\) of the time domain \(v(t)\) can be computed using the following equation: \[ S_v(f) = \frac{1}{f_s} \sum_{m=-\infty}^{\infty} R_{xx}(m) e^{-j 2 \pi m f / f_s} \ \left[\frac{(m/s)^2}{Hz}\right] \] where

• \(f_s\) is the sampling frequency in Hz
• \(R_{xx}\) is the autocorrelation

The PSD represents the distribution of the (average) signal power over frequency. \(S_v(f)df\) is the infinitesimal power in the band \((f-df/2, f+df/2)\) and the total power in the signal is obtained by integrating these infinitesimal contributions.

To compute the Power Spectral Density with matlab, we use the pwelch function (documentation). The use of the pwelch function is: [pxx,w] = pwelch(x,window,noverlap,nfft, fs) with:

• x is the discrete time signal
• window is a window that is used to smooth the obtained PSD
• overlap can be used to have some overlap from section to section
• nfft specifies the number of FFT points for the PSD
• fs is the sampling frequency of the data x in Hertz

As explained in (Schmid 2012), it is recommended to use the pwelch function the following way:

First, define a window (preferably the hanning one) by specifying the averaging factor na.

nx = length(v);
na = 8;
win = hanning(floor(nx/na));

Then, compute the power spectral density \(S_v\) and the associated frequency vector \(f\) with no overlap.

[Sv, f] = pwelch(v, win, 0, [], Fs);

The obtained PSD is shown in figure 4.

The Amplitude Spectral Density (ASD) is defined as the square root of the Power Spectral Density and is shown in figure 5.

Instead of computing the time domain velocity before computing the Power Spectral Density, we could have directly computed the PSD of the measured voltage \(x\) and then take into account the sensitivity of the measurement devices to have the PSD of the velocity.

To do so, we use the fact that a signal \(u\) with a PSD \(S_{uu}\) going through a LTI system \(G_(s)\) (figure 6) will generate a signal \(y\) with a PSD:

Similarly, the ASD of \(y\) is:

Thus, we could have computed the PSD of \(x\) and then obtain the PSD of the velocity with: \[ S_{v}(\omega) = |G_a(j\omega) G_g(j\omega)|^{-1} S_{x}(\omega) \]

The PSD of \(x\) is computed below.

nx = length(x);
na = 8;
win = hanning(floor(nx/na));
[Sx, f] = pwelch(x, win, 0, [], Fs);

And the PSD of \(v\) is obtained with the below code.

Svv = Sx.*squeeze(abs(freqresp(inv(Gg*Ga), f, 'Hz'))).^2;

The result is compare with the PSD computed from the \(v\) signal obtained with the lsim command in figure 7.

Similarly to what has been done in the last section, we can consider the displacement \(d\) can be obtained from the velocity \(v\) by going through an LTI system \(1/s\) as shown in figure 8.

We then have the relation between the PSD of \(d\) and the PSD of \(v\):

Using a frequency variable \(f\) in Hz:

For the Amplitude Spectral Density:

Note here that the PSD (and ASD) of one variable and its derivatives/integrals are equal at one particular frequency \(f = 1\ rad/s \approx 0.16\ Hz\):

With Matlab, the PSD of the displacement can be computed from the PSD of the velocity with the following code.

Sd = Sv.*(1./(2*pi*f)).^2;

The obtained PSD of the displacement can be seen in figure 9.

The Cumulative Power Spectrum is the cumulative integral of the Power Spectral Density:

It is also possible to integrate from high frequency to low frequency:

The Cumulative Power Spectrum taken at frequency \(f\) thus represent the power in the signal in the frequency band \(0\) to \(f\) or \(f\) to \(\infty\) depending on the above definition taken.

The choice of the integral direction depends on the shape of the PSD. If the power is mostly present at low frequencies, it is preferable to use equation \eqref{orgb6111fe}.

The Cumulative Amplitude Spectrum is defined as the square root of the Cumulative Power Spectrum: \[ CAS_v(f) = \sqrt{CPS_v(f)} = \sqrt{\int_f^\infty PSD_v(\nu) d\nu} \quad [m/s] \]

The Root Mean Square value of the velocity corresponds to the Cumulative Amplitude Spectrum when integrated at all frequencies: \[ v_{\text{rms}} = \sqrt{\int_0^\infty PSD_v(\nu) d\nu} = CAS_v(0) \quad [m/s \ \text{rms}] \]

With Matlab, the Cumulative Power Spectrum can be computed with the below formulas and the results are shown in figure 10.

