Multi-ANN approach for forecasting joint hardness and process variability in the friction stir welding process of AA2024-T3

Any process exhibits a certain intrinsic variability that cannot be completely eliminated. For this reason, it could be very useful to be able to estimate and, even more, to forecast this variability inherent to the process. This research aims to propose a prediction approach able to provide the exact value and its confidence interval. The investigation is based on the use of artificial neural networks (ANN), an essential tool in machine learning, by leveraging their ability to train slightly differently each time. For this reason, it is referred to as a multi-ANN approach. By exploiting their capability of generating variable predictions, it becomes possible to estimate the variability of the process. In other words, the authors developed a new prediction algorithm that employs multiple ANNs in parallel and integrates statistics to estimate confidence intervals. The developed approach is applied to the prediction of the joint hardness values and trend in the friction stir welding (FSW) process demonstrating good reproducibility of the predicted results and the confidence interval, thus verifying the ability of the approach to simulate the inherent variability of the processes. Specifically, the forecasted results exhibit an average percentage error of a mere 0.07% with minimum and maximum values of 0.01% and 3.13%, respectively.