WuotD: radio measurement and DTFT

This commit is contained in:
Eric Teunis de Boone 2023-06-19 23:43:47 +02:00
parent 9f7ead9a41
commit 41ac01696f
2 changed files with 189 additions and 87 deletions

View file

@ -10,132 +10,137 @@
\begin{document}
\chapter{Measuring with Radio Antennas}
\label{sec:waveform}
Electric fields,
Antenna Polarizations,
Frequency Bandwidth,
\\
Time Domain,
Sampling,
Waveform + Time vector,
\\
Analysis:
Fourier Transforms,
Correlation
\hrule
%Electric fields,
%Antenna Polarizations,
%Frequency Bandwidth,
Radio antennas are sensitive to changes in their surrounding electric fields.
Depending on the antenna geometry, multiple polarisations of the electric field can be recorded simultaneously.
The polarisations 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 (called channels) are incorporated into a single unit to obtain complementary polarisation recordings.
Additionally, the shape and size of antennas affect how well the antenna responds to certain frequency ranges, resulting in different designs meeting different criteria.
\\
Recording
%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.
In this waveform, the $n$-th sample is associated with a fixed timestamp $t_n = t_0 + n/f_s$ after the initial sample at $t_0$.
\\
% Filtering before ADC
The finite sampling rate of the waveform means that very high frequencies are not observed by the \gls{ADC}.
However, frequencies just at or above half of sampling rate will affect the sampling and appear in the waveform at lower frequencies as aliases.
This frequency at half the sampling rate is known as the Nyquist frequency.
To prevent such aliases, these frequencies must be removed by a filter before sampling.
\\
For air shower radio detection, very low frequencies are also not of interest.
Therefore, this filter is generally a bandpass filter.
For example, in \gls{Auger} the filter removes all of the signal except for the frequency interval between $30 \text{--} 80\MHz$.\Todo{citation}
\\
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} notch filters are introduced to suppress radio signals in the FM-radio band which lies in its $20 \text{--} 200\MHz$ band.\Todo{citation}
\\
\section{Analysis Methods}% <<<
\label{sec:waveform:analysis}
\subsection{Correlation}% <<<<
\label{sec:correlation}
\Todo{intro}
The correlation is a measure of how similar two signals $u(t)$ and $v(t)$ are as a function of a time delay $\tau$.
It is defined as
\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.
% Filter and Antenna response
From the above it is clear that the digitised waveform is a measurement of the electric field that is sequentially convoluted by the antenna's and filter's response.
Thus to reconstruct properties of the electric field signal from the waveform, both responses must be known.
\begin{figure}
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_12_correlation.pdf}
\caption{
Correlation
}%
\label{subfig:correlation}
\end{subfigure}%
\\
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_1.pdf}
\caption{
Waveform 1
}
\label{}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_2.pdf}
\caption{
Waveform 2
}
\label{}
\end{subfigure}
\caption{
Top: Correlation of Waveform 1 and Waveform 2
}
\label{fig:correlation}
\caption{Antenna arms of a GRAND detector, a noisy waveform and a filter response}
\end{figure}
% >>>
\subsection{Fourier Transform}% <<<<
% Analysis, properties, frequencies, pulse detection, shape matching,
\bigskip
Different methods are available for the analysis of the waveform, and the antenna and filter responses.
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 \acrlong{FT}s these frequency spectra can be produced.
This technique is especially important for the sine beacon of Section~\ref{sec:beacon:sine}, as it forms the basis of the phase measurement.
\\
The detection and identification of more complex signals can be achieved using the cross correlation\Todo{rephrase},
which is the basis for the pulsed beacon method of Section~\ref{sec:beacon:pulse}.
%\section{Analysis Methods}% <<<
%\label{sec:waveform:analysis}
\section{Fourier Transform}% <<<<
\label{sec:fourier}
\Todo{intro}
The \gls{FT} allows 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 complex amplitudes that depend on frequency.
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 formulation of the \acrlong{FT} takes the following form,
The continuous \gls{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}
It decomposes the signal $x(t)$ into complex-valued plane waves $X(f)$ of frequency $f$.
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.
\\
When $x(t)$ is sampled at discrete times, the integral of \eqref{eq:fourier} is discretized in time to result in the \acrlong{DTFT}:
\bigskip
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) \in \mathcal{R} $ is sampled at times $t[n]$.
Considering a finite sampling size $N$ and periodicity of the signal, the bounds of the integral in \eqref{eq:fourier} collapse to $t[0]$ up to $t[N]$.
where $x(t)$ is sampled a finite number of times $N$ at some timestamps $t[n]$.
Note that the amplitude $A(f)$ will now scale with the number of samples~$N$, and thus should be normalised to $A(f) = 2 \left| X(f) \right| / N$.
\\
From this it follows that the lowest resolvable frequency is $f_\mathrm{lower} = \tfrac{1}{T} = \tfrac{1}{t[N] - t[0]}$.
