Relationship Discovery for Heterogeneous Time Series Integration: A Comparative Analysis for Industrial and Building Data

Cyber-physical systems like buildings and power plants are monitored with ever-increasing numbers of sensors, gathering massive and heterogeneous time-series datasets collected in data lakes. Appropriate meta-data, describing both the function and location of each sensor, is essential for any profitable use of the data but is often not available or incomplete. Particularly, information about related sensors, meaning sensors belonging to the same functional subsystem, might be hard to derive if appropriate meta-data is unavailable. While various approaches exist for automatic meta-data extraction from relational databases, the unique characteristics of heterogeneous time-series data necessitate specialized algorithms. Among the general algorithms developed for time-series meta-data inference, only a few are concerned with relationship discovery despite the critical importance of this information in many meta-data formats. Nevertheless, other domains offer a variety of measures for pairwise relationship discovery in homogeneous time-series collections. This paper consolidates these measures and evaluates their performance for identifying related but heterogeneous time series from the same functional subsystem within industrial facilities. We evaluate the methods on a collection of different datasets to extract promising relationship measures from the literature and show that there are other better-performing candidates than the common Pearson Correlation Coefficient.