Chapter 35 Novelty Detection

Encyclopedia of Structural Health Monitoring. Edited by Christian Boller, Fu-Kuo Chang and Yozo Fujino  2009 John Wiley & Sons, Ltd. ISBN: 978-0-470-05822-0. of possible failure modes, the effects of which on observable (sensor) data are often poorly defined. To compound this, examples of abnormal behavior in high-integrity systems are few and far between; usually, there are insufficient examples of failure to construct accurate fault-detection systems. As a result, conventional fault-specific failure-detection schemes are usually limited to identifying a small subset of known, well-understood modes of failure. An alternative to identifying rare and unexpected modes of failure is the novelty detection paradigm [1–5], in which a model of normality is constructed from normal system data. Departures from normal behavior are classified as novel events. Novelty detection is alternatively known as one-class classification [6] or outlier detection [7].