117 / 2025-05-15 17:57:28

Melt Pool State Monitoring Method for Laser Wire Arc Additive Manufacturing Based on MKSE-MAML

Laser and wire additive manufacturing; Nickel-based high-temperature alloys; Melt pool; Cladding quality; Meta learning

Laser Wire Additive Manufacturing (LWAM) is a critical technology for the fabrication of high-precision and complex structural components, wherein the forming quality is directly governed by the state of the molten pool. However, conventional convolutional neural network (CNN) employed for molten pool recognition demand large-scale annotated datasets, and exhibit limited predictive accuracy when applied to cladding quality estimation across heterogeneous processes and multi-source signals. To overcome these limitations, the experimental molten pool data of a GH4169 nickel-based superalloy were categorized into seven distinct classes: dripping, stubbing, incomplete melting, oscillating, unmelted, normal melting, and overmelting. Coaxial and off-axis molten pool images acquired via Laser Directed Energy Deposition (LDED) were utilized as the training set, whereas coaxial molten pool images obtained from LWAM were designated as the testing set. An MKSE-MAML model was subsequently developed to perform molten pool state classification. To assess the effectiveness of the proposed method, its predictive performance was systematically compared with that of ProNet, Relation Network, MAML, and SE-MAML models. Experimental results revealed that the MKSE-MAML model achieved an accuracy of 99.48% in few-shot target domain evaluations, thereby demonstrating that multi-source data augmentation significantly enhances the model's generalization capability.