Pipeline Leak Detection via Machine Learning

Time to read
1 minute
Read so far

Pipeline Leak Detection via Machine Learning

Wed, 11/20/2019 - 13:23
Pipeline Leak Detection via Machine Learning
Pipeline Leak Detection via Machine Learning

As physical entities, pipelines are subject to numerous points of failure including corrosion, mechanical damage, and natural hazards. Despite being infrequent, pipeline failure can have disproportionate consequences resulting from environmental clean-up and lost production. Best practices in pipeline risk management employ both leak-prevention and leak-detection strategies, the latter to reduce leak impacts via earlier detection, resolution, and remediation. However, sensor systems for leak detection (e.g. fiber optics) c an be prohibitively costly to install on legacy pipelines. Inferential (soft) sensing approaches using hydraulic modeling can be effective, but are vulnerable to measur ement uncertainties, noise, and calibration drifts. There is a clear need for models that can tolerate such phenomena while minimizing detection time and false-positive and false-negative errors.


We propose an inferential sensing framework using machine l earning as a cost-effective leak detection system. We treat leak detection as an instance of anomaly detection; a model of normal behavior is built, and deviations from that model trigger alarms. Intelligent anomaly detection designs are two-stage models, with normal behavior and deviations from it learned separately, often via completely different algorithms. In this instance, sensor data streams (temperature, pressure, etc.) are treated as time series, and forecasting models (deep neural networks) are learned from the delayed normal behavior of the pipeline. These forecasts are designed to predict the current, rather than future, pipeline behavior from past observations. The anomaly detector (support vector machines or shallow neural networks) learns to compare the prediction against the actual, current observation and raise an alarm when predictions and reality diverge significantly. When tested against real-world pipeline data with a nominal flow rate of 350 m3/hr, the inferential sensing framework correctly identified a 35 m3/hr leak within 5 minutes and a 5 m3 /hr leak within 48 minutes, with no false positives.

Contributing Companies and Authors

Article Categories

View the full Article within our Journal: