Performances of regression model and artificial neural network in monitoring welding quality based on power signal

Abstract

In this study, a systematic research was conducted to compare the performances of the regression model and artificial neural network in predicting the nugget diameter of spot-welded joints by monitoring the dynamic power signature. The TC2 titanium alloy with a thickness of 0.4mm was used as the welding material, and a high-frequency precision spot welder was used to join the titanium alloy sheets. The dynamic welding current curve was obtained using the Rogowski coil, while the voltage curve was detected via two leads clipped onto the upper and lower electrodes during the entire welding process. The variations in the welding power signal in the welding process were investigated, and the characteristics of the power signals for different welding currents and electrode forces were analyzed. The power signals of different types of welding joints varied significantly. Five characteristics were extracted from the power signal to describe the shape of the curve. The stepwise regression analysis and back propagation neural network were respectively used to classify the welding joints into three categories: bad welds, good welds, and welds with expulsion. The performances of the two established prediction models were compared, and their behavioral discrepancies were attributed to their own data-mapping capabilities.