Model-Based Reasoning with Temporal Belief Networks

Conventional approaches to spacecraft fault detection and isolation (FDI) suffer from a combinatorial explosion in trying to represent the myriad number of ways failures can occur and propagate in a modern complex spacecraft. To get around this problem, we propose to explore the feasibility of developing a model based reasoning (MBR) technique using temporal belief networks (TBN) to model only the normal operation of the spacecraft subsystems. The natural graphic representation of this approach can then take advantage of the human supervisor¿s natural ability to recognize normal and abnormal operating modes, without requiring the necessity of a detailed and lengthy search through candidate failure modes. The addition of a temporal dimension to traditional Bayesian belief network structures facilitates analysis of time-varying commands and their current and predicted effects on the spacecraft. Our proposed incorporation of sensor measurements as belief network ¿evidence¿ will allow the user to detect faults, while a ¿what-if¿ type of speculative analysis mechanism on the same network will diagnose faults. Efficient, single-pass evidence propagation algorithms, which are consistent with probabilistic semantics for representing uncertainty in complex systems, makes this TBN approach particularly suitable for real-time operational environments such as the spacecraft control stations envisioned under this approach.