A neural network approach to failure diagnostics for underwater vehicles

The author addresses the proposed use of Kalman filters and artificial neural networks to provide the detection and isolation of impending system failures. Such system health diagnosis is necessary for the overall success of mission controllers for autonomous underwater vehicles (AUVs). Two examples of network designs are given. The first addresses the identification of anomalous changes to the vehicle's acceleration behavior resulting from possible propulsion system changes or loss of propulsion efficiency from fouling. The second example relates to the identification of excessive frictional loads in the propulsion drive train that may cause motor failure. In each case, the training method and the resulting decision surface characterization of the networks so designed are described.<<ETX>>