Abstract: This paper presented a Bayesian network-based method to implement autonomous decision making of welding positions with T-joints in gas metal arc multipass welding. First, the vision sensing system was used to detect the feature points of the weld seam profile, and the current filling states and the requirements of the weld leg were respectively transformed into visual feature representations, with which the welding state was defined as backing, filling and cosmetic welding. Second, a Bayesian network was built using the identified feature points and the determined welding state, and prior probability used in this network was designed in a computational way based on welding experience. Finally, the decision-making result was implemented using the adaptive importance sampling algorithm based on the maximum posteriori probability criterion. Experimental results showed that the proposed Bayesian network here possessed better decision stability of the welding position during the multipass welding process with T-joints of different thicknesses compared with the analytic hierarchy process. The proposed Bayesian network method can overcome the effect of fake possible feature points on the decision-making result of welding positions. The decision-making results by the proposed Bayesian network are consistent with those by welding experience. The effectiveness of the decision-making results is up to about 95%. This method contributes to high welding efficiency and provides a reference for improving the automatic welding level in thick plate gas metal arc welding.Highlights: (1)The vision sensing system was used to detect the feature points of the weld seam profile, and the current filling states and the requirements of the weld leg were respectively transformed into visual feature representations, with which the welding state was defined as backing, filling and cosmetic welding.(2)A Bayesian network was built using the identified feature points and the determined welding state.(3)Prior probability used in this network was designed in a computational way based on welding experience.