Improving the Reliability, Detection, and Accuracy Capabilities of Existing Leak Detection Systems (CPMs) Using Machine Learning

Overview

Main Objective

This project will address three primary leak detection systems gaps: [1] the ability to find smaller leaks, [2] the ability to find leaks faster and [3] the ability to find leaks more reliably (higher confidence, lower false alarms) than is possible with conventional CPM systems. The resulting machine learning (ML) CPM development will not be a replacement of current CPM algorithms, but will be an enhancement that will augment current CPM algorithms with highly sophisticated machine learning models to find smaller leaks than are possible today by going below the threshold (or noise floor) of those algorithms.

Public Abstract

Computational Pipeline Monitoring (CPM) is by far the most prominent leak detection system (LDS) used for liquid leak detection in North America, since most of the needed infrastructure (e.g., SCADA) is already in place. Current SCADA methods suffer from limitations, including difficulty in detecting leaks during line transients and slack conditions, less reliable localization of the leak, and difficulty in detecting small leaks (<1% pipeline throughput) (U.S. Department of Transportation) (Frost, 2010).

A recent study by the U.S. Department of Transportation Pipeline and Hazardous Materials Safety Administration (PMHSA) that analyzed leak incidents between January 2010 and July 2012 for hazardous liquid pipelines concluded that CPM and SCADA systems identified only 20% and 28%, respectively, of leaks. Specifically, the study concluded that, "with internal Leak Detection Systems (LDS), false alarms are often a tradeoff with sensitivity. With a higher threshold for detection, fewer random and transient effects will have an impact on the imbalance or pressure deviations. Of course, this also affects reliability by increasing the probability of misses."

This project outlines a promising approach, using proven machine learning techniques (Araujo), together with CPM data, for reliably identifying much smaller leaks than are possible with typical CPM threshold methods, improving the leak alarming reliability by decreasing the probability of false alarms.

PRCI Catalog No. PR-015-203900-R01

Summary and Conclusions

The project's main success was development of the algorithm and framework that describes how pipeline operators can integrate machine learning algorithms with computational pipeline monitoring data (i.e. pressure/volume). The work within the project also provides much perspective on how much data is needed to adequately train a model on a pipeline. Unfortunately, a limited data set was utilized in this work as volunteered by participating operators so future work should still be funded. The further research should enhance the capabilities developed during the project and presented in the final report. To continue this work, more data is needed. Optimally, the data would cover long periods and include multiple examples of commodity withdrawals spread throughout the data and occurring in a variety of operational configurations.

Relevant Files & Links