Files in this item

FilesDescriptionFormat

application/pdf

application/pdfBALAKRISHNAN-THESIS-2021.pdf (2MB)Restricted to U of Illinois
(no description provided)PDF

Description

Title:Recurrence-based models for improving coverage within GPS and satellite-denied mobile sensor networks
Author(s):Balakrishnan, Rahul
Advisor(s):Caesar, Matthew
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:M.S.
Genre:Thesis
Subject(s):UAV relay network
reinforcement learning
mobile sensing network
drone network
MANET
coverage maximization
recurrent algorithms
optimization
decentralized control
distributed network
Abstract:Adversarial GPS-denial and coordinate spoofing, as well as satellite jamming, serve as common obstacles to disaster-recovery and military-based teams. While such teams are often supported by UAVs that connect multiple personnel by serving as relay devices, UAV position schemes that rely on centralized controllers are also disrupted by such hurdles, as GPS and satellite-based denial will effect the ability of UAVs to communicate with the controller and drones outside line of sight. In this thesis, we design and implement a drone-relay system that allows drones to cooperatively maximize coverage in a GPS-denied and satellite-denied scenario. Here, we define coverage as the ability of one entity to speak to another entity using a drone network as an intermediate relay system and is measured as the ratio of fulfilled entity-to-entity connections to all possible entity-to-entity connections. We maximize coverage over an extended experiment duration, consisting of 300–1000 timesteps, by devising algorithms that rely on a centralized controller with full knowledge of drone and ground entity position. Particle Swarm Optimization (85% coverage), Reinforcement Learning on a Recurrent Neural Network (75% coverage), and a Time-Series based Inference Optimizer (71% coverage) were amongst the best performing movement algorithms, improving upon movement models in related works by up to 40%. We then design a distributed backend that disperses commands from a centralized controller using a distributed drone-to-drone communication scheme, also collecting observations made by each drone and relaying them to the centralized controller. This backend is additionally integrated with failure recovery and security-based protocols to ensure recovery in drone-downtime and drone-compromised scenarios; both systems feature minimal overhead, allowing drone recovery from downtime in 7% of the simulated episode length and featuring a constant time-addition from encryption that does not increase as drone count increases. Finally, we remove all notions of centrality by designing and implementing a fully decentralized system, where drones operate in squads and house their own models and decision-making protocols. In this system, drone squads are able to share observations of their surroundings with neighboring drone squads to improve predictive performance. This final system complies with GPS and satellite-denial limitations as drones only perform observations in a surrounding vision radius, assuming a camera to be mounted on each drone, and drones can only send messages to neighbors in a transmission radius, avoiding the need of sending satellite-based messages.
Issue Date:2021-04-28
Type:Thesis
URI:http://hdl.handle.net/2142/110737
Rights Information:Copyright 2021 Rahul Balakrishnan
Date Available in IDEALS:2021-09-17
Date Deposited:2021-05


This item appears in the following Collection(s)

Item Statistics