Publication
Modern IoT platforms implement a plethora of techniques for both improving performance and energy efficiency. Because of this, the leading cause of energy consumption in IoT devices is moving from raw computation to data acquisition, storage, and transmission. The energy consumption of these tasks comes from the fact that almost every signal processing technique relies on the assumption that the acquired samples are equally spaced in time (or in the generic dependent variable upon which we assume the signal to vary). This space is fixed by the Nyquist theorem that uses this sufficient but not necessary assumption. Therefore, by modulating the space between each sample, we are not breaking a necessary condition of the Nyquist theorem. Moreover, even if part of the information content may be lost, if the space between each sample is computed using local properties of the signal, we can control the amount of information discarded. This method to treat signals, often referred to as event-based sensing, can vastly decrease the amount of data needed by IoT applications and even by datacenters, allowing smaller memories and fewer data to process. Our previous work on a variety of bio-signal related tasks (QRS complex detection, postural estimation, and respiration rate estimation) already shows how we can obtain comparable experimental results to the same techniques applied to uniformly sampled signals. At the same time, we obtain a decrement of more than 90% of the overall amount of processed data. In this Ph.D., we aim to develop several techniques to explore the potentiality of this framework regarding computation efficiency and memory management. To do so, we already started to analyze several event-based sensing techniques like level-crossing sampling and the Wall-Danielson algorithm and their applications. In particular, we divide our objective into two tasks: in the first one, we will study the behavior of this framework when compared to a classic signal processing framework acting on uniformly sampled signals, in the second one, we aim to develop new techniques that fully make use of the event-based sensing properties.