Multi-objective optimization of MAG process parameters based on ensemble models

Three-factor and three-level full factor bead-on plate and butting welding tests were carried out with welding voltage, welding speed and wire feed rate being taken as considerable process parameters. Based on the data of the tests, error back propagation neural network, radial basis function neural network and kriging model were established to predict weld reinforcement, the tensile strength and impact energy of the weld generated in the tests. The results of model prediction show that the models can predict the weld performance well, but there are no models that can predict the three kinds of weld performance at the same time, and the prediction of each model fluctuates greatly. In order to further improve the precision and stability of prediction, the three stand-alone models mentioned above were combined into ensemble models in the manner of linear weighting method. Then, the multi-objective optimization of process parameters was achieved by NSGA-II based on the ensemble models. Finally, the non-inferior solutions between the weld reinforcement, the tensile strength and impact energy of the joints are obtained and verified. It is of great significance to realize the overall optimization of weld comprehensive performance and the fine application of welding.