|About this Abstract
||MS&T21: Materials Science & Technology
||Additive Manufacturing of Metals: Equipment, Instrumentation and In-Situ Process Monitoring
||Melt Pool Level Flaw Detection in Laser Hot Wire Additive Manufacturing Using a Trained Convolutional Long Short Term Memory Autoencoder
||Brandon Abranovic, Sulagna Sarkar, Jack Lee Beuth
|On-Site Speaker (Planned)
This work focuses on deep learning enabled process monitoring for large-scale laser hot wire additive manufacturing using video data, which was collected using camera mounted on the robot arm pointed at the melt pool. Initial work consists of the unsupervised training of a convolutional long short term memory autoencoder to reconstruct footage from anomaly free single beads. The trained architecture was used to reconstruct footage where anomalies including arcing and wire stubbing occurred. Wire stubbing is a condition where un-melted wire impacts the solid bottom of the melt pool, leading to jittering of the wire. The model’s ability to faithfully reconstruct the video was quantified by computing a regularity score between the raw frames and model outputs, with low regularity scores being indicative of an anomaly. Preliminary results have demonstrated the model’s robustness in detecting the anomaly classes of interest to this study.