CPS_v  = cumtrapz(f, Sv); % Cumulative Power Spectrum from low to high frequencies
CPS_vv = flip(-cumtrapz(flip(f), flip(Sv))); % Cumulative Power Spectrum from high to low frequencies

figure;
ax1 = subplot(1, 2, 1);
hold on;
plot(f, CPS_v);
hold off;
set(gca, 'xscale', 'log'); set(gca, 'yscale', 'log');
xlabel('Frequency [Hz]'); ylabel('Cumulative Power Spectrum [$(m/s)^2$]')

ax2 = subplot(1, 2, 2);
hold on;
plot(f, CPS_vv);
hold off;
set(gca, 'xscale', 'log'); set(gca, 'yscale', 'log');
xlabel('Frequency [Hz]');

linkaxes([ax1,ax2],'xy');
xlim([0.1, 500]); ylim([1e-15, 1e-11]);

We load the PSD of the signal we wish to replicate.

load('./mat/dist_psd.mat', 'dist_f');

We remove the first value with very high PSD.

dist_f.f = dist_f.f(3:end);
dist_f.psd_gm = dist_f.psd_gm(3:end);

The PSD of the signal is shown on figure fig:psd_ground_motion.

Using sisotool or any other tool, we create a transfer function \(G\) such that its magnitude is close to the Amplitude Spectral Density \(\Gamma_x = \sqrt{S_x}\): \[ |G(j\omega)| \approx \Gamma_x(\omega) \]

G_gm = 0.002*(s^2 + 3.169*s + 27.74)/(s*(s+32.73)*(s+8.829)*(s+7.983)^2);

We compare the ASD \(\Gamma_x(\omega)\) and the magnitude of the generated transfer function \(|G(j\omega)|\) in figure [[]].

We know that a signal \(u\) going through a LTI system \(G\) (figure 13) will have its ASD modified according to the following equation:

Thus, if we create a random signal with an ASD equal to one at all frequency and we pass this signal through the previously defined LTI transfer function G_gm, we should obtain a signal with an ASD that approximate the original ASD.

To obtain a random signal with an ASD equal to one, we use the following code.

Fs = 2*dist_f.f(end); % Sampling Frequency [Hz]
Ts = 1/Fs; % Sampling Time [s]

t = 0:Ts:500; % Time Vector [s]
u = sqrt(Fs/2)*randn(length(t), 1); % Signal with an ASD equal to one

We then use lsim to compute \(y\) as shown in figure 13.

y = lsim(G_gm, u, t);

The obtained time domain signal is shown in figure 14.

We now compute the Power Spectral Density of the computed time domain signal \(y\).

nx = length(y);
na = 16;
win = hanning(floor(nx/na));

[pxx, f] = pwelch(y, win, 0, [], Fs);

Finally, we compare the PSD of the original signal and the obtained signal on figure fig:psd_comparison.

One advantage of this technique is that it can be easily integrated into simulink.

The corresponding schematic is shown in figure 16 where the block Band-Limited White Noise is used to generate a random signal with a PSD equal to one (parameter Noise Power is set to 1).

Then, the signal generated pass through the transfer function representing the wanted ASD.

We simulate the system shown in figure 16.

out = sim('matlab/generate_signal_psd.slx');

And we compute the PSD of the generated signal.

nx = length(out.u_gm.Data);
na = 8;
win = hanning(floor(nx/na));

[pxx, f] = pwelch(out.u_gm.Data, win, 0, [], 1e3);

Finally, we compare the PSD of the generated signal with the original PSD in figure 17.

We load the PSD of the signal we wish to replicate.

load('./mat/dist_psd.mat', 'dist_f');

We remove the first value with very high PSD.

dist_f.f = dist_f.f(3:end);
dist_f.psd_gm = dist_f.psd_gm(3:end);

The PSD of the signal is shown on figure fig:psd_original.

We define some parameters that will be used in the algorithm.

Fs = 2*dist_f.f(end);    % Sampling Frequency of data is twice the maximum frequency of the PSD vector [Hz]
N  = 2*length(dist_f.f); % Number of Samples match the one of the wanted PSD
T0 = N/Fs;               % Signal Duration [s]
df = 1/T0;               % Frequency resolution of the DFT [Hz]
                         % Also equal to (dist_f.f(2)-dist_f.f(1))

We then specify the wanted PSD.

phi = dist_f.psd_gm;

We create amplitudes corresponding to wanted PSD.

C = zeros(N/2,1);
for i = 1:N/2
  C(i) = sqrt(phi(i)*df);
end

Finally, we add some random phase to C.

theta = 2*pi*rand(N/2,1); % Generate random phase [rad]

Cx = [0 ; C.*complex(cos(theta),sin(theta))];
Cx = [Cx; flipud(conj(Cx(2:end)))];;

The time domain data is generated by an inverse FFT.

