In situ monitoring plasma arc additive manufacturing process with a fully convolutional network

Plasma arc additive manufacturing (PAM) is an additive manufacturing technology that has been widely used in the past and has practical applications in many fields. The non-destructive testing methodology of PAM workpiece quality monitoring requires high level of accuracy and real-time capability. The characteristics of the melt pool and plasma arc are the keys to characterizing the dynamic manufacturing process during the powder feed PAM process, allowing for process prediction and real-time feedback control. This paper describes a new image recognition system that uses a fully convolutional network (FCN) to acquire melt pool and plasma arc morphologies simultaneously. Its image segmentation performance is compared with four conventional methods and three artificial intelligence methods. Results show that the FCN method can extract melt pool and plasma arc quickly and accurately, even in a complex manufacturing environment. The image segmentation-based FCN’s accuracy is 95.1%, and the average processing time is only 84 ms, according to the results. The performance is far superior to the existing seven methods. The relationship between average captured areas (melt pool and plasma arc) and parameters of the plasma arc (current intensity and scanning speed) is then analyzed. Finally, the quality of products in terms of sample surface roughness is measured, and its relationship with the average areas of the melt pool and plasma arc is clarified.