m-thesis-documentation/documents/thesis/chapters/radio_measurement.tex

276 lines
13 KiB
TeX
Raw Normal View History

% vim: fdm=marker fmr=<<<,>>>
\documentclass[../thesis.tex]{subfiles}
\graphicspath{
{.}
{../../figures/}
{../../../figures/}
}
\begin{document}
\chapter{Waveform Analysis Techniques}
\label{sec:waveform}
2023-06-19 23:43:47 +02:00
%Electric fields,
%Antenna Polarizations,
%Frequency Bandwidth,
2023-06-19 23:43:47 +02:00
Radio antennas are sensitive to changes in their surrounding electric fields.
The polarisation of the electric field that a single antenna can record is dependent on the geometry of this antenna.
Therefore, in experiments such as \gls{Auger} or \gls{GRAND}, multiple antennas are incorporated into a single unit to obtain complementary polarisation recordings.
2023-06-19 23:43:47 +02:00
Additionally, the shape and size of antennas affect how well the antenna responds to certain frequency ranges, resulting in different designs meeting different criteria.
\\
2023-06-19 23:43:47 +02:00
%Waveform time series data
%Sampling,
%Waveform + Time vector,
In each radio detector, the antenna presents its signals to an \gls{ADC} as fluctuating voltages.
In turn, the \gls{ADC} records the analog signals with a specified samplerate $f_s$ resulting in a sequence of digitised voltages or waveform.
The $n$-th sample in this waveform is then associated with a time $t[n] = t[0] + n/f_s = t[0] + n\cdot\Delta t$ after the initial sample at $t[0]$.
%In reality, the sampling rate will vary by very small amounts resulting in timestamp inaccuracies called jitter.
2023-06-19 23:43:47 +02:00
\\
2023-06-19 23:43:47 +02:00
% Filtering before ADC
The sampling is limited by the \gls{ADC}'s Nyquist frequency at half its sampling rate.
In addition, various frequency-dependent backgrounds can be reduced by applying a bandpass filter before digitisation.
For example, in \gls{AERA} and in AugerPrime's radio detector \cite{Huege:2023pfb}, the filter attenuates all of the signal except for the frequency interval between $30 \text{--} 80\MHz$.
2023-06-19 23:43:47 +02:00
\\
In addition to a bandpass filter, more complex filter setups are used to remove unwanted components or introduce attenuation at specific frequencies.
For example, in \gls{GRAND} \cite{GRAND:2018iaj}, the total frequency band ranges from $20\MHz$ to $200\MHz$.
2023-11-04 17:34:14 +01:00
such that the FM broadcasting band ($87.5\MHz \text{--} 108\MHz$) falls within this range.
Therefore, notch filters have been introduced to suppress signals in this band.
\\
2023-06-19 23:43:47 +02:00
% Filter and Antenna response
From the above it is clear that the digitised waveform is a measurement of the electric field that is sequentially convolved with the antenna's and filter's response.
2023-06-19 23:43:47 +02:00
Thus to reconstruct properties of the electric field signal from the waveform, both responses must be known.
\\
% Analysis, properties, frequencies, pulse detection, shape matching,
Different methods are available for the analysis of the waveform, and the antenna and filter responses.
2023-06-19 23:43:47 +02:00
A key aspect is determining the frequency-dependent amplitudes (and phases) in the measurements to characterise the responses and, more importantly, select signals from background.
With \glspl{FT}, these frequency spectra can be produced.
This technique is especially important for the sinewave beacon of Section~\ref{sec:beacon:sine}, as it forms the basis of the phase measurement.
2023-06-19 23:43:47 +02:00
\\
The detection and identification of more complex time-domain signals can be achieved using the cross correlation,
2023-06-19 23:43:47 +02:00
which is the basis for the pulsed beacon method of Section~\ref{sec:beacon:pulse}.
\\
\section{Fourier Transforms}% <<<<
\label{sec:fourier}
\glspl{FT} allow for a frequency-domain representation of a time-domain signal.
In the case of radio antennas, it converts a time-ordered sequence of voltages into a set of complex amplitudes that depend on frequency.
2023-06-19 23:43:47 +02:00
By evaluating the \gls{FT} at appropriate frequencies, the frequency spectrum of a waveform is calculated.
This method then allows to modify a signal by operating on its frequency components, i.e.~removing a narrow frequency band contamination within the signal.
\\
% DTFT from CTFT
The continuous \acrlong{FT} takes the form
\begin{equation}
\label{eq:fourier}
\phantom{.}
X(f) = \int_\infty^\infty \dif{t}\, x(t)\, e^{-i 2 \pi f t}
.
\end{equation}
2023-06-19 23:43:47 +02:00
It decomposes the signal $x(t) \in \mathcal{R}$ into plane waves with complex-valued amplitude $X(f)$ at frequency $f$.
\\
From the complex amplitude $X(f)$, the phase $\pTrue(f)$ and amplitude $A(f)$ are calculated as
\begin{equation*}
\begin{aligned}
\phantom{.}
\pTrue(f) = \arg\left( X(f) \right), && \text{and} && A(f) = 2 \left| X(f) \right|
.
