Abstract: There is a certain correspondence between the frontal molten pool and the penetration state. However, due to the numerous potential weld pool features related to the penetration state, how to remove redundant features and establish a concise knowledge model that reflects the corresponding relationship between weld pool features and penetration state is of great significance for achieving online control of welding penetration. A rough-fuzzy control method for welding penetration state is proposed and verified in variable gap MAG welding experiments. To solve the minimal feature set of the melt pool that characterizes the penetration state, we provide two attribute reduction algorithms based on variable precision rough sets. A decision information system for penetration state is established through variable gap-current welding experiments, and two shape parameters are introduced to describe the degree of sharpness at the tail of the molten pool. The classification rules of penetration status are obtained using rough set knowledge reduction and rule extraction algorithms. We establish a fuzzy control model for the width coefficient of the molten pool tail, and use the minimization of fuzzy entropy to construct membership functions of the error domain. The proposed control model is validated through two sets of variable gap welding experiments, and the results show that under closed-loop control, the weld back width is uniform and consistent, which can meet the requirements of welding specifications.