Abstract
This letter introduces a new method for fault diagnosis of electrohydraulic actuators (EHA). Common faults include abnormal supply pressure, viscous friction, bulk modulus, and oil leakage. First, a validated model of an EHA is used to simulate faults at varying percent perturbations to generate data. The harmonics of the corresponding servo-valve control signal are calculated using sliding discrete Fourier transform (DFT) and stored. A fully connected neural network is then trained to predict fault cases using harmonic information. Simulation results show high accuracy for diagnosis of individual cases and ability to diagnose mixed cases. The actuator motion reference signal is also shown to impact overall diagnosability. Finally, the method was tested on experimental data consisting of nominal and low-pressure cases. Random disturbances were added to simulation training data to allow more robustness and resulted in 80% accuracy on the 10 experimental datapoints. It is demonstrated that a validated model can be used to diagnose faults in actual experimental setup. Future work will be done to diagnose non-pressure experimental cases.
