Melt Pool Size Prediction of Laser Powder Bed Fusion by Process and Image Feature Fusion

Real-time monitoring and control of the melt pool size during the laser powder bed fusion (L-PBF) can potentially improve the forming quality of the parts. Most existing studies predict the size based on process features, but the same building conditions may lead to different melt pool evolutions due to the inherent randomness of the L-PBF process. A novel prediction model based on process and image feature fusion is proposed in this article. First, process features that reflect the complex characteristics of the scanning process are extracted according to the process parameters and scanning strategy. Subsequently, the melt pool sizes are determined by the methods of three-scale threshold and least-square fitting. Finally, process features and melt pool features from previous scanning time periods are integrated by inputting them into recurrent neural networks (RNNs) in scanning order. The testing results indicate that the approach could better capture both the overall change trend and the inherent randomness of the melt pool. In addition, the gated recurrent unit (GRU) with a forgetting mechanism and fewer training parameters has better prediction performance compared with other typical RNNs, and the mean absolute percentage error (MAPE) of the melt pool area is 14.8%.