Optimization of neural network parameters using Taguchi Robust Design: Application in plasma arc cutting process

Abstract

In this study, an experimental investigation of plasma arc cutting process was performed, according to the full factorial design of experiments. The examined process parameters were the cutting speed, torch standoff distance and arc voltage. Cut surface quality was identified by measuring the surface roughness and dimensional accuracy of the machined specimens. The obtained experimental data were used to train a feed forward back propagation NN in order to predict the quality indicators of the plasma cutting process. Training and architecture parameters of artificial neural network were optimized by the implementation of Taguchi method. Nine different NN were developed and tested according to the L9 (3^3) orthogonal array. Finally, by utilizing the analysis of means and the analysis of variances, the optimum levels of NN parameters were determined, and as a consequence, improved prediction ability was achieved.