\end{aligned}
\end{equation*}
Note the factor $2$ in this definition of the amplitude.
It is introduced to compensate for expecting a real valued input signal $x(t) \in \mathcal{R}$ and mapping negative frequencies to their positive equivalents.
\\
2023-06-19 23:43:47 +02:00
When $x(t)$ is sampled at discrete times, the integral of \eqref{eq:fourier} is discretized in time to result in the \gls{DTFT}:
\begin{equation}
%\tag{DTFT}
\label{eq:fourier:dtft}
X(f) = \sum_{n=0}^{N-1} x(t[n])\, e^{ -i 2 \pi f t[n]}
\end{equation}
where $x(t)$ is sampled a finite number of times $N$ at times $t[n]$.
2023-06-19 23:43:47 +02:00
Note that the amplitude $A(f)$ will now scale with the number of samples~$N$, and thus should be normalised to $A(f) = 2 \left| X(f) \right| / N$.
\\
2023-06-19 23:43:47 +02:00
Considering a finite sampling size $N$ and periodicity of the signal, the bounds of the integral in \eqref{eq:fourier} have collapsed to $t[0]$ up to $t_{N-1}$.
It follows that the lowest resolvable frequency is $f_\mathrm{lower} = 1/T = 1/(t_{N-1} - t[0])$.
\\
2023-06-19 23:43:47 +02:00
Additionally, when the sampling of $x(t)$ is equally spaced, the $t[n]$ terms can be written as a sequence, $t[n] - t[0] = n \Delta t = n/f_s$, with $f_s$ the sampling frequency.
Here the highest resolvable frequency is limited by the Nyquist~frequency.
\\
% DFT sampling of DTFT / efficient multifrequency FFT
Implementing the above decomposition of $t[n]$, \eqref{eq:fourier:dtft} can be rewritten in terms of multiples of the sampling frequency $f = k f_s/N$, becoming the \gls{DFT}
\begin{equation*}
\label{eq:fourier:dft}
2023-06-19 23:43:47 +02:00
\phantom{.}
X(k)
% = \sum_{n=0}^{N-1} x(t[n])\, e^{ -i 2 \pi f (t[0] + n/f_s)}
% = \sum_{n=0}^{N-1} x(t[n])\, e^{ -i 2 \pi f t[0]}\, e^{ -i 2 \pi f n/f_s}
% = e^{ -i 2 \pi f t[0]}\, \sum_{n=0}^{N-1} x(t[n])\, e^{ -i 2 \pi k n}
= e^{ -i 2 \pi f t[0]} \sum_{n=0}^{N-1} x(t[n])\, \cdot e^{ -i 2 \pi \frac{k n}{N} }
2023-06-19 23:43:47 +02:00
.
\end{equation*}
2023-06-19 23:43:47 +02:00
The direct computation of this transform takes $2N$ complex multiplications and $2(N-1)$ complex additions for a single frequency $k$.
When computing this transform for all integer $0 \leq k < N$, this amounts to $\mathcal{O}(N^2)$ complex computations.
\acrlong{FFT}s (\acrshort{FFT}s) are efficient algorithms that derive all $X( 0 \leq k < N)$ in $\mathcal{O}( N \log N)$ calculations.
\begin{figure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{methods/fourier/waveforms.pdf}%
%\caption{}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{methods/fourier/spectrum.pdf}%
\label{fig:fourier:dtft_dft}
%\caption{}
\end{subfigure}
\caption{
\textit{Left:} A waveform sampling a sine wave with white noise.
\textit{Right:}
The frequency spectrum of the waveform.
Comparison of the \gls{DTFT} and \gls{DFT} of the same waveform.
The \gls{DFT} can be interpreted as sampling the \gls{DTFT} at integer multiple of the waveform's sampling rate $f_s$.
}
\label{fig:fourier}
\end{figure}
% Linearity fourier for real/imag
2023-06-19 23:43:47 +02:00
In the previous equations, the resultant quantity $X(f)$ is a complex amplitude.
Since a complex plane wave can be linearly decomposed as
\begin{equation*}
\phantom{,}
\label{eq:complex_wave_decomposition}
\begin{aligned}
e^{-i x}
&
= \cos(x) + i\sin(-x)
%\\ &
= \Re\left(e^{-i x}\right) + i \Im\left( e^{-i x} \right)
,
\end{aligned}
\end{equation*}
the above transforms can be decomposed into explicit real and imaginary parts aswell,
i.e.,~\eqref{eq:fourier:dtft} becomes
\begin{equation}
\phantom{.}
\label{eq:fourier:dtft_decomposed}
\begin{aligned}
X(f)
&
= X_R(f) + i X_I(f)
%\\ &
\equiv \Re(X(f)) + i \Im(X(f))
\\ &
= \sum_{n=0}^{N-1} \, x(t[n]) \, \cos( 2\pi f t[n] )
- i \sum_{n=0}^{N-1} \, x(t[n]) \, \sin( 2\pi f t[n] )
.
