Towards Robust Monitoring of the Laser Powder Bed Fusion Process based on Acoustic Emission combined with Machine Learning Solutions

Laser Powder Bed Fusion (LPBF) is an Additive Manufacturing (AM) process consolidating parts layer by layer, from a metallic powder bed. It allows no limitation in terms of geometry and is therefore of particular interest to various industries. Metallic LPBF samples can achieve near-full density and high resistance. However, to prevent defects from deteriorating the quality of the workpiece, process costly and time-consuming parameters optimization is required. The LPBF process lack of repeatability it limits its applications. The Ph.D. thesis proposes an alternative solution to trial-and-error optimization, based upon real-time acoustic in-situ monitoring combined with state-of-the-art machine learning. We first investigated the use of a low-cost microphone AM41 combined with machine learning (ML) algorithms. Three regimes (lack of fusion pores, conduction mode, and keyhole pores) and three alloys (316L stainless steel, bronze CuSn8, and Inconel 718) were selected. Three conventional ML algorithms and a Convolutional Neural Network (CNN) were chosen to perform the classification tasks. We proved that the acoustic emissions features are related to the laser-material interaction and don't originate from undesired machine or environmental noises. The regimes are classified with high accuracy (> 87%) regardless of the algorithms and materials. The AE features used for the classifications are material and regime dependent. Finally, a CNN multi-label architecture for classifying the material and the process regimes simultaneously reached a very high classification accuracy (93%), which is of great interest for multi-materials LPBF systems. We also introduce alternative AI methods to reduce the amount of data needed to train the algorithms, as well as to transfer the knowledge from one material to another. Saliency maps are used to determine the frequencies responsible for the ML algorithm classification. The analysis of saliency maps allows the quality of the trained model to be evaluated. A second microphone with a flat frequency response from 2 kHz to 200 kHz is compared with the AM41, whose response is restricted to bands around 10, 20, and 40 kHz. The information needed for the classification is scattered in the range 2 kHz-200 kHz, which allows both types of microphones to predict the regimes. Detailed investigation shows that most of the information is confined below 30 kHz, leading to an easier classification and a better model with the flat response microphone. The need to have a robust ML database is highlighted, with the flat response microphone. Twelve different laser parameter sets are chosen for each processing regime, to construct a database for the training of a CNN, applied to stainless steel 316L. We prove the possibility of generalization, i.e. using the model for classifying the regimes with high confidence (>96%), from AE signals recorded with unseen laser parameters. The influence of the "distance" in terms of power, speed, and