科研之友
微信
新浪微博
Facebook
分享链接
摘要
In the process of selective laser melting, spatter and melt pool contain important information which can reflect the processing quality. It is one of the research emphases in recent years to obtain this information from the melt pool images, which are collected in the processing process, and then realize the process monitoring of selective laser melting. In order to extract the information of melt pool and spatter more accurately and effectively, a target detection model based on YOLOv5 is proposed to realize the real-time location and capture of the spatter and melt pool from the processing image. Firstly, based on YOLOv5s target detection network, the depth and width of backbone network are adjusted, and the number of detection heads is modified. After that, Self-calibrated convolutions and CBAM attention module are introduced to design a new feature integration structure. Through the above steps, the detection performance of the network is improved. The images collected by industrial camera are made into target detection datasets for model’s training and testing. The results show that the network can accurately locate the spatter and melt pool targets from the original image. With better detection accuracy, the network model has few parameters, which is more in line with the needs of industrial applications. The detection accuracy of mAP@0.5:0.95 reaches 0.466, thus provides a new method for the monitoring of selective laser melting process based on images. ? 2023 Editorial Office of Chinese Journal of Mechanical Engineering.
- 单位
