New Fault Diagnosis Procedure and Demonstration on Hydraulic Servo-Motor for Single Faults

This article presents a new online approach for predicting component degradation in hydraulic systems using a few distributed sensors. The procedure uses neural network nonlinear autoregressive exogenous (NARX) models to model a healthy hydraulic system and then finds the distributions of NARX prediction errors as the system operates in a variety of degraded states. Kullback–Leibler methods are then used to compute relative entropies between online error distributions and the various degraded state distributions. Finally, these data are then used to train a classification neural network, which takes relative entropies between known health state distributions and an online distribution as inputs and predicts component-level degradation. The procedure has been tested on an industrial hydraulic system under a variety of degraded conditions and has demonstrated a high level of accuracy in predicting the level of degradation. By being able to predict a health state, the cost and time required for maintenance can be improved, since knowledge of the system's health condition will improve the repair time and prevent unnecessary removal of healthy subcomponents.