Efficient Estimation for Diagnosis Using Factored Dynamic Bayesian Network

Robust and efficient estimation of hidden state variables of a system in the presence of measurement noise and modeling errors is crucial for online model based fault diagnosis of continuous systems. Dynamic Bayesian Networks (DBNs) provide generalized and systematic methods for reasoning under uncertainty. This paper presents an approach to improve estimation efficiency by partitioning the DBN into smaller factors and invoking estimation algorithms on each factor independently. The factors are generated by replacing some state variables with algebraic functions of some measurement variables, thus reducing the across-time links between these state variables. Hence, given the measurements, these state variables become conditionally independent of the state variables in other factors, and the states of each factor can be estimated separately. This paper derives an algorithm for generating these factors and presents experimental results to demonstrate the effectiveness of our factoring approach for accurate estimation of system behavior.