This research investigates a novel data-driven approach to condition monitoring of Electrical-Mechanica
l Actuators (EMAs) consisting of feature extraction and fault classification. The approach is designed to accommodate varying loads and speeds since EMAs typically operate under non-steady conditions. Since many common faults in rotating machinery produce unique frequency components, the approach is based on signal analysis in the frequency domain of both inherent EMA signals and accelerometers. The feature extraction process exposes fault frequencies in the signal data that are synchronous with motor position through a series of signal processing techniques consisting of digital re-sampling to the position domain, Power Spectral Density (PSD) computation to the frequency domain, and feature reduction. The reduced dimension feature is then used to determine the condition of the EMA with a trained Bayesian Classifier. Signal data collected from EMAs in known health configurations is used to train the algorithms so that the condition of EMAs with unknown health may be predicted. A passive, linear load test fixture is used to provide a known load (2,400-lbf) on a MOOG industrial MaxForce EMA used for the testing. A seeded fault testing methodology is used to induce known faults in the ball screw and then used as training and validation data for the proposed work. Various desired driving commands are utilized to simulate "real-world" conditions. Laboratory results show that EMA condition can be determined over multiple operating conditions. Although the process was developed for EMAs, it can be used generically on other rotating machine applications as a Health and Usage Management System (HUMS) tool.
This research investigates a novel data-driven approach to condition monitoring of Electrical-Mechanical Actuators (EMAs) consisting of feature extraction and fault classification. The approach is designed to accommodate varying loads and speeds since EMAs typically operate under non-steady conditions. Since many common faults in rotating machinery produce unique frequency components, the approach is based on signal analysis in the frequency domain of both inherent EMA signals and accelerometers. The feature extraction process exposes fault frequencies in the signal data that are synchronous with motor position through a series of signal processing techniques consisting of digital re-sampling to the position domain, Power Spectral Density (PSD) computation to the frequency domain, and feature reduction. The reduced dimension feature is then used to determine the condition of the EMA with a trained Bayesian Classifier. Signal data collected from EMAs in known health configurations is used to train the algorithms so that the condition of EMAs with unknown health may be predicted. A passive, linear load test fixture is used to provide a known load (2,400-lbf) on a MOOG industrial MaxForce EMA used for the testing. A seeded fault testing methodology is used to induce known faults in the ball screw and then used as training and validation data for the proposed work. Various desired driving commands are utilized to simulate "real-world" conditions. Laboratory results show that EMA condition can be determined over multiple operating conditions. Although the process was developed for EMAs, it can be used generically on other rotating machine applications as a Health and Usage Management System (HUMS) tool.
