Detecting novel fault conditions with hidden Markov models and neural networks

• Published in: Machine Intelligence and Pattern Recognition Vol. 16; pp. 525 - 536
• Main Author: Smyth, Padhraic
• Format: Book Chapter
• Language: English
• Published: 1994
• ISBN: 9780444818928; 0444818928
• ISSN: 0923-0459
• DOI: 10.1016/B978-0-444-81892-8.50050-X

This chapter presents the detection of novel fault conditions with hidden Markov models and neural. Fault detection and isolation for the antennas is often a complicated and lengthy process because of the fact that it can be difficult to establish the root cause of a problem in the communications chain. Loss of a spacecraft signal during a planetary encounter can result in the irretrievable loss of scientific data. Hence, there is considerable motivation to be able to quickly detect and isolate anomalous conditions. Similar scenarios occur in other applications, such as industrial plant process monitoring, biomedical health monitoring, and on-board vehicle fault diagnosis. A key point in the Markov monitoring approach is that the transition probabilities are not estimated from the data but rather are chosen a priori based on the long-term temporal characteristics of the system and prior knowledge concerning the system failure modes. In speech modelling, estimation algorithms are used to estimate the probabilities through maximum likelihood methods directly from data.