This paper presents a general approach for building a robust and efficient supervision system for fault detection and recovery. The approach uses a set of monitors that obtain information about the system state and, instead of detecting fault states directly, detects significant differences between perceived and expected states. To deal with uncertainty in the knowledge about the system state and the result of some actions, it uses a POMDP model to decide when it is worthwhile to take recovery actions. We present the general approach and show its application with an indoor mobile robot by reporting and evaluating comparative results for different solutions.