Quality prediction and rivet/die selection for SPR joints with artificial neural network and genetic algorithm

Abstract

• Three artificial neural network (ANN) models were developed and proved capable of predicting the quality of SPR joints. • Genetic algorithm (GA) combined with ANN models can automatically generate optimal rivet/die combinations for new joints. • Application range maps of rivet/die combinations offer an effective methodology for selection of rivet/die for new joints. • Interaction effects between different joint parameters can be easily visualized by plotting contour graphs with ANN models. In this study, artificial neural network (ANN) was adopted to predict the quality of SPR joints. Three ANN models were developed respectively for the key joint quality indicators: the interlock, the remaining bottom sheet thickness at the joint center ( T cen ) and under the rivet tip ( T tip ). Experimental SPR tests were performed and the results verified the high prediction accuracy of the ANN models. The mean absolute errors (MAE) between the experimental and prediction results for the interlock, T cen and T tip reached 0.058mm, 0.075mm and 0.059mm respectively, and the corresponding mean absolute percentage errors (MAPE) were 14.2 %, 22.4 % and 10.9 %. Moreover, two innovative approaches were proposed to simplify the selection of rivet and die for new joint designs. One was realized by combining the genetic algorithm (GA) with the ANN models, and can generate optimal rivet and die combinations for different joint quality standards. The second was achieved by plotting application range maps of different rivet and die combinations with the help of ANN models, and can quickly select the suitable and accessible rivet and die. Furthermore, interaction effects between different joining parameters on the joint quality were also discussed by analyzing the contour graphs plotted with the ANN models.