Deep Learning Based Anomaly Detection in Water Distribution Systems

Abstract

Water distribution system (WDS) is one of the most essential infrastructures all over the world. However, incidents such as natural disasters, accidents and intentional damages are endangering the safety of drinking water. With the advance of sensor technologies, different kinds of sensors are being deployed to monitor operative and quality indicators such as flow rate, pH, turbidity, the amount of chlorine dioxide etc. This brings the possibility to detect anomalies in real time based on the data collected from the sensors and different kinds of methods have been applied to tackle this task such as the traditional machine learning methods (e.g. logistic regression, support vector machine, random forest). Recently, researchers tried to apply the deep learning methods (e.g. RNN, CNN) for WDS anomaly detection but the results are worse than that of the traditional machine learning methods. In this paper, by taking into account the characteristics of the WDS monitoring data, we integrate sequence-to-point learning and data balancing with the deep learning model Long Short-term Memory (LSTM) for the task of anomaly detection in WDSs. With a public data set, we show that by choosing an appropriate input length and balance the training data our approach achieves better F1 score than the state-of-the-art method in the literature.