Unsupervised Statistical Learning for In-Situ Anomaly Detection in LPBF | School of Engineering

In-situ optical spectroscopy during Laser Powder Bed Fusion (LPBF) captures high-resolution signals encoding the physical state of the melt pool. However, this data presents a formidable statistical challenge: observations are strongly non-i.i.d., exhibiting spatiotemporal autocorrelation, directional asymmetry from moving laser paths, and complex thermal dynamics. Because labeled defect data is rarely available at scale, this project bypasses supervised approaches to build principled, physics-informed unsupervised learning frameworks.

This project will build on the supervisors’ research in stochastic modelling, statistical learning, numerical analysis, and mathematically grounded anomaly detection. The research will focus on formulating flexible, label-free statistical machine learning for real-time anomaly detection and defect classification. Depending on candidate interest and project evolution, potential methodological avenues include combining spatial modelling with non-parametric hypothesis testing to separate systematic process variation from distributional shifts. To account for complex spatiotemporal dependencies, you will investigate robust calibration schemes (such as block-permutation or wild-bootstrap methods), while exploring techniques ranging from sparse matrix decomposition to geometric clustering and manifold learning to isolate and classify physical anomaly regimes like porosity, lack of fusion, or keyhole collapse.

Bridging spatial statistics, rough path theory, functional data analysis, and advanced process engineering, this project translates rigorous mathematical methodology into immediate industrial impact for metal additive manufacturing.