Neural networks approach to electric machine on-line diagnostics

A neural network approach to the diagnosis of induction machine rotor faults is presented. The neural network substitutes the diagnostic indexes obtained through a complete model of the investigated machine. This model requires a large number of simplifying assumptions in conjunction with many design parameters not easy obtainable. In this work, the machine model is used to the aim of evidencing the input and the output patterns to train the neural network. The attention is focused on the conditioning of machine variables and parameters in order to select the most suitable inputs and outputs to realize a diagnostic system applicable to machine of a definite power range. The training process convergence and the evolution of the network weights put in evidence the effectiveness of the chosen inputs and outputs and the feasibility of this diagnostic procedure