Sequential Scheduling of Observations in Diagnosis of Continuous Dynamic Systems

A fault in a continuous dynamic system is a non nominal value of a physical parameter. Parameter values can usually be estimated with uncertainty given the values of the output variables of the system that have been observed over a time interval. A diagnosis consists of several probability density functions, one for each parameter. Such a diagnosis may be noisy, especially if it is unlikely that the system is affected by several faults at the same time. In order to reduce the uncertainty about the diagnosis, new observations are needed. However, in some contexts observations are expensive, thus there is a maximum number of observations that can be gathered. This paper faces the problem of choosing the best schedule of a limited amount of observations within a given time horizon, in order to minimize the uncertainty of the diagnosis. Several scheduling policies have been tried in order to experimentally compare them.