Asynchronous sampling via time encoding strategies
File(s)
Author(s)
Hilton, Marek
Type
Thesis
Abstract
Time encoding is an alternative approach to sampling that encodes a signal in the timing of samples and not their amplitudes. Time encoding generates inherently asynchronous samples and in this thesis I look at how such encodings can be used to more efficiently sample and reconstruct signals which themselves exhibit significant sporadic behaviour. By sporadic behaviour, I refer to signals that are comprised of periods of activity interspersed with periods of quiescence. Such signals are typically oversampled by existing time encoding methods which generate samples at a rate similar to conventional sampling techniques.
I focus largely on signals with finite rate of innovation as a model for sporadic signals. These include bursts of pulses and piecewise constant signals. I extend existing results regarding time encoding and reconstruction of signals with finite rate of innovation to novel sampling kernels, namely, hyperbolic kernels and the biologically inspired α-synaptic function. Time encoding is inherently a non-uniform sampling technique and consequently, the majority of reconstruction algorithms proceed by applying standard non-uniform reconstruction techniques. In this sense, the signal driven timing of samples is not exploited to the fullest extent in standard approaches to time encoding.
Taking an alternative perspective, I present a new time encoding architecture called a time dilating integrate-and-fire encoder (TDIF) that uses dilation of time measurements to affect a form of filtering of an equivalent signal in amplitude-domain. I show that time encodings derived from such a sampler can be treated as uniform samples and thus, the standard techniques of uniform sampling can be applied. Since the local sampling rate of such a system is determined solely by the local power of the signal, it is ideally suited to sampling highly sporadic signals.
Finally I present some multi-encoder networks that complement the complement the reconstruction strategies presented in this thesis.
I focus largely on signals with finite rate of innovation as a model for sporadic signals. These include bursts of pulses and piecewise constant signals. I extend existing results regarding time encoding and reconstruction of signals with finite rate of innovation to novel sampling kernels, namely, hyperbolic kernels and the biologically inspired α-synaptic function. Time encoding is inherently a non-uniform sampling technique and consequently, the majority of reconstruction algorithms proceed by applying standard non-uniform reconstruction techniques. In this sense, the signal driven timing of samples is not exploited to the fullest extent in standard approaches to time encoding.
Taking an alternative perspective, I present a new time encoding architecture called a time dilating integrate-and-fire encoder (TDIF) that uses dilation of time measurements to affect a form of filtering of an equivalent signal in amplitude-domain. I show that time encodings derived from such a sampler can be treated as uniform samples and thus, the standard techniques of uniform sampling can be applied. Since the local sampling rate of such a system is determined solely by the local power of the signal, it is ideally suited to sampling highly sporadic signals.
Finally I present some multi-encoder networks that complement the complement the reconstruction strategies presented in this thesis.
Version
Open Access
Date Issued
2024-05-10
Date Awarded
01/01/2025
License URL
Advisor
Dragotti, Pier Luigi
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)