Comparison of Early Stopping Neural Network and Random Forest for In-Situ Quality Prediction in Laser Based Additive Manufacturing
Matthew Behnke, Shenghan Guo, Weihong Guo
Abstract
Laser-Based Additive Manufacturing (LBAM) is a promising process in manufacturing that allows for capabilities in producing complex parts with multiple functionalities for a large array of engineering applications. Melt pool is a well-known characteristic of the LBAM process. Porosity defects, which have hampered the expansive adoption of LBAM, is correlated with the melt pool characteristic that occurs throughout the LBAM process. High-speed monitors that can capture the LBAM process have created the possibility for in-situ monitoring for defects and abnormalities. This paper focuses on augmenting knowledge of the relation between the LBAM process and porosity and providing models that could efficiently, accurately, and consistently predict defects and anomalies in-situ for the LBAM process. Two models are presented in this paper, Random Forest Classifier and Early Stopping Neural Network, which are used to classify pyrometer images and categorize if those images will result in defects. Both methods can achieve over 99% accuracy in an efficient manner, which would create an in-situ method for quality prediction in the LBAM process.