Anomaly detection in directed energy deposition: a comparative study of supervised and unsupervised machine learning algorithms

Directed Energy Deposition (DED) is a promising additive manufacturing technology increasingly utilized in critical industries such as aerospace and biomedical engineering for fabricating complex metal components. However, ensuring the structural integrity of DED-fabricated parts remains a significant challenge due to the emergence of in-process defects. To address this, we propose a comprehensive anomaly detection framework that leverages in-situ thermal imaging of the melt pool for defect identification. Our approach encompasses both supervised and unsupervised machine learning techniques to capture diverse defect patterns and accommodate varying levels of labeled data availability. Supervised methods—including ensemble classifiers and deep neural networks—are employed to learn from annotated thermal data, while unsupervised methods, such as autoencoders and clustering algorithms, are used to detect anomalies in unlabeled scenarios and uncover previously unknown defect patterns. The pipeline incorporates essential preprocessing techniques—such as feature extraction, normalization, and class rebalancing—to enhance model robustness. Experimental evaluations offer a detailed comparison between the supervised classifiers and unsupervised models utilized in this work, emphasizing the predictive performance and practical applicability of each learning paradigm. Notably, the supervised classification-based framework achieved high performance in detecting porosity-related anomalies, with an F1 score of up to 0.88 and accuracy reaching 99%.