Fault detection remains a key problem for reducing the operation and maintenance costs of marine hydrokinetic (MHK) turbines. With this in mind, a deep bidirectional long short-term memory (Bi-LSTM) network for rapid detection of MHK turbine faults is presented. The effectiveness of the proposed scheme is validated using simulated time-series sensor data gathered from a novel Fatigue, Aerodynamics, Structures, and Turbulence (FAST)-based MHK turbine simulation platform. Four-factor analysis of variance is performed along with post-hoc testing at the 95% confidence level to establish the model's capabilities. Operating conditions, input length, training pitch, and generalization pitch are used as factors in this experiment. Models are trained on data consisting of 500 baseline and 500 faulty examples of a single blade pitch imbalance. To evaluate generalization performance, models are applied, without fine-tuning, on the remaining levels of fault. Over 70% mean generalization accuracy in all cases, with an optimal accuracy over 95% given 1 second lengths of data, is observed. No significant difference is found due to varying operating conditions, indicating a robust model. Training on lower levels of fault demonstrated greater generalization capability. To the best of our knowledge, this is the first time Bi-LSTM has been applied as a building block, and that a deep learning-based method has been applied to only time-series sensor measurements for fault detection in MHK turbines.