Robert Posey

Excited Group Particle Filters for Fault Detection and Identification

Date: WEDNESDAY, DEC 21 at 10: AM

ABSTRACT:

This Thesis presents new particle filter called Excited Group Particle Filter (EGPF), which is optimized for Fault Detection and Identification (FDI) in complex systems with large numbers of fault modes. The EGPF uses the Rao-Blackwellised concept in scheme that controls both state and particle resolution dynamically in non-steady state conditions. We first show that the FDI performance of several popular Rao-Blackwellised based particle filters degrades rapidly as the number of potential system modes increases. Then we will show by discussion and demonstration that the EGPF's performance is nearly independent of the number of potential system modes. We will also show and demonstrate that any attempt to detect transitionally excited system faults can dramatically increase false alarms rates in current Rao-Blackwellised Particle Filters in fault free systems, while going undetected when actually present. In contrast, we show that the EGPF detects transitionally excite system faults when they exist, and rejects them when they are absence. The EGPF does all this while reducing computational complexity, and improving noise rejection.

Thesis Committee:
Bahram Shafai (advisor)
Dana Brooks
Eric Miller