The design and evaluation via simulation of an observer for nuclear reactor fault detection is reported. The method used is an extension of that proposed by Beard (in 1971) that allows actuator, sensor, and system dynamic faults to be detected and localized by studying the asymptotic response of an error signal. Signal noise and modeling errors can cause false alarms and/or failure to detect real faults. These types of error are characterized by their respective first hitting times of a decision threshold. Cost functions are then defined for each error and optimization is used to select observer parameters. The final design was evaluated by simulation on both a one-group linear reactor model and a six-group nonlinear one. The method was shown both to detect and localize faults and to be robust against measurement noise and modeling errors.