Effectiveness of saliency-based methods in optimization of NN structure for induction motor fault diagnosis

The choice of optimal topology of neural network (NN) is one of the most important factor for the success of any application. Generally the optimization of neural network (NN) has based on cross validation method which requires more learning and test procedures. This paper proposes the use of sophisticated methods, it is one of the pruning NN methods as: "Optimal Brain Damage" (OBD) and "Optimal Brain Surgeon" (OBS) methods. The aim of the use of these methods is to firstly to improve the performances of the fault location procedure in the stator winding of the induction motor by NN and secondly the optimization of the NN structure. These methods are based on the use of the second derivative result of the cost function in order to obtain a compromise between network complexity and error learning. In this paper, a comparison between IM fault diagnosis results with the classic method (cross validation) and with the OBD and OBS methods is presented. The results with these two last techniques show their ability to locate the faulty phase with the best test performance using the least neurones number avoiding the over fitting.