Study on two-stage detection of keyhole tungsten inert gas welding penetration states

Keyhole tungsten inert gas welding has been widely applied in medium and thick plate connections due to its high efficiency. This study proposes a two-stage model, PSP-SquNet, for accurate classification of the weld penetration state. Knowledge distillation is applied to enhance the performance of the segmentation network, which is subsequently used to extract a coloured mask of the weld pool. After that, the mask is fed into SqueezeNet, which recognises four different stages of penetration: burn-through, complete penetration, partial penetration, and lack of penetration. This approach shows good stability and performance in real-time applications, achieving a detection accuracy of 98.49% while keeping the inference speed at 80 FPS.