Semi-supervised anomaly detection in directed energy deposition using thermal images

Official URL: https://dx.doi.org/10.5220/0013729700003982

Directed Energy Deposition (DED) is a crucial additive manufacturing process used in aerospace and health-care applications, among others. However, ensuring defect-free production remains a challenge due to the difficulty in detecting defect-related anomalies in real-time