Investigation of Deep Learning for Real-Time Melt Pool Classification in Additive Manufacturing

Abstract

Consistent melt pool geometry is an indicator of a stable laser powder bed fusion (L-PBF) additive manufacturing process. Melt pool size and shape reflect the impact of process parameters and scanning path on the interaction between the laser and the powder material, the phase change and the flow dynamics of the material during the process. Current L-PBF processes are operated based on predetermined toolpaths and processing parameters and consequently lack the ability to make reactions to unexpected melt pool changes. This paper investigated how melt pool can be characterized in real-time for feedback control. A deep learning-based melt pool classification method is developed to analyze melt pool size both fast and accurately. The classifier, based on a convolutional neural network, was trained with 2763 melt pool images captured from a laser melting powder fusion build using a serpentine scan strategy. The model is validated through 2926 new images collected from a different part in the same build using ‘island’ serpentine strategy with predictive accuracy of 91%. Compared to a traditional image analysis method, the processing time of the validation images is reduced by 90 %, from 9.72 s to 0.99 s, which gives the feedback control a reaction time window of 0.34 ms/image. Results show the feasibility of the proposed method for a real-time closed loop control of L-PBF process.