Abstract

Innovations in communication systems, compute hardware, and deep learning algorithms have led to the advancement of smart industry automation. Smart automation includes industrial sectors such as intelligent warehouse management, smart infrastructure for first responders, and smart monitoring systems. Automation aims to maximize efficiency, safety, and reliability. Autonomous forklifts can significantly increase productivity, reduce safety-related accidents, and improve operation speed to enhance the efficiency of a warehouse. Forklifts or robotic agents are required to perform different tasks such as position estimation, mapping, and dispatching. Each of the tasks involves different requirements and design constraints. Smart infrastructure for first responder applications requires robotic agents like Unmanned Aerial Vehicles (UAVs) to provide situation awareness surrounding an emergency. An immediate and efficient response to a safety-critical situation is crucial, as a better first response significantly impacts the safety and recovery of parties involved. But these UAVs lack the computational power required to run Deep Neural Networks (DNNs) that are used to provide the necessary intelligence. In this dissertation, we focus on two applications in smart industry automation. In the first part, we target smart warehouse automation for Intelligent Material Handling (IMH), where we design an accurate and robust Machine Learning (ML) based indoor localization system for robotic agents working in a warehouse. The localization system utilizes millimeter-wave (mmWave) wireless sensors to provide feature information in the form of a radio map which the ML model uses to learn indoor positioning. In the second part, we target smart infrastructure for first responders, where we present a computationally efficient adaptive exit strategy in multi-exit Deep Neural Networks using Deep Reinforcement Learning (DRL). The proposed adaptive exit strategy provides faster inference time and significantly reduces computations.

Library of Congress Subject Headings

Automation--Technological innovations; Millimeter wave devices; Reinforcement learning; Deep learning (Machine learning); Neural networks (Computer science)

Publication Date

7-2022

Document Type

Dissertation

Student Type

Graduate

Degree Name

Engineering (Ph.D.)

Department, Program, or Center

Engineering (KGCOE)

Advisor

Amlan Ganguly

Advisor/Committee Member

Andres Kwasinski

Advisor/Committee Member

Michael E. Kuhl

Campus

RIT – Main Campus

Plan Codes

ENGR-PHD

Share

COinS