A Deep Learning Approach to Defect Detection in Additive Manufacturing of Titanium Alloys

In Additive Manufacturing (AM) of titanium alloys, the formation of defects in parts is typically related to the stability of the melt pool. With increased instability and size of the melt pool comes an increase in the level of emissions generated as the laser processes the material. Recent developments with in-situ monitoring and process control allows the collection of large amounts of data during the printing process. This includes data about emissions, which are made available as 2D representations in the form of colour images. However, it is still a manual process to inspect these 2D representations to identify defects, which hinders scalability. Given recent advances in Deep Learning for computer vision and the availability of large amounts of data collected from in-situ monitoring, our approach leverages Deep Learning techniques for characterizing abnormal emissions to automatically identify defects during the printing process. One of the challenges to apply deep learning in AM is the lack of proper labelled data for training the models. In this paper, we tackle this challenge by proposing an approach that uses transfer learning and fine-tuning on a pre-trained Convolutional Neural Network (CNN) model called VGG 16 to successfully train the deep model with a small labelled dataset. Results show good classification accuracy on the emission images obtained from the in-situ monitoring system, and improvements in classification of defects on a public industrial benchmark datasets named DAGM (Deutsche Arbeitsgemeinschaft für Mustererkennung e.V., German chapter of the IAPR).