Thermal image analytics for industrial anomaly detection

• Main Author: Goyal, Shreyas
• Other Authors: Jagath C Rajapakse
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; hotspot detection; thermal images; self-supervised learning; GAN; Intrusion detection
• Online Access: https://hdl.handle.net/10356/175574
• Institution: Nanyang Technological University

Thermal imaging has become pervasive in a number of applications. These range from healthcare (medical screening, temperature monitoring), security and defence (nightvision), mobility (autonomous vehicles), monitoring of equipment in several industries (solar panels, transformers, electric equipment), etc. Hotspots develop in equipment as early indicators of oncoming failures. This is why hotspot detection is an important step in the early diagnosis of equipment for faults while they are developing, to prevent complete failures through remedial preventative action. However, the application of equipment monitoring using thermal imagery suffers from problems like lack of precision in abnormality or hotspot detection and isolation. Moreover, there is a paucity of suitable thermal imaging datasets for abnormality detection. This thesis presents a novel and comprehensive exploration of deep learning techniques applied to thermal image analysis, addressing critical issues in the domains of hotspot detection. Comprising of three chapters, the research delves into the realms of self-supervised learning, generative adversarial networks (GANs), and their applications in the thermal imaging domain. Chapter 2 introduces a novel approach to hotspot detection using self-supervised learning. Leveraging the power of unsupervised learning paradigms, the proposed methodology autonomously identifies thermal hotspots in images without the need for labelled data. With this work we propose a solution to the problem of paucity of thermal image datasets, as well as the problem of low precision in thermal hotspot isolation. This is done by using self-supervised learning, as well as the contribution of a new thermal imaging dataset. The chapter not only establishes the efficacy of self-supervised learning in thermal image analysis but also lays the foundation for subsequent advancements in the field. Chapter 3 builds further on the problem of hotspot detection and isolation to make it more accurate. It explores the application of conditional generative adversarial networks (cGANs) in hotspot detection within thermal images. GANs, known for their ability to generate realistic data, are employed to enhance the detection accuracy of thermal hotspots. The research demonstrates the synergy between GANs and thermal imaging, providing a sophisticated and effective solution for accurate hotspot identification in diverse thermal scenarios. Chapter 4 shifts the focus to human-intrusion detection as a special case of hotspot detection, through the lens of thermal imaging. Harnessing the unique attributes of infrared signatures of humans, the research investigates deep learning algorithms to discern and classify different types (intrusion actions) of human presence in thermal images. This chapter contributes to the evolving field of surveillance and security, offering a robust solution for identifying and monitoring human intrusions in various scenarios. This work proposes a solution for the shortcomings of visual surveillance in lowlight conditions by utilizing thermal images for human action detection and localization. Collectively, these chapters form a cohesive narrative that showcases the transformative potential of deep learning in thermal image analysis for hotspot detection. The integration of self-supervised learning, human-intrusion detection, and GANs reflects a holistic approach to addressing challenges in thermal imaging. This thesis not only contributes to the academic discourse but also offers practical solutions with wide-ranging applications in fields such as industrial monitoring, security, and environmental sensing.