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 $t[N]$.
It follows that the lowest resolvable frequency is $f_\mathrm{lower} = 1/T = 1/(t[N] - t[0])$.
\\
Additionally, when the sampling of $x(t)$ is equally spaced, the $t[n]$ terms can be written as a sequence, $t[n] - t[0] = n \Delta t = \tfrac{n}{f_s}$, with $f_s$ the sampling frequency.
The highest resolvable frequency, known as the Nyquist frequency, is limited by this sampling frequency as $f_\mathrm{nyquist} = \tfrac{f_s}{2}$.
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 at $f_\mathrm{nyquist} = f_s/2$.
\\
% 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 $k$ of the sampling frequency, becoming the \acrlong{DFT}
Implementing the above decomposition of $t[n]$, \eqref{eq:fourier:dtft} can be rewritten in terms of multiples of the sampling frequency $k = f/f_s$, becoming the \gls{DFT}
\begin{equation*}
\label{eq:fourier:dft}
\phantom{,}
X(k) = \sum_{n=0}^{N-1} x[n]\, \cdot e^{ -i 2 \pi {\frac{k n}N} }
,
\phantom{.}
X(k) = e^{ -i 2 \pi t[0]} \sum_{n=0}^{N-1} x[n]\, \cdot e^{ -i 2 \pi \frac{k n}{N} }
.
\end{equation*}
with $k = \tfrac{f}{f_s}$.
For integer $0 \leq k < N $, efficient algorithms exist that derive all $X( 0 \leq k < N )$ in $\mathcal{O}( N \log N )$ calculations, a~\acrlong{FFT}, sampling a subset of the frequencies.\Todo{citation?}
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 (\gls{FFT}s) are efficient algorithms that derive all $X( 0 \leq k < N)$ in $\mathcal{O}( N \log N)$ calculations.
\Todo{citation?}
%For integer $0 \leq k < N $, efficient algorithms exist that derive all $X( 0 \leq k < N )$ in $\mathcal{O}( N \log N )$ calculations instead of $\mathcal{O}(kcalled \acrlong{FFT}s, sampling a subset of the frequencies.\Todo{citation?}
\begin{figure}
\includegraphics[width=\textwidth]{fourier/dtft_dft_comparison.pdf}
\caption{
Comparison of the \gls{DTFT} and \gls{DFT} of the same waveform.
The \gls{DFT} can be interpreted as sampling the \gls{DTFT}
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:dtft_dft}
\end{figure}
\bigskip
\bigskip
% Linearity fourier for real/imag
In the previous equations, the resultant quantity $X(f)$ is a complex value.
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{,}
@ -172,18 +177,20 @@ The normalised amplitude at a given frequency $A(f)$ is calculated from \eqref{e
\begin{equation}
\label{eq:complex_magnitude}
\phantom{.}
A(f) \equiv \frac{ 2 \sqrt{ X_R(f)^2 + X_I(f)^2 } }{N}
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{.}
\pTrue(f) \equiv \arctantwo\left( X_I(f), X_R(f) \right)
\pTrue(f)
\equiv \arg( X(f) )
= \arctantwo\left( X_I(f), X_R(f) \right)
.
\end{equation}
Note the factor $2$ in the definition of the amplitude in \eqref{eq:complex_magnitude}.
It is introduced to compensate for expecting a real input signal $x(t)$ and mapping negative frequencies to their positive equivalents.
\\
% Recover A\cos(2\pi t[n] f + \phi) using above definitions
@ -191,9 +198,104 @@ Applying \eqref{eq:fourier:dtft_decomposed} to a signal $x(t) = A\cos(2\pi t[n]
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.
\bigskip
% % 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 $2N+1$, with the additional.
% .. 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 phases in realtime.\Todo{figure?}
% >>>>
\subsubsection{Hilbert Transform (optional)}% <<<<
%\section{Pulse Detection}
\section{Cross-Correlation}% <<<<
\label{sec:correlation}
\Todo{intro}
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 is turned into a function of this time delay.
It is defined as
\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.
\\
\bigskip
% 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 in effect shifting one waveform by some number of samples.
\\
% Approaching analog \tau; or applying upsampling and interpolation
Techniques such as upsampling or interpolation can be used
When the sampling rates are
\begin{figure}
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_12_correlation.pdf}
\caption{
Correlation of two Waveforms as a function of time.
}%
\label{subfig:correlation}
\end{subfigure}%
\\
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_1.pdf}
\caption{
Waveform 1
}
\label{}
\end{subfigure}
\hfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pulse/waveform_2.pdf}
\caption{
Waveform 2
}
\label{}
\end{subfigure}
\caption{
Top: Correlation of Waveform 1 and Waveform 2
}
\label{fig:correlation}
\end{figure}
% >>>
\section{Hilbert Transform}% <<<<
A variation of the Fourier Transform of Section~\ref{sec:fourier} is the Hilbert Transform.
With it, the analytic signal $s_a(t)$ of a waveform $x(t)$ can be obtained through
\begin{equation}
\label{eq:analytic_signal}
\phantom{,}
s_a(t) = x(t) + \hat{x}(t)
,
\end{equation}
where $\hat{x}(t)$ is the Hilbert Transformed waveform.
The Hilbert Transform corresponds to a \gls{FT} where positive frequencies $f > 0$ are phase-shifted by $-\pi/2$ and negative frequencies are phase-shifted by $+\pi/2$.
\bigskip
The envelope of a waveform $x(t)$ is determined by taking the absolute value of its analytic signal $s_a(t)$.
Figure~\ref{fig:hilbert_transform} shows an envelope with its original waveform.
\begin{figure}
\includegraphics[width=\textwidth]{pulse/hilbert_timing_interpolation_template.pdf}
\caption{
Timing information from the maximum amplitude of the envelope.
}
\label{fig:hilbert_transform}
\end{figure}
% >>>>
% >>>
\end{document}