Coaxial melt pool monitoring with pyrometer and camera for hybrid CNN-based bead geometry prediction in directed energy deposition

Abstract

During the blown powder directed energy deposition (DED) process, optimizing key parameters such as laser power, travel speed, and powder feed rate is crucial for maintaining process stability. However, these conditions often require real-time adjustments due to thermal accumulation and excessive cooling over prolonged operations. To achieve this, accurately predicting bead geometry through real-time monitoring is essential. This study presents a coaxial melt pool monitoring approach that integrates a two-color pyrometer and a CMOS vision camera on the deposition head, enabling the simultaneous acquisition of temperature and image data. This configuration provides a comprehensive understanding of melt pool dynamics, improving predictive performance in bead geometry estimation. Given that precise bead geometry prediction (i.e., width, height, and depth) is critical for ensuring deposition quality and final component performance, we propose a hybrid CNN regression model that combines 1D CNN-based temporal analysis with 2D CNN-based spatial feature extraction. The proposed model outperforms both unimodal CNNs and conventional regression models, achieving high R2 values of 0.988, 0.970, and 0.978 for bead width, height, and depth, respectively, with notably low RMSE values. This multi-modal data-driven hybrid model demonstrates strong potential for advancing real-time melt pool monitoring in DED, contributing to improved process stability and part quality.