Anomaly Detection in Real-Valued Multidimensional Time Series

We present a new algorithm for detecting anomalies in real valued multidimensional time series. Our algorithm uses an exemplar-based model that is used to detect anomalies in single dimensions of the time series and a function that predicts one dimension from a related one to detect anomalies in multiple dimensions. The algorithm is shown to work on a variety of different types of time series as well as to detect a variety of different types of anomalies. We compare our algorithm to other algorithms for both one-dimensional and multidimensional time series and demonstrate that it
improves over the state-of-the-art.