Application of multi-layered fuzzy inference based on backpropagation method to the robotic welding

This paper deals with the problem concerning the sensing of weld pool phenomena in robotic welding. In order to obtain a good quality of the welding result, it is important to control the penetration depth of the weld pool in robotic welding regardless of the disturbance such as variation of the gap and so on. It is difficult to directly measure the penetration depth. Moreover it may be difficult to describe the state equations for the penetration depth, since welding phenomena are described by partial differential equations. In order to estimate the penetration depth, a new knowledge based method is proposed, i.e. the depth is estimated from information such as welding current, the surface shape of the weld pool, and gap. The performance of the fuzzy inference depends on the fuzzy variables. The authors propose a new method to tune up the fuzzy variables. The method is based on the backpropagation method used to train feedforward neural networks. The validity of the fuzzy estimator is verified by carrying out the welding experiments.