Prediction of welding residual stresses using Artificial Neural Network (ANN)

Abstract

Weld locations are significant sources of failure in components. Therefore, evaluating the mechanical properties such as residual stresses and eigenstates in these components is essential. This report describes the procedure to calculate residual stresses developed in welded components using Artificial Neural Networking (ANN). Artificial Neural networks are highly efficient and get us the result in the least possible time. A data-driven model was developed from using ANN; this made the procedure of calculating Residual stress in weld locations more efficient and less time-consuming. The predicted result was compared with the dataset and accuracy; the mean square error (mse) was calculated. This model was then made parametric so that it could be used to predict different mechanical properties for different databases. Hence a general regression model was developed. The model was also tested with dynamic layers and nodes to arrive at an optimum number giving higher accurate results; hence a more precise picture is drawn concerning how ANN can be optimally used in calculating material properties.