Bayesian Approaches to Rare Event Prediction in Multivariate Time Series

The problem of rare events in large multivariate time series is explored within the context of a condition monitoring and failure prediction problem. A data set from DataPath’s MaxView software is used. We establish a baseline for prediction on this data set. This report examines feature extraction and similarity metrics for multivariate time series. The aspect of rare events is most relevant in sections discussing methods for algorithm performance evaluation and similarity metrics, which are used for exploratory data analysis. The modeling sections pay attention to the sensitivity of the methods to outliers and non-Gaussian behavior. After reviewing standard time series models in the linear Gaussian framework, extensions are explored; common extensions include the ability to handle rare or non-Gaussian events, non-parametric versions, and factorial models. I propose various extensions to standard time series models, which are tailored to the failure prediction problem.