Arc sensor model using multiple-regression analysis and a neural network

Abstract Experimental arc sensor models are developed by consideration of the welding conditions and characteristics of each welding process, and developing the model is significant because of its applicability to various welding environments. In this study, different types of regression model were developed for the current area difference method, the current integration difference method, and the weaving end current difference method, which are commonly used for the arc sensor. The characteristics of each regression model were examined, and a multiple-regression model was subsequently suggested, integrating all the conventional model characteristics. The multiple-regression model used the welding current signal of each model as the regressor and the offset distance as the response variable. In addition, an artificial neural network employing the current variable of each model as the input variable and the offset distance as the output variable was suggested as a new arc sensor model. A seam tracking simulation with a fuzzy controller implemented was constituted to facilitate the optimization of the scaling factor, and the scaling factor minimizing the tracking error was determined through the grid search method. Conventional models and the models suggested in this study were consequently compared with each other through the seam tracking experiment using the optimized scaling factors.