Back propagation neutral network based modelling and optimization of thermal conductivity of mild steel welds Agglutinated by Tungsten Inert Gas welding technique

Abstract

In this study, Central Composite Design (CCD) was used to develop a Design of Experiment (DOE) for predicting thermal conductivity, allowing for the selection of weld input variables based on the ranges found in literature. Therefore, a design matrix having six (6) centre points, six (6) axial points and eight (8) factorial points, resulting in twenty (20) experimental runs was employed as TIG welding input variables, which included welding current ranging from 199.77-250.23 A, voltage ranging from 20.98-26.0 V and gas flow rate ranging from 11.98-16.0 L/min. Artificial Neural Network (ANN) algorithm was employed to compute the thermal conductivity generated DOE to optimum range. It was observed that thermal conductivity obtained from experimental and ANN prediction depended on the input process parameters which included welding current, voltage and gas flow rate. The regression plot showed R = 0.9919 as progress of training, R = 0.8982 as progress of validation and R = 0.9979 as progress of the training test. This led to overall correlation coefficient (R) of 0.8768 which signified that ANN is a robust tool for predicting the percentage of weld dilution. To test the reliability of the trained network, the ANN model was employed to predict its own value of percentage dilution using the same input parameters generated from the central composite design. Based on the observed and the predicted values of percentage dilution, a regression plot of outputs was thereafter generated, and R2 value of 0.9884. There was correlation in the results obtained, as both experimental and predicted weld thermal conductivity were 55.84 and 49.99 W/m.k minimum as well as 79.75 and 76.04 W/m.k maximum, indicating that ANN can be applied in actual welding scenario for prediction of thermal conductivity with minimal inacuracies.