Hyperspectral In-Situ Monitoring for Deep Learning-Based Anomaly Classification in Metal Additive Manufacturing

Abstract

Metal Additive Manufacturing processes such as Directed Energy Deposition (DED) require process monitoring to ensure the highest part quality. Detecting and avoiding material defects to meet high material requirements remains a challenge due to the complexity of the process. To address this challenge, this study presents a novel approach that combines hyperspectral imaging with convolutional neural networks to classify process anomalies. Hyperspectral in-situ monitoring captures the light emitted from the melt pool over the 2 spatial axis, but also over the spectral axis. The resulting hypercube image contains a lot of information over the thermal state of the melt pool but is very high-dimensional, which is not a problem for Convolutional Neural Networks. The proposed classification model reaches an accuracy in excess of 94% over the validation set. The classification mechanism of the proposed model is investigated thanks to the Guided GradCAM visualization method and links with the melt pool temperature distribution are formulated. The inference speed of the optimized model is measured and shown to be compatible with real-time applications. This study is a stepping stone towards smart control of the DED process based on the identified thermal state of the melt pool, with the goal of improving the part quality.