Neural network-based assessment of the stress concentration factor in a T-welded joint

Abstract

The capability of an artificial neural network-based model in calculation of the stress concentration factor in non-load carrying T-welded joints was investigated in this study. A number of numerical models using the finite element method were assessed in order to evaluate the effect of changes in different geometrical parameters on the stress concentration factor of the configuration under investigation. The joint was analyzed in as-welded and TIG-dressed conditions for three cases of membrane, bending and membrane-bending loading, which are typical loading cases experienced by this type of joint. Taguchi methods were used for the design of experiments, leading to >320 models with different variable values. An artificial neural network technique was utilized to evaluate data from the finite element models and establish a relationship between the affecting variables in each condition and loading case. Optimizations were then performed using a genetic algorithm in order to establish the best combination of variables leading to the optimized stress concentration factor at each joint condition and loading case. Training and validation of the neural network-based model enables prediction of situations that have not been modeled. Prediction of stress concentration factors by the proposed model yielded perfect agreement with finite element results even for configurations in which the local weld parameters were outside the ranges for which the network was trained. The use of empirical equations for stress concentration factor calculation, however, gave clearly erroneous results.