Abstract: To help an automatic welding machine on reasoning dynamic welding process, a Kalman Filter Gaussian Process Regression (KF-GPR) model was proposed, and its theoretical basis was annualized. A prediction model was established later. Compared to conventional statistic method, the KF-GRP method can better estimate the distributed form and parameters for a dynamic welding process, which had higher robustness and fault tolerance. TIG welding experiment of the 304 stainless steel was carried out to verify the method. Totally 8 423 pairs of experiment data were collected and used for the model. The modeling results showed the proposed KF-GPR can suppress noises and provide fast and accurate model, which is essential for future online control experiment.