Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών
Πρόγραμμα Προπτυχιακών Σπουδών
ΠΑΡΟΥΣΙΑΣΗ ΔΙΠΛΩΜΑΤΙΚΗΣ ΕΡΓΑΣΙΑΣ
Ασύγχρονος Συμπερασμός σε Ασύρματα Δίκτυα Αισθητήρων με Ενέργεια από το Περιβάλλον
Asynchronous Inference in Ambiently-Powered Wireless Sensor Networks
Καθηγητής Άγγελος Μπλέτσας (επιβλέπων)
Καθηγητής Γεώργιος Καρυστινός
Αναπληρωτής Καθηγητής Μιχαήλ Λαγουδάκης
Wireless Sensor Networks (WSNs) are cost effective and ultra-low power networks that have recently become an integral part of many Internet-of-Things (IoT) applications. They consist of a certain number of nodes (or terminals), each of which is connected to a large number of sensors. Typically, the ambient information that the sensors are able to collect is wirelessly tranferred to some kind of centralized processing unit, which usually involves cloud or edge technologies.
In this work, we consider a WSN that is batteryless and solely powered by the environment. Our goal is to utilize such a network removing the centralized processing unit, and, by carefully balancing the computation and communication cost of modern inference algorithms, allow it to make autonomous, in network decisions itself; all that, exploiting its asynchronous operation that stems from the fact that it is ambiently powered: at some point of time certain WSN nodes may fail to operate.
In particular, we consider a linear fixed point problem and mathematically formulate its asynchronous variant, aiming to capture the asynchronous operation of the WSN, according to which some parts of the calculated vector may not be updated at some iterations. We propose a k-means based clustering method of assigning different parts of a vector to different WSN nodes. Next, we describe two algorithms that are both expressed as linear fixed point problems: a) Gaussian Belief Propagation and b) Average Consensus, as well as their asynchronous variants introduced in this work. Analysis as well as numerical results of this work show that the asynchronous operation of a WSN can be a key in the convergence of Gaussian Belief Propagation; indeed, we show that different asynchronous schedulings vastly affect its convergence speed, and – in some cases – asynchrony can make a divergent instance (in synchronous operation) to converge. On the other hand, in the case of Average Consensus, we derive a statistical condition that, when satisfied, leads to in expectation convergence of the algorithm. Hence, it is possible to execute Average Consensus in an ambiently powered WSN; the caveat here is an increased delay, since independent repetitions of the algorithm are necessary for an accurate result.
Meeting ID: 936 6662 2925