Steffen Illium

In collaboration with Munich's municipal utility provider, Stadtwerke München (SWM), this project explored the feasibility of using acoustic monitoring for early leak detection in water pipe infrastructure. The primary goal was to develop machine learning models capable of identifying leak-indicating sound patterns within a real-world operational environment.

The objective was to investigate and develop methods for the automated detection and localization of leaks in urban water distribution networks using acoustic sensor data.

Methodology & Activities

Data Acquisition: Sensor networks comprising contact microphones were deployed across sections of Munich's suburban water network to capture continuous acoustic data.
Signal Processing: Raw audio signals were pre-processed and transformed into mel spectrograms, converting the time-domain audio data into image-like representations suitable for analysis with computer vision techniques.
Model Development: Various machine learning approaches were evaluated. Deep neural networks, particularly Convolutional Neural Networks (CNNs), were trained on the spectrogram data to classify segments as containing leak sounds or normal background noise.
Analysis & Validation: The performance of the models was assessed against ground truth data provided by SWM, identifying both the successes and challenges of applying these methods in a complex, noisy, real-world setting.

A field technician using an acoustic sensor rod to detect leaks on a water pipe — Me representing a field technician using an acoustic leak detection device equipped with headphones to pinpoint potential water leaks.

Team performing acoustic leak detection around a manhole cover — SWM personnel conducting on-site acoustic measurements around a manhole, illustrating data acquisition for the leak detection project.

Key Findings & Outcomes

The project demonstrated the potential of deep learning models applied to mel spectrograms for identifying relevant acoustic features indicative of water leaks.
CNN-based approaches showed advantages over traditional machine learning methods in capturing the complex patterns associated with leak sounds amidst background noise.
Significant insights were gained regarding the practical challenges of sensor deployment, data quality variability, and noise interference in real-world utility networks.
The research conducted within this project formed the basis for several scientific publications and a paper writeup (Müller et al., 2021).

Workers performing actions on a pipe section — On-site 'arificial leak' creation in the proximity of deployed acoustic sensors.

Project team next to an SWM utility vehicle after field work — Project team from SWM and LMU after a session of field data collection.

This applied research project provided valuable experience in handling real-world sensor data, adapting machine learning models for specific industrial challenges, and collaborating effectively with industry partners.

ErLoWa Leak Detection

Methodology & Activities

Key Findings & Outcomes