In underwater environment it is not easy to predict all situations and phenomena such as current force, sudden change of temperature and so on. Installation of a proper scheme to cope with un-expected troubles is, therefore, essential when Autonomous Underwater Vehicles (AUVs) carry out their mission out of human's reach. In order to supervise whether the vehicle operates itself in a appropriate way, this paper proposes a model based approach to the self diagnosis for AUVs. This system includes self diagnosis which is carried out based on a dynamics model of an AUV and an active mechanism to get desirable information for diagnosis. The dynamics model is constructed by artificial neural networks taking advantage of its flexible learning ability. When a sensor is found to be defective, dead reckoning using the corresponding output of the dynamics model can be introduced. The performance of the proposed system was examined by implementing it to “The Twin-Burger”, a test-bed AUV. It is shown that the system detects failures of onboard sensors and actuators, and then select proper action schemes to minimize the damage to the AUV.