Anomaly detection for multivariate time series in water distribution IoT systems

Water is a common good and a limited and strategic resource that needs to be protected and used in a sustainable way. Water resources are under unprecedented stress levels because of the combination of demographic evolution, the trend of increasing urbanisation, and climate change. These conditions force water operators to manage the water demand sustainably and efficiently, and identify and manage water leakages properly. In the last few years, ICT greatly improved and matured, producing novel data-based methodologies, new devices, and smart Decision Support Systems to help improve water resource management in man-made or natural environments, especially regarding monitoring and reporting quality, quantity, reuse of water, extreme events, and leakages. Some water utilities have already put into practise new digital technologies, such as smart sensors and actuators, drones and robots, and advanced data analytics models that provide detailed data, allowing deeper insights into water availability, quality, and use. In particular, many water distribution networks are continuously monitored through installed sensors that measure hydraulic parameters and water quality parameters. This continuous monitoring allows the collection of raw data to be used in multiple engineering applications, with flow rate and pressure data being the most widely used time series by water utilities. Performing leakage control, detecting changes in customer water demand or manoeuvring in the network can be summarised as physical anomaly detection. To perform this task, the time series of flow rate and pressure need to be analysed and passed through an anomaly detection model. The aim of this work is to create an anomaly detection model based on semi-supervised methods of deep learning to discover both sensor anomalies, caused by malfunctions and failures of the devices, and physical anomalies