Optimisation of ultrasonically welded joints through machine learning

Abstract

The quality of joint achievable through ultrasonic welding is highly dependent on the process input parameters. In this study an artificial neural network (ANN) is combined with a genetic algorithm (GA) to develop a high-fidelity model for predicting the strength of ultrasonically welded joints. Initial weights of the ANN were optimized using the GA. The model was then trained by the Levenberg-Marquardt algorithm on 27 training experiments and validated on 10 experiments. The model demonstrated a high level of accuracy with a mean relative error of 6.79% on validation data and a correlation coefficient of 0.9827 for all 37 experiments.