\end{aligned}
\end{equation}
% FT term to phase and magnitude
The normalised amplitude at a given frequency $A(f)$ is calculated from \eqref{eq:fourier:dtft} as
\begin{equation}
\label{eq:complex_magnitude}
\phantom{.}
2023-06-19 23:43:47 +02:00
A(f)
\equiv \frac{2 \left| X(f) \right| }{N}
= \frac{ 2 \sqrt{ X_R(f)^2 + X_I(f)^2 } }{N}
.
\end{equation}
Likewise, the complex phase at a given frequency $\pTrue(f)$ is obtained by
\begin{equation}
\label{eq:complex_phase}
\phantom{.}
2023-06-19 23:43:47 +02:00
\pTrue(f)
\equiv \arg( X(f) )
= \arctantwo\left( X_I(f), X_R(f) \right)
.
\end{equation}
% Recover A\cos(2\pi t[n] f + \phi) using above definitions
Applying \eqref{eq:fourier:dtft_decomposed} to a signal $x(t) = A\cos(2\pi t f + \pTrue)$ with the above definitions obtains
an amplitude $A$ and phase $\pTrue$ at frequency $f$.
When the minus sign in the exponent of \eqref{eq:fourier} is not taken into account, the calculated phase in \eqref{eq:complex_phase} will have an extra minus sign.
\\
Figure~\ref{fig:fourier} shows the frequency spectrum of a simulated waveform that is obtained using either a \gls{DFT} or a \gls{DTFT}.
It shows that the \gls{DFT} evaluates the \gls{DTFT} only at certain frequencies.
By missing the correct frequency bin for the sine wave, it estimates both a too low amplitude and the wrong phase for the input function.
\\
2023-06-19 23:43:47 +02:00
% % Static sin/cos terms if f_s, f and N static ..
When calculating the \gls{DTFT} for multiple inputs which share both an equal number of samples $N$ and equal sampling frequencies $f_s$, the $\sin$ and $\cos$ terms in \eqref{eq:fourier:dtft_decomposed} are the same for a single frequency $f$ upto an overall phase which is dependent on $t[0]$.
Therefore, at the cost of an increased memory allocation, these terms can be precomputed, reducing the number of real multiplications to $2N+1$.
2023-06-19 23:43:47 +02:00
% .. relevance to hardware if static frequency
Thus, for static frequencies in a continuous beacon, the coefficients for evaluating the \gls{DTFT} can be put into the hardware of the detectors,
opening the way to efficiently measuring the amplitude and phase in realtime.
2023-06-19 23:43:47 +02:00
% >>>>
2023-06-19 23:43:47 +02:00
\section{Cross-Correlation}% <<<<
\label{sec:correlation}
The cross-correlation is a measure of how similar two waveforms $u(t)$ and $v(t)$ are.
By introducing a time delay $\tau$ in one of the waveforms it turns into a function of this time delay,
2023-06-19 23:43:47 +02:00
\begin{equation}
\label{eq:correlation_cont}
\phantom{,}
\Corr(\tau; u, v) = \int_{-\infty}^{\infty} \dif t \, u(t)\, v^*(t-\tau)
,
\end{equation}
where the integral reduces to a sum for a finite amount of samples in either $u(t)$ or $v(t)$.
Still, $\tau$ remains a continuous variable.
% Figure example of correlation and argmax
Figure~\ref{fig:correlation} illustrates how the best time delay $\tau$ between two waveforms can thus be found by finding the maximum cross-correlation.
\\
2023-06-19 23:43:47 +02:00
% Discrete \tau because of sampling
In reality, both waveforms have a finite size, also reducing the time delay $\tau$ resolution to the highest sampling rate of the two waveforms.
When the sampling rates are equal, the time delay variable is effectively shifting one waveform by a number of samples.
\\
% Upsampling? No
Techniques such as upsampling or interpolation can be used to effectively change the sampling rate of a waveform up to a certain degree.
2023-06-19 23:43:47 +02:00
\\
% Approaching analog \tau; or zero-stuffing
Since zero-valued samples do not contribute to the integral of \eqref{eq:correlation_cont}, they can be freely added (or ignored) to a waveform when performing the calculations.
This means two waveforms of different sampling rates can be correlated when the sampling rates are integer multiples of each other, simply by zero-stuffing the slowly sampled waveform.
This allows to approximate an analog time delay between two waveforms when one waveform is sampled at a very high rate as compared to the other.
2023-06-19 23:43:47 +02:00
\begin{figure}
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{methods/correlation/waveforms.pdf}
%\caption{
% Two waveforms.
%}%
\label{subfig:correlation:waveforms}
2023-06-19 23:43:47 +02:00
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{methods/correlation/correlation.pdf}
%\caption{
% The correlation of two Waveforms as a function of time.
%}%
\label{subfig:correlation}
\end{subfigure}%
2023-06-19 23:43:47 +02:00
\caption{
\textit{Left:} Two waveforms to be correlated with the second waveform delayed by $5$.
\textit{Right:} The correlation of both waveforms as a function of the time delay $\tau$.
Here the best time delay (red dashed line) is found at $5$, which would align the maximum amplitudes of both waveforms in the left pane.
2023-06-19 23:43:47 +02:00
}
\label{fig:correlation}
\end{figure}
% >>>
\end{document}