The ifft Matlab does not take into account the sampling frequency, thus we need to normalize the signal.

u = N/sqrt(2)*ifft(Cx); % Normalisation of the IFFT
t = linspace(0, T0, N+1); % Time Vector [s]

We duplicate the time domain signal to have a longer signal and thus a more precise PSD result.

u_rep = repmat(u, 10, 1);

We compute the PSD of the obtained signal with the following commands.

nx = length(u_rep);
na = 16;
win = hanning(floor(nx/na));

[pxx, f] = pwelch(u_rep, win, 0, [], Fs);

Finally, we compare the PSD of the original signal and the obtained signal on figure fig:psd_comparison.

Let’s first define the number of sample and the sampling time.

N  = 10000; % Number of Sample
dt = 0.001; % Sampling Time [s]

t = dt*(0:1:N-1)'; % Time vector [s]

We generate of signal that consist of:

• a white noise with an RMS value equal to anoi
• two sinusoidal signals

The parameters are defined below.

asig = 0.1; % Amplitude of the signal [V]
fsig = 10;  % Frequency of the signal [Hz]

ahar = 0.5; % Amplitude of the harmonic [V]
fhar = 50;  % Frequency of the harmonic [Hz]

anoi = 1e-3; % RMS value of the noise

The signal \(x\) is generated with the following code and is shown in figure 21.

x = anoi*randn(N, 1) + asig*sin((2*pi*fsig)*t) + ahar*sin((2*pi*fhar)*t);

Let’s compute the PSD of the signal using the blackmanharris window.

nx = length(x);
na = 8;
win = blackmanharris(floor(nx/na));

[pxx, f] = pwelch(x, win, 0, [], 1/dt);

Normalization of the PSD.

CG = sum(win)/(nx/na);
NG = sum(win.^2)/(nx/na);
fbin = f(2) - f(1);

pxx_norm = pxx*(NG*fbin/CG^2);

We determine the frequency bins corresponding to the frequency of the signals.

isig = round(fsig/fbin)+1;
ihar = round(fhar/fbin)+1;

The theoretical RMS value of the signal is:

srmt = asig/sqrt(2); % Theoretical value of signal magnitude

And we estimate the RMS value of the signal by either integrating the PSD around the frequency of the signal or by just taking the maximum value.

srms = sqrt(sum(pxx(isig-5:isig+5)*fbin)); % Signal spectrum integrated
srmsp = sqrt(pxx_norm(isig) * NG*fbin/CG^2); % Maximum read off spectrum

We see that the integrated value gives a good approximate of the true RMS value whereas the maximum value give an poor approximate. This effect is called scallop loss. Thus, always the integrated method should be used.

The noise level can also be computed using the integration method.

The theoretical RMS noise value is.

nth = anoi/sqrt(max(f)) % Theoretical value [V/sqrt(Hz)]

We can estimate this RMS value by integrating the PSD at frequencies where the power of the noise signal is above the power of the other signals.

navg = sqrt(mean(pxx_norm([ihar+10:end]))) % pwelch output averaged

The estimate of the noise level is quite good.

This is taken from here.

Let’s note:

• \(q\) is the corresponding value in [V] of the least significant bit (LSB)
• \(\Delta V\) is the full range of the ADC in [V]
• \(n\) is the number of ADC’s bits
• \(f_s\) is the sample frequency in [Hz]

Let’s suppose that the ADC is ideal. The only noise comes from the quantization error. Interestingly, the noise amplitude is uniformly distributed.

The quantization noise can take a value between \(\pm q/2\), and the probability density function is constant in this range (i.e., it’s a uniform distribution). Since the integral of the probability density function is equal to one, its value will be \(1/q\) for \(-q/2 < e < q/2\) (Fig. 22).

Now, we can calculate the time average power of the quantization noise as

The other important parameter of a noise source is the power spectral density (PSD), which indicates how the noise power spreads in different frequency bands. To find the power spectral density, we need to calculate the Fourier transform of the autocorrelation function of the noise.

Assuming that the noise samples are not correlated with one another, we can approximate the autocorrelation function with a delta function in the time domain. Since the Fourier transform of a delta function is equal to one, the power spectral density will be frequency independent. Therefore, the quantization noise is white noise with total power equal to \(P_q = \frac{q^2}{12}\).

Thus, the two-sided PSD (from \(\frac{-f_s}{2}\) to \(\frac{f_s}{2}\)), we should divide the noise power \(P_q\) by \(f_s\):

Finally: