Rotor system fault diagnosis based on simulation data

General regression neural network (GRNN) was used to diagnose three types of rotor system faults, namely, misadjusted trim-tab, misadjusted pitch control rod, and imbalanced mass. Three cascaded levels of networks were used to identify fault type, location, and extend, respectively. Simulation results, which include faulty rotor responses, hub loads, and fuselage vibration, from a coupled rotor-fuselage analytical model were used for training and testing. Artificial noises were added to simulation data to enhance network generality. Results show that: 1) trained GRNN is capable of diagnosing faults from noisy helicopter responses, which indicating the feasibility of simulation data-trained GRNN being used in rotor health and usage monitoring system (HUMS) development; 2) using noise-added training data can significantly improve GRNN's generality; 3) properly selecting the spread of network is important for fault diagnosis accuracy.