Neural Networks Applied for Predictive Parameters Analysis of the Refill Friction Stir Spot Welding Process of 6061-T6 Aluminum Alloy Plates

Abstract

Refill friction stir spot welding (RFSSW) technology is a solid-state joint that can replace conventional welding or riveting processes in aerospace applications. The quality of the new welding process is directly influenced by the welding parameters selected. A finite element analysis was performed to understand the complexity of the thermomechanical phenomena during this welding process, validated by controlled experiments. An optimization model using neural networks was developed based on 98 parameter sets resulting from changing 3 welding parameters, namely pin penetration depth, pin rotation speed, and retention time. Ten parameter sets were used to verify the learning results of the optimization model. The 10 results were drawn to correspond to a uniform distribution over the training domain, with the aim of avoiding areas that might have contained distortions. The maximum temperature and normal stress reached at the end of the welding process were considered output data.