Unsupervised Machine Learning for Anomaly Detection in Electron Beam Powder Bed Fusion

Abstract

Electron Beam Melting is used in high-stakes industries, such as in the aerospace and medical sectors, thus requiring rigorous quality control processes. As some builds consist of over 7,000 images, manual analysis is not feasible. Current approaches to counter these issues rely on heuristic-based methods that require large annotated datasets. This study suggests an unsupervised machine learning approach for anomaly detection, trained only on nominal data, removing the need for labelled defect examples. This work conducts an ablation study exploring different input modalities across four neural network models: an autoencoder and three image-to-image models. Results show that the autoencoder trained on reconstructing Layer images, and the image-to-image model taking Rake images together with CAD Masks perform best by a notable margin, where the image-to-image model achieves an F1 score of 98% on a baseline build containing both nominal and anomalous layers. The anomaly score behaviour remains consistent across multiple build cycles, suggesting the approach is scalable beyond a single build. However, validation across additional geometries and process parameters remains an important direction for future work.