Robust and Adaptive Filtering of Multivariate Online-Monitoring Time Series

We propose a new regression-based filter for multivariate time series that separates signals from noise and outliers in real time. The new method merges the advantageous properties of two existent filtering procedures for online-monitoring time series. Our multivariate and robust procedure yields signal estimations at the right end point of a moving time window whose width is adapted to the current data situation. Since the proposed filter works in real time, it can be used, e.g., to lower the rate of false positive threshold alarms of intensive care online-monitoring systems. Keywords— Online-monitoring, multivariate time series, signal extraction, robust regression, window width adaptation.