Relation Inference among Sensor Time Series in Smart Buildings with Metric Learning

Abstract:

Smart Building Technologies hold great potential for improving residents' comfort while reducing energy footprints. These technologies require the knowledge about the sensing and control points in a building, including what they measure, where they are located, how they are connected, and more. However, this contextual information, often referred to as “metadata”, varies significantly in vocabulary and structure from one building to another. Consequently, obtaining this information is a costly, laborious process requiring cross-domain collaboration and necessitates an automated solution.

We will share results from a specific use case with relevance to efficient energy usage: the functional and spatial relationship between a Variable Air Volume (VAV) Box and associated Air Handling Unit (AHU). These two pieces of equipment are exposed to the same real-world events, e.g., a fan turning on or a person entering the room, thus exhibiting correlated changes in their sensor reading time series. This represents a common set of relationships across the Earth and Space Sciences, which gives our approach high potential for the AGU community. Indeed, as the underlying idea of our method is general, we expect it to apply to a broader set of relation inference for time-series data.

Our solution works in the frequency domain where analysis is less sensitive to noise in the data, compared to the time domain. Specifically, the method exploits metric learning to identify correlated streams: it learns to characterize pairs of similar or dissimilar time series with a neural network such that the similar ones are mapped to closer locations in a projected “space”. This learning paradigm could generally apply to data in the AGU community, e.g., to learn to recognize correlated geomagnetically induced current measurements and power grid observations.