Automatic Anomaly Detection for Sewage Network Sensors

The DONUT project (Ahm et. al, 2020) aims to provide a holistic view of the urban water system by making continuous water level measurements using a low-cost sensor network. A crucial instrument for ensuring the reliability of the network is the ability to detect anomalous events/ observations, e.g., due to faulty sensors or anomalous water conditions. In this paper, we propose two approaches (both based on machine learning models) to address this issue: an anomaly detection model for isolated sensors and a more expressive anomaly detection model relying on measurements from neighboring sensors. Our preliminary experiments show promising results for both approaches.