In-situ capture of melt pool signature in high-speed laser cladding using fully convolutional network

High-speed laser cladding is a promising additive manufacturing technology, and the quality of cladding forming is closely related to the forming state of the melt pool. The result of melt pool segmentation is the key to understand the dynamics of the melt pool, and the feature segmentation of the melt pool can provide an effective basis for the subsequent quality diagnosis and real-time feedback control. Due to the traditional image processing algorithm is difficult to adapt to the melt pool image uneven brightness and splash and other disturbing factors, the traditional image processing methods have limitations on the adaptability of the melt pool segmentation. In this paper, a novel method for detecting the features of high-speed laser cladding melt pool based on fully convolutional neural network of FCN-Resnet50 is introduced, and the corresponding melt pool image dataset is produced. The FCN-Resnet50 image segmentation performance is compared with traditional image segmentation methods and other semantic segmentation algorithms. The results show that the validation accuracy rate of FCN-Resnet50 is 96.7 %, which realizes a better image segmentation effect with high robustness. In addition, the relationship between the average melt pool area and melt pool width captured by the melt pool and the laser power, scanning speed and powder feeding speed is analyzed.