This study investigates the use of a low cost microphone combined with state-of-the-art machine learning (ML) algorithms as online process monitoring to differentiate various materials and process regimes of Laser-Powder Bed Fusion (LPBF). Three processing 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 resulting in five main findings. First, we proved that the AE features are related to the laser-material interaction and not from undesired machine or environmental noise. Second, the process regimes are classified with high accuracy (> 87%) regardless of the algorithms and materials. Third, it is possible to build a single model from the three materials and still reach high classification accuracy (>86%) of the different regimes. Forth, the AE features used for the classifications are material and regime dependent. Finally, with LPBF processing of multi-materials on the rise, a strategy for classifying the material and the process regimes simultaneously using a CNN multi-label architecture reached a very high classification accuracy (≈ 93%). The results demonstrate the potential of our approaches for online LPBF process monitoring of different materials and regimes.

Metal-based Laser Powder Bed Fusion (LPBF) suffers from a lack of repeatability and is challenging to model, making their quality monitoring essential and demanding. The reason lies in the high dynamics taking place during the interaction of the laser with metallic powders. To bring this technology to mass production, industries are only interested in the process regime where the built layer's quality meets their standards. All other process regimes leading to poor mechanical properties and/or defect formation such as balling, Lack of Fusion (LoF) pores, keyhole pores, delamination, and crack propagation irrespective of their different regimes are considered anomalies. Today, the common methodology for monitoring uses conventional/supervised Machine Learning (ML) algorithms for the classification task. However, it requires collecting a balanced dataset corresponding to each investigated regime from the sensors, which is very expensive and time-consuming. As an alternative, the article proposes a semi-supervised approach where the defect-free regime can be differentiated from the anomalies by familiarising the ML algorithms only with the distribution of acoustic signatures corresponding to the defect-free regime. This work presents two generative Convolutional Neural Network architectures based on Variational Auto-Encoder and General Adversarial Network. As a result, we could classify the anomaly regimes with 96 and 97% accuracy, respectively.