Training a neural observer using a hybrid approach

Abstract

In this work, we use the approach based on observers such as the neural observer in order to introduce the diagnosis of nonlinear systems. There are different techniques for training the neural networks. Among these techniques, we quote the backpropagation technique, the backpropagation technique with momentum and the hybrid one which is a mixture between the backpropagation technique and the sliding variable structure. The robustness of this kind of training for neural observer is tested through a physical example. The obtained results show that the third type of training is better than using a classic kind of training especially concerning the rapidity of convergence.