Welding deformation adversely affects the quality and precision of structural components, and traditional methods require significant material resources and time. Machine learning has demonstrated exceptional accuracy and efficiency in solving complex problems. Thus, the use of machine learning to predict welding deformations is a novel approach. In this study, laser welding experiments were conducted on a TC4 titanium alloy to establish a welding deformation dataset. The deep neural network (DNN) and convolutional neural network (CNN) models were designed and constructed, with average prediction errors of 0.85 mm and 0.94 mm on the validation set, respectively. To further optimize the network parameters, a differential evolution algorithm was employed through mutation, crossover, and selection. The results indicated that after optimization, the prediction errors of the DNN and CNN models reduced to 0.75 mm and 0.85 mm, respectively. These represent accuracy improvements of 14.8 % and 9.6 %, respectively. The optimized models exhibited superior predictive performances for the validation set.
Cheng et al. (Thu,) studied this question.