Advanced Deep Learning Systems for Monitoring and Analyzing Flare Activities in Video Data

Abstract

Flare stack is an important safety protection component in a petrochemical system that requires continuous inspection and monitoring. This kind of plant can transform the potentially polluted gases from an oil refinery plant into a more environmentally friendly form before emitting them into the air. However, because of the high-temperature, toxic gas-encompassed circumstance, the flare stack is among the least inspected plants despite its great importance. Therefore, to complete a flare stack real-time monitoring without making a worker exposed to the harsh environment, analyzing data collected from remote sensing devices such as camera-mounted Unmanned Aerial vehicles (UAV) and common CCTV cameras is an appealing topic.

Thisprojectaimstodevelopanadvanceddeep-learningmodelthatcanautonomously extract and analyze important features of flare and smoke activities hidden within collected video data. The major tasks can be concluded as flare-and-smoke classification, detection, segmentation, and a specialized focus on incipient-stage smoke detection. To start with, this thesis paper presents an in-depth review of significant achievements in Flare stack monitoring over the past decade and highlights ongoing challenges in Artificial Intelligence (AI)-enhanced remote inspection. Through the literature review, the writer gained a comprehensive knowledge of field improvements in the realm of AI, remote sensing, and flare stack inspection, and understood the trend and prevailing challenges of flare stack monitoring. The thesis then implemented both traditional and up-to-date technologies for flare stack classification, detection, and segmentation tasks, and demonstrated how cutting-edge machine learning and deep learning algorithms can significantly benefit flare stack monitoring. Later, the writer built up on these new techniques, making them better for the flare stack monitoring problem as well as more competitive in comparison with benchmark methodologies. Finally, this thesis presents limitations in current approaches and lists several aspects to which further attention should be paid.

Date of Award	26 Apr 2024
Original language	American English
Supervisor	Naoufel Werghi (Supervisor)