Automated detection of anomalies in complex real-world time-series
data is typically performed using a combination of knowledge
engineering approaches that suffer from frequent false alarms and/or
inability to detect anomalies before they lead to critical system
failures. The domains of interest often involve thousands of sensors
over millions of time samples, making it very difficult to overcome
the limitations of such knowledge intensive approaches using
straightforward application of standard regression techniques
(e.g. neural networks with error bars). To address such concerns, I
developed the ELMER (Envelope Learning and Monitoring via Error
Relaxation) system. ELMER automatically learns high and low limit
functions ("envelopes") from data. Toward reducing operations costs
while improving reliability and automation, ELMER focusses on learning
envelopes which maintain low false alarm rates. It does so in an
anytime fashion, beginning with relatively wide limits, such as
constant red-lines which are not context-sensitive. As promising
context-defining sensors are identified during learning, it forms
tighter and more input-conditional envelopes. ELMER is being
evaluated for a variety of NASA domains --- including both ground and
onboard operations and both earth-orbiting and deep space missions.
In this talk, I will present examples and discuss some of the
technical advances behind the ELMER work, which includes work on
asymmetric regression cost functions, learning Bayes nets, and feature
selection/construction.
Date: Thurs., February 19; Time: 4:15-5:30PM; Place: Gates 100
Return to seminar schedule.