Unsupervised Anomaly Detection in Water System Networks using Recurrent Neural Networks

This was my thesis project where I worked with time series produced by flow and pressure sensors in a water distribution system. The purpose of this work was to automatically detect an anomaly. Interesting anomalies can be bursts, unusual demands, and illegal consumption. Water leakage detection and location is a very difficult problem, due to the lack of information about the water system, and a leak might not be easily detected or confused with other events. Thus, the great challenge of this work was the lack of labels. The methodology proposed detects anomalies in the water system distributions, with a focus on bursts, through the use of deep learning architectures, in particular, encoder-decoder architectures based on LSTM such as LSTM autoencoder, CNN-LSTM, CNN-BiLSTM, and SCB-LSTM. Predictions are adjusted by using a temperature correction model. Simulation analysis and experimental results in real data show the pitfalls of the unsupervised anomaly detection task in water distribution systems. It also highlights that the proposed methodology, although yielding some properties of interest, needs to be complemented with additional principles to the targeted end. Finally, it pinpoints meaningful differences between recurrent architectures.