This paper examines the development of a data-driven anomaly detection methodology for servo-actuated hydraulic valves installed in a gas turbine fuel delivery system. Degraded operation of these valves is a leading cause of unavailability for gas turbine driven power plants. Nearlyeighty potential features were generated from the limited raw sensors and control system signals through a combination of domain expertise, statistical feature extraction, and insight gains from prior physics-based simulations. Important features were down-selected by examining the behavior of the features using several years of operating data in conjunction with known field failures. Univariate statistical techniques were used to eliminate candidate features with limited capability to distinguish healthy from abnormal operation. A final machine learning model was generated using a process of recursive feature elimination. This paper will also touch on the practical implications of deploying a machine learning model in a real-time production environment.