As the capabilities of autonomous underwater vehicles (AUVs) improve, the missions become longer, riskier, and more complex. For AUVs to succeed in complex missions, they must be reliable in the face of subsystem failure and environmental challenges. In practice, fault detection activities carried out by most AUVs employ a rule-based emergency abort system that is triggered by specific events. AUVs equipped with the ability to diagnose faults and reason about mitigation actions in real time could improve their survivability and increase the value of individual deployments by replanning their mission in response to failures. In this paper, we focus on AUV autonomy as it pertains to self-perception and health monitoring and argue that automatic classification of state-sensor data represents an important enabling capability. We apply an online Bayesian nonparametric topic modeling technique to state-sensor data in order to automatically characterize the performance patterns of an AUV, then demonstrate how in combination with operator-supplied semantic labels these patterns can be used for fault detection and diagnosis by means of nearest-neighbor classifier. The method is applied in post-processing to diagnose faults that led to the temporary loss of the Monterey Bay Aquarium Research Institute's Tethys long-range AUV in two separate deployments. Our results show that the method is able to accurately identify and characterize patterns that correspond to various states of the AUV, and classify faults with high probability of detection and no false detects.