A neural network model is developed for on-line nugget size estimation in resistance spot welding. The variables used consist of features extracted from both controllable process input variables and on-line signals. A systematic signal and feature selection procedure is developed. The three commonly observed on-line signals, dynamic resistance, force, and electrode displacement, have been proven to carry similar information. Thus, only dynamic resistance is used in the model. The obtained model has been demonstrated to be robust over various welding conditions including electrode wear.