III: Small: Uncovering the Myths of Unlikelihood: Granger Graphical Models for Anomaly Detection in Multivariate Time-Series Data

点击翻译按钮获取中文摘要

The project aims to develop effective approaches to anomaly detection from high-dimensional time series data, motivated by applications such as oil drilling, semiconductor fabrication, and railroad operation.The proposed approach takes advantage of Granger Graphical models, which uncover the temporal dependencies between variables, to efficiently compute a robust correlation anomaly score for each variable and obtain insights regarding the causes of anomalies. The project develops effective approaches to addresses several specific challenges that arise in real-world applications of anomaly detection, including (1) nonlinear temporal dependencies; (2) hidden variables; and (3) massive amounts of data. The resulting algorithms will be evaluated on two real production systems: an oil-field mechanical system and a semi-conductor fabrication system.The project is expected to advance the state of the art in anomaly detection for high-dimensional time series data that arise in many application domains. It offers research-based training opportunities at the intersection of machine learning, data mining, and intelligent production management, as well as operational research in general.Workshops and mini-courses will be organized to introduce advanced machine learning techniques to students, practitioners, and researchers in production management. The anomaly detection code and data sets will be freely disseminated to the broader research and educational community. Additional information about the project can be found at: http://www-bcf.usc.edu/~liu32/ggm.htm.