Using Seed Transducers for Describing and Synthesizing Structural Time-Series Constraints
We introduce a new way to describe a large family of global constraints over time-series data. These constraints describe various features of the time series, like the number of peaks, or the steepest ascent, which together characterise the time-series. Those features often are linked to physical limits of the generator of the series, for example a power plant, or a data sensor. We provide a formalisation for the description of the constraints based on "seed transducers". Based on the description we can automatically generate constraint propagators, based on automata with counters, and efficient constraint checkers. This new description not only unifies the structure of the existing thirty time-series constraints in the global constraint catalog, but in addition leads to a total of over 400 new constraints and predicates. As a practical use case we show how to detect problems in the time series data coming from sensors in the Campus21 project with the new constraints.
Researcher at Insight Centre for Data Analytics
Wednesday, 18 February, 2015 (All day)