The experimental model of arc sensor in $CO_2$ arc welding has been individually developed according to welding condition and welding procedure. Therefore, the development of new arc sensor having the features of all conventional arc sensor is important in point of its application to various welding environment. In this study, the arc sensor experimental models comprised of a regression model and noise term were formulated using conventional arc sensing algorithm such as current area difference, current integration difference and weaving end current difference method, and their features were observed. The new regression arc sensor model was suggested using multiple linear regression analysis using current variables as independent variables of regression analysis. The artificial neural network model was also suggested where current variables and offset distance was used input/out variables of input/output node.