Automated Detection of Small Hazardous Liquid Pipeline Leaks, 10-R8552
Maria S. Araujo
Sue A. Baldor
Shane P. Siebenaler
Edmond M. DuPont
Daniel S. Davila
Inclusive Dates: 04/01/15 – Current
Background — The prevailing leak detection systems used today (e.g., computational pipeline monitoring) are simply unable to detect small leaks (less than one percent of the line throughput). False alarms of any leak detection system are a major industry concern, as they lead to alarms being ignored, resulting in leak detection systems that are ultimately ineffective. The focus of this research is to detect small leaks while also characterizing and rejecting non-leak events to significantly reduce false positive rates.
Approach — The objective of this project is to develop and evaluate a technology that can be suitable to meet the goals set by the Pipeline Research Council International (PRCI) for small liquid leak detection — Detect small leaks (1 percent of line throughput) in less than 5 minutes, with 95 percent confidence level, under all operating conditions (i.e. steady-state, pump start/stop, line pack, take-offs, etc.).
The approach is focused on fusing input from the different sensors in a variety of different combinations (hyperspectral, infrared, and visual), and use feature extraction and classifier training techniques to identify unique features that could provide a more reliable "fingerprint" of not only small liquid leaks, but also non-leak events in a variety of operating conditions, to substantially reduce false positive rates. Leaks considered in this research include crude oil and refined products such as diesel, gasoline, and jet fuel, which would cover a large percentage of hazardous liquid pipelines in the U.S. Leaks and non-leak events are being simulated in various lighting and weather conditions that would allow for signatures of leaks and non-leak events to be well characterized.
Accomplishments — Small leaks were simulated using various fluids (e.g., crude oil, kerosene, etc.) on a variety of representative surfaces (e.g., gravel, grass, dirt, etc.). All imagers recorded data simultaneously, so as to allow for the construction of a composite image using data from all or a subset of the sensors. Scenarios that could trigger false alarms (i.e., non-leak events) were also simulated and characterized. Focus was given to highly reflective and highly absorbent materials/conditions that are typically found near pipelines such as water pools, presence of highly-reflective surfaces (e.g., insulation sheeting on pipes), concentrated zones of heat (e.g., from the sun and other warm fluids).
Feature characterization is being performed via a set of feature characterization techniques running in parallel. The first of these, mean spectral vector analysis, is currently under development. The second characterization technique under development is textural analysis, which involves a three-dimensional generalization of the ubiquitous two-dimensional Gabor filters. These filters mimic the human visual system for the purpose of edge and texture detection.