Self-Supervised Bayesian representation learning of acoustic emissions from laser powder bed Fusion process for in-situ monitoring

[Display omitted] •This study addresses LPBF monitoring robustness amid diverse data distributions across process parameters.•To label complex datasets into discrete process dynamics, this study suggests an ML strategy using Acoustic Emission from the process zone.•The study suggests a self-supervised Bayesian Neural Network for LPBF process dynamics identification without ground-truth data.•The framework excels in classification, anomaly detection, and transfer learning, even with varying AE data distribution due to LPBF parameters.•The study shows improved ML model generalizability through the self-supervised learning method's accurate predictions in a different environment. This study presents a self-supervised Bayesian Neural Network (BNN) framework using air-borne Acoustic Emission (AE) to identify different Laser Powder Bed Fusion (LPBF) process regimes such as Lack of Fusion, conduction mode, and keyhole without ground-truth information. The proposed framework addresses the challenge of labelling datasets with semantic complexities into discrete process dynamics. This novel AE-based in-situ monitoring approach provides a promising alternative to quantify part density in LPBF process. The study demonstrates the effectiveness of a Bayesian encoder backbone for learning the manifold representations of LPBF regimes, which were visually separable in a lower-dimensional representation using t-distributed stochastic neighbour embedding. The generalized representations learned by the Bayesian backbone allowed traditional classifiers trained on smaller datasets to exhibit high classification accuracy. The feature map computed using pre-trained Bayesian encoder on other datasets was also effective in anomaly detection, achieving 92% accuracy with one-class Support Vector Machine. Additionally, the representation learned by the BNN facilitates transfer learning, where it can be fine-tuned for classification tasks on different process maps, which is also demonstrated in this work. Our proposed framework improves the generalization and robustness of the LPBF monitoring, particularly in the face of varying data distribution across multiple process parameter spaces.