Robust anomaly detection for time series data in sensor-based critical systems.

Critical systems are defined as systems where a malfunction can cause significant economic damage, harm to the ecosystem, or endanger human life, potentially resulting in the loss of lives. Financial, Energy, Healthcare, Transportation, Telecommunication, Water Supply, and Defense systems are just some of the more prominent examples of such critical systems. A particular role is played by sensor-based critical systems, which are deeply intertwined between the digital and physical worlds, and where one or more physical quantities are sampled at given time intervals, forming the so called time series, for monitoring and/or control purposes. In both cases, it is important to detect potential anomalies as they may indicate either a malfunction of a sensor (which can lead to apply a control algorithm on unreliable data) or a problem with the system dynamics itself (and thus posing a risk to its integrity). Detecting anomalies in a timely and reliable manner in such systems, which typically operate in ``harsh" conditions like humidity, moisture, dust, vibration, shock, etc., and under strict operational requirements, is still an open problem. The main challenges are related to the data quality, the resilience required by the anomaly detection processes, the robustness of the algorithms, and the need to provide clear indications to the operators about the potential anomalies and their severity; while keeping the false positives under control and performing the detection in a timely manner, usually in real-time or near real-time. This thesis work aim at addressing such challenges, and in particular it brings four main contributions: an extensive and critical overview of anomaly detection, a methodological framework for reliably identifying the anomalies in the context of sensor-based critical systems, a software library implementing such methodology, and a real-world case study in the Water Distribution Systems domain.