An unscented Kalman filter based statistical failure detector

In the paper an approach for fault detection of information systems is presented. The characteristic of the underlying system is assumed to be unknown. The method is based on an adaptive unscented Kalman filter which models are derived from process output data. The ability to track an unknown evolving system over time and predict its internal state is covered by this approach within limits. Statistical techniques such as χ², generalized log-likelihood ratios or distance to standard deviation detect deviations from normal conditions. These techniques are used to classify faulty behavior.