Near Real-Time Automated Detection of Small Hazardous Liquid Pipeline Leaks Using Remote Optical Sensing and Machine Learning

The prevailing leak detection systems used today on hazardous liquid pipelines (computational pipeline monitoring) do not have the required sensitivities to detect small leaks smaller than 1% of the nominal flow rate. False alarms of any leak detection system are a major industry concern, as such events will eventually lead to alarms being ignored, rendering the leak detection system ineffective [1]. This paper discusses the recent work focused on the development of an innovative remote sensing technology that is capable of reliably and automatically detecting small hazardous liquid leaks in near real-time. The technology is suitable for airborne applications, including manned and unmanned aircraft, ground applications, as well as stationary applications, such as monitoring of pipeline pump stations. While the focus of the development was primarily for detecting liquid hydrocarbon leaks, the technology also shows promise for detecting gas leaks. The technology fuses inputs from various types of optical sensors and applies machine learning techniques to reliably detect “fingerprints” of small hazardous liquid leaks. The optical sensors used include long-wave infrared, short-wave infrared, hyperspectral, and visual cameras. The utilization of these different imaging approaches raises the possibility for detecting spilled product from a past event even if the leak is not actively progressing. In order to thoroughly characterize leaks, tests were performed by imaging a variety of different types of hazardous liquid constitutions (e.g. crude oil, refined products, crude oil mixed with a variety of common refined products, etc.) in several different environmental conditions (e.g., lighting, temperature, etc.) and on various surfaces (e.g., grass, pavement, gravel, etc.). Tests were also conducted to characterize non-leak events. Focus was given to highly reflective and highly absorbent materials/conditions that are typically found near pipelines. Techniques were developed to extract a variety of features across the several spectral bands to identify unique attributes of different types of hazardous liquid constitutions and environmental conditions as well as non-leak events. The characterization of non-leak events is crucial in significantly reducing false alarm rates. Classifiers were then trained to detect small leaks and reject non-leak events (false alarms), followed by system performance testing. The trial results of this work are discussed in this paper.