Time-series clustering for sensor fault detection in large-scale Cyber-Physical Systems

Large-scale Cyber-Physical Systems (CPSs) are information systems that involve a vast network of sensor nodes and other devices that stream observations in real-time and typically are deployed in uncontrolled, broad geographical terrains. Sensor node failures are inevitable and unpredictable events in large-scale CPSs, which compromise the integrity of the sensors measurements and potentially reduce the quality of CPSs services and raise serious concerns related to CPSs safety, reliability, performance, and security. While many studies were conducted to tackle the challenge of sensor nodes failure detection using domain-specific solutions, this paper proposes a novel sensor nodes failure detection approach and empirically evaluates its validity using a real-world case study. This paper investigates time-series clustering techniques as a feasible solution to identify sensor nodes malfunctions by detecting long-segmental outliers in their observations' time series. Three different time-series clustering techniques have been investigated using real-world observations collected from two various sensor node networks, one of which consists of 275 temperature sensors distributed around London. This study demonstrates that time-series clustering effectively detects sensor node's continuous (halting/repeating) and incipient faults. It also showed that the feature-based time series clustering technique is a more efficient long-segmental outliers detection mechanism compared to shape-based time-series clustering techniques such as DTW and K-Shape, mainly when applied to shorter time-series windows.