With these features a set of input data was built for the neural network (input vector), making up sample totals of 14 undercut (UC), 15 lack of penetration (LP), 17 porosity (PO) and 49 slag inclusion, divided in 24 of nonlinear slag inclusion (NLSI) and 25 of linear slag inclusion (LSI). As the quantity of data was unequal between classes, some classes with less data were duplicated at random in order to obtain 25 samples, the largest quantity available for any one class, in order not to favor any one class during the training of the neural network. However, when considering the slag inclusion (SI) class as only one class, a total of 50 samples were used.