Abstract: Control system of weld quality in real time,which can ensure an excellent weld,is an important sub-system of welding flexible manufacturing cell (WFMC). A sub-system for weld quality control in real time and reliable communication between penetration control sub-system and the center computer of WFMC were investigated in this paper. Based on measurement of weld pools' characteristics parameters,artificial neural network models for topside and backside of welding pool were established. With the models,backside weld width could be deduced from topside parameters of welding pool in real time. A neuron self learning proportional summational differential (PSD) controller was designed,which could control backside weld width though sensing topside parameters of welding pool and adjusting welding peak current during robotic pulsed GTA welding. Controlling experiments verified that the controllers were effective.