|Stochastic Systems Group|
Distributed Change Detection in Dynamic Systems
In modern systems such as airplanes or manufacturing plants, it is anticipated that a large number of distributed sensing and processing nodes will be employed. The ability to handle massive data and detect deviations from normal or expected system behavior in a timely manner is required. This brings new challenges. Most of the existing change detection strategies for dynamic systems are centralized. This may not be feasible for large systems due to bandwidth and energy constraints. In dynamic systems, the distribution of observations is not readily available, but evolves as the dynamics proceeds. The correlation between observations in a general dynamic system is difficult to characterize. Also, the derivation of optimal decision rules when the inputs are spatially correlated in the classical distributed detection problem formulation largely remains an open question. Nevertheless, even if optimal cooperative rules prove intractable, seeking effective suboptimal decision rules is worthwhile due to the prevalence of correlated signals in many applications.
In this talk, we consider change detection in dynamic systems using multiple redundant sensors. System behaviors before and after change can be modeled as two hypotheses. The dynamic systems may exhibit complex nonlinearity and non-Gaussianity. Therefore, particle filters are implemented at the local sensors to predict the system state. Under the assumptions of independent, full correlated or partial correlated observations, efficient distributed fault detection algorithms are proposed, including local detector design and decision fusion rule design. Illustrative examples are presented to demonstrate the effectiveness of our approaches.
Problems with this site should be emailed to firstname.lastname@example.org