Session: 04-01: Leak Detection Session - Part A

Submission Number: 185761

An Acoustic Machine Learning Method for Leak Detection in Short-Run Liquid Pipelines 

Pipeline leak detection is a crucial need for operational safety and environmental stewardship. Leak detection (LD) technology has advanced significantly in recent years, but most systems rely on similar existing flow and pressure measurements. Alternatives such as fiber optic remain difficult and costly to install on existing infrastructure. There remains a significant gap in methodologies for short pipelines that lack high-quality flow measurements, such as stub lines or laterals.

This study introduces a novel leak detection methodology using a dynamic pressure sensor as a hydrophone, collecting high frequency acoustic data and using a machine learning model in an edge computing platform to determine leak presence. This acoustic machine learning (AML) model continuously adapts to changing operating conditions.

In the proposed AML model, data is collected using a high-precision piezoelectric dynamic pressure sensor at approximately 10 kHz. A data acquisition system streams the data to an on-site edge computer for analysis. The data is processed in two frequency ranges to distinguish mechanical pipeline vibrations from high-frequency acoustic phenomena associated with leaks. Spectral analysis using the Welch Power Spectral Density (PSD) method extracts frequency-domain features including peak frequency, band power, entropy, and RMS. These features are input into a gradient-boosted decision tree classifier, trained on labeled datasets representing leak and non-leak conditions. The system monitors baseline signatures and automatically recalibrates as conditions change, deploying challenger models to maintain detection performance.

Two case studies are presented: a controlled flow loop with water at 200 psi and a 1-mile pipeline segment with produced water at 400 psi. The system detected 14 induced leaks in each (15 - 40% leaks in the test loop, 3 - 5% in the field). The AML model achieved >90% validation accuracy, >99% recall, and >95% precision while maintaining low computational overhead, making it suitable for real-time inference on ARM-based edge devices in remote areas.

The AML model is applicable to high-pressure, short-distance liquid pipelines where acoustic signals propagate effectively. A single sensor can detect leaks 1-2 miles upstream and downstream (pressure-dependent due to acoustic attenuation). It can operate with external accelerometers when internal mounting is not possible. The approach functions as a standalone module, a confirmation layer for existing flow-balance systems, or an early-warning tool for high-consequence areas such as river crossings.

LD technology integrating sensing, signal processing, and automated classification within a closed-loop system represents the next generation of pipeline safety tools, expanding coverage across high-risk, short-run pipelines lacking traditional instrumentation.

Presenting Author: Chris Johnston, Pipewise Technology

Presenting Author Biography: Chris Johnston is an R&D engineer based in Alberta, Canada with a passion for developing new solutions for the energy sector. He holds three patents in sand separation technology and has spent the last three years developing new leak detection methods and technology for the modern oil and gas industry. Chris also serves on multiple boards for education and works to advance STEM programs in northern Alberta to foster the next generation of problem solvers and innovators.

Authors:

Chris Johnston Pipewise Technology
Akshay Rajesh Pipewise Technology Ltd.
Yashar Rahimi Pipewise Technology Ltd.