Partially Observable Markov Decision Processes for Fault Management in Autonomous Underwater Vehicles

Abstract

This paper reports on a novel solution for fault detection, identification, and recovery in autonomous underwater vehicles using partially observable Markov decision processes (POMDP). It explains the reasoning for using - POMDP model-based fault manager over traditional static look-up tables. A generic long-endurance AUV was simulated with selected sub-systems as a proof-of-concept. The POMDP was applied to the fault management system to determine the actions of the vehicle given a fault in either the environment or vehicle. The simulation test bed that was developed is described. The results from simulations were performed using both synthetically generated and actually measured bathymetry data. The results are very encouraging.