Favor the tortoise over the hare: a study on an efficient detection algorithm for wireless sensor networks.

Author(s)

Allan Eduardo Feitosa

• Document type: Doctoral Thesis
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Escola Politécnica (EP/BC)
• Defense date: 2023-08-30

Examining board members

Vitor Heloiz Nascimento; Fernando Gonçalves de Almeida Neto; Márcio Holsbach Costa; Rodrigo Caiado de Lamare; Magno Teófilo Madeira da Silva

• Advisor: Vitor Heloiz Nascimento; Cássio Guimarães Lopes

Abstract

This doctoral thesis describes the results of a detailed research conducted between January 2019 and July 2023 on a new distributed detection algorithm. In general terms, this study deals with the statistical detection problem using smart Wireless Sensor Networks (WSNs). In this context a network of sensors is distributed over a site in order to monitor the environment and decide the current state of nature based on observations under Gaussian noise. The sensors use embedded computation capabilities to locally process data and communicate wirelessly with closest sensors, enabling the exploitation of cooperative algorithms. More specifically, this study focused on a situation where WSNs are deployed over sites under stringent power conditions therefore, low computational complexity and low power consumption is highly desired, which led to the development of a detection algorithm suitable for real applications and with a performance that tends to optimum under such restrictions. Moreover, in a world increasingly connected through the Internet of Things (IoT) paradigm, algorithms that perform indispensable tasks such as detection and operate with minimum energy consumption are highly sought after. Not incidentally, the main contribution of this thesis is the description of a detector with low computational complexity that approximates the performance expected from an optimal detector in terms of the average probability of error, provided certain conditions are met. The most crucial is maintaining a slow learning rate of the distributed algorithm that drives the detection routine, specifically the diffusion LMS (Least Mean Square), a wellknown adaptive estimation algorithm for distributed networks. The diffusion LMS is used in this context for data processing and sharing information among sensors throughout the network. Notably, like the Tortoise of the fable, the performance of the detector improves as the value of the LMS step size is reduced, without penalizing the convergence rate in terms of probability of error, and despite the slowing down of the estimation routine at the core of the developed detection algorithm. This counterintuitive result is explained theoretically and confirmed by simulations. The detection problem presented herein is modeled as a multiple hypothesis test using the Bayesian formulation, and it extends the research conducted during my master degree to a more general situation. Additionally, this thesis includes interesting insights about the value of the initial estimate of the LMS algorithm, paving the way for promising future research. (AU)

• FAPESP's process: 18/26040-2 - Study and development of distributed detectors with fast convergence
• Grantee: Allan Eduardo Feitosa
• Support Opportunities: Scholarships in Brazil - Doctorate