Please use this identifier to cite or link to this item: https://open.ni.ac.rs/handle/123456789/1687
Title: Assessing the sensitivity of the artificial neural network to experimental noise: A case study
Authors: Madić, Miloš 
Marinković V.
Issue Date: 1-Jan-2010
Journal: FME Transactions
Abstract: © Faculty of Mechanical Engineering, Belgrade. This case study deals with modeling of plasma cutting process using artificial neural network (ANN), with the aim of simulation the impact of noise on its performance. The input parameters used in this study were reduced to three cutting parameters which consisted of strength of current (I), cutting speed (V), and material thickness (s). The ten-point height of irregularities (Rz), which is one of the basic characteristics of the surface quality, was adopted as the output parameter (response). The data for this research were gathered from literature. A feed-forward three-layer ANN was created, with backpropagation and algorithm for supervised learning. For the hidden layer neurons sigmoidal type of non-linearity was selected, while a linear activation function was selected for the output layer. ANN training was carried out using Levenberg-Marquardt algorithm with Bayesian regularization. The trained and tested ANN on the original data set showed a satisfactory level of prediction accuracy. In order to simulate an experiment with noise, measured values of the surface roughness were corrected. The correction was performed by adding randomly selected numbers to each measured value, within the range - 0.1 to + 0.1 μm. With the previously selected architecture and the same other parameters, retraining of ANN was carried out. The analysis showed that the ANN model trained on the data with noise has similar performance as the ANN model trained on the original data, which indicates the robustness of this type of ANN.
URI: https://open.ni.ac.rs/handle/123456789/1687
ISSN: 14512092
Appears in Collections:Naučne i umetničke publikacije

Show full item record

SCOPUSTM   
Citations

6
checked on Sep 5, 2020

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.