An artificial neural network model for predicting joint performance in ultrasonic welding of composites

Abstract Ultrasonic welding is a robust and cost-effective joining method. It can be applied to joining of metals, polymers and composites. Several approaches have been used in the identification of the relationships between input parameters of the ultrasonic welding (e.g., welding energy, time) and the joint strength, including experimental methods and finite element based prediction. Both these two methods are accurate but require long lead time for model development. The empirically identified relationships may be very complex and the accuracy of prediction is sensitive to the variations of welding conditions. Hence, a more robust and intelligent method is needed for predicting the joint strength. In this paper, an artificial neural network (ANN) method is proposed for predicting the joint strength in ultrasonic welding of polymer composites based on experiments. The input parameters are welding energy, plunging speed, trigger force, surface condition, and annealing temperature. The output parameter is the maximum loading force in a lap shear test. A supervised algorithm of training the network is applied because the equations governing the process are very complex and the conditions of the welding are continuously changing. The trained network is used for determining the quantitative relationship between significant inputs and output parameters as well as for predicting the performance for other inputs which are not experimentally tested or modeled, but fit in the range of training. The ultrasonic welding ANN is a dynamic model which can be continuously trained with new cases, serving as a flexible and cost-effective tool for designing welding procedure.