Multi-temporal-scale event detection and clustering in IoT systems

Abstract

Sensor-based Internet of Things (IoT) systems detect events from the data stream and take appropriate actions through event processing. The core of event processing, event rules, are typically defined manually by domain experts. However, there are limitations to domain experts manually setting rules for all the unlabeled events in the runtime of IoT systems. Therefore, there is a need for methods that support the generation of rules for unlabeled events. This study addresses this issue by adding two phases to the existing event processing. The first phase is the detection of unlabeled events from the data stream. Considering the characteristics of IoT systems, we propose Multi-Temporal-Scale Sampling (MulTemS), an extension of anomaly detection techniques that can detect events of various temporal-scales from multivariate time-series data. The second phase is the formation of clusters among similar events. We propose Feature-based Clustering Number prediction and Clustering (FeatCNC), which predicts the number of clusters through feature extraction and performs domain-neutral clustering. Through experiments, we demonstrate that MulTemS can effectively detect events of multiple temporal-scales, and FeatCNC can reliably cluster events across diverse domains. Additionally, we verify that the integration of these two phases results in the better formation of clusters that capture the characteristics of the events.