Rajrup Mitra

Date of Award

May 2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering

First Advisor

Naima Kaabouch

Abstract

As integral parts of different infrastructure constructions, steel structures must be monitored regularly for their durability, serviceability, and safety. However, environmental exposure of steel structures leads to defects over time that may compromise the structural integrity and lead to colossal repair costs if not detected early. Traditional manual inspection methods have limitations like accuracy constraints, inefficiency, and non-scalability. Hence, there is a need for cutting-edge technology-based solutions to streamline inspection processes. Thus, this thesis proposes a pioneering exploration into the development and integration of an Unmanned Aerial Vehicle (UAV) system and an intelligent Graphical User Interface (GUI) for real-time defect detection in steel structures. The suggested system uses modern visual and thermal imaging technologies combined with artificial intelligence algorithms and interactive operations aimed at real-time and smart structural integrity assessment. The UAV payload is a lightweight, low-power, and seamlessly integrated solution. A comprehensive evaluation and feasibility of the usage of two deep learning models, namely InceptionResnetV2 and ResNet152V2, for structural inspection is studied in this thesis. These models are trained and tested on two different datasets. The results show that ResNet152V2 outperforms InceptionResnetV2 with an average accuracy of 95% and a misdetection rate of 5%. Additionally, there is also the smart GUI that complements the UAV payload system by enhancing operator interactions as well as facilitating easy and real-time defect detection. On average it takes as low as less than a minute for an entire run to complete for a specific type of defect detection from an image click to saving the results to the repository. The effectiveness and robustness of the proposed system are demonstrated, showcasing its potential to boost structural inspection practices and enhance safety standards across various industries. Apart from this, for safeguarding the UAV's safety and integrity against GPS spoofing attacks, a comparative evaluation of lightweight artificial intelligence (AI) is also conducted for real-time detection. Results showing the Random Forest machine learning model outperforms three other AI models in GPS spoofing attack detection while a multi-layer perceptron (MLP) model takes the least processing time. By conducting extensive assessments and carrying out experiments, this thesis demonstrates the efficiency and reliability of the design.

Download

Share

COinS