This study addresses intelligent techniques for fulfilling quality control of bead-on-plate welding. A new visual double-sided sensing system capable of imaging the weld pool topside and backside simultaneously in a frame was provided to determine the weld pool geometry parameters. The imaging principle was analyzed with spectrum distribution, in which the weld pool was illu minated by arc light emission to receive a clear image under base current. Double-sided size parameters describing weld pool geometry were defined anc determined in real time with the devel oped image processing algorithm. The influences of welding parameters such as pulse duty ratio and travel speed on weld bead geometry were identified by step response. Based on the analysis, a neural network model of the dynamic process was established for predicting the back side width with the welding parameters and topside size parameters. The simula tion results indicated the accuracy of the model, and the characteristics of the welding process were analyzed carefully, Aiming at the bead-on-plate pulsec GTAW process, conventional and intelli gent control methods of single input and single output were investigated, and the neuron self-learning PSD control was verified with better performance for practical application through comparisons.