Application of Artificial Neural Network Algorithm to Predict Percentage Dilution of Gas Tungsten Arc Weld

Abstract

In this study a predictive, model was developed for percentage

dilution of mild steel welds using the artificial neural network expert

system. Predicting responses beyond experimentation boundaries is

a disadvantage to some other expert systems like the response

surface methodology. The same percentage dilution data collected

from the central composite experimental design was used for the

ANN model. The data was normalized, trained and tested. The neural

network architecture comprises, three (3) inputs, which is the

current voltage and gas flow rate and one output which is percentage

dilution, ten (10) neurons in the hidden layers and two (2) neurons

in the output layer. Lavenberg-Marquardt algorithm was used for

the data training. A performance evaluation plot showed that both

the test data set and the validation data set have similar

characteristics. The predicted values showed high correlation to the

observed data.