Comparison of Signal Pre-processing and Machine Learning Modelling for Water-leak Detection Using Vibration and Pressure Data

Leak detection is important to enable automatic and early identification of leakages in water distribution systems, which may prevent water wastage, reduce the environmental impact of leakages, and also avoid structural damages to pipe networks. While there are different sensors, hardware and software-based methods for leak detection, this field still faces issues, especially regarding data scarcity and imbalance. In this paper, a comparison of different Machine Learning (ML) models for leak detection and leak type classification has been carried out, addressing data augmentation and balancing issues, and providing a comparison on the performance of pressure and accelerometric data for leak monitoring with ML. Compared to the state-of-the-art, mainly focusing on leak detection, the focus has also been on leak type identification, which may be useful for diagnosing pipe damage. Results indicate that accelerometer data is more consistent across tasks and performs better than pressure data for multi-class leak classification. The best-performing model is an Ensemble Bagged Trees classifier using accelerometer data and features extracted from 5 s long windows, and it achieved 91.7% accuracy in multi-class leak classification. While the proposed approach still faces issues in identifying longitudinal cracks, it provides a solid base for leak classification with ML and it identifies the accelerometer as a viable, non-invasive solution for leak monitoring.