Abstract: To exploring the causes and rules of rotary arc spatter, a combination method of multi threshold and BP neural network based on mask was proposed to identify welding spatter in accordance with the welding images of rotary arc flat surfacing collected by high-speed camera. The multi threshold method was used to obtain the spatter position and contour, and then the spatter was identified by establishing a BP neural network model with five characteristic values The recognition accuracy of this combined method can reach 95.76% for rotating arc spatter images with complex background. At the same time, through the phase analysis of spatter and welding wire position, the average phase of the maximum number of spatters is 241.4°, that is, about 0.14 cycle position after the end of welding wire enters the molten pool. This is mainly due to the current surge caused by the contact between the droplet at the end of welding wire and the molten pool, and the insufficient current suppression, The research results provide a basis for controlling spatter in rotating arc welding.