Semi-supervised Gaussian and t-distribution hybrid mixture model for water leak detection

The last few years have seen a great number of announcements and projections on cities of the future, where technological interconnected metering infrastructure is the main smart-grid unit, promoting higher sustainability due to its more efficient management capability. The water supply network is one of the grids that has been given additional attention due to the problem of waste caused by water leakage, usually requiring rapid detection for fast intervention to prevent high costs. With centralised information coming from the grid, like the measurement of pressure and flow, it is revealed that anomaly detection could be an important tool for quick automatic detection without needing permanent analysis by a human operator. However, there is a need for a more robust approach, especially when noisy data are present. In this paper, we propose the implementation of a new approach based on a hybrid expectation maximization (EM) Gaussian model combined with a t-distribution mixture. This approach is compared to both a pure EM Gaussian mixture model and a t-distribution mixture model that can use labelled data or not. Each EM algorithm was applied to real data acquired from a water supply grid with the aim of automatically detecting water leaks. Using the newly developed approach, the results show that detection is both possible and more accurate for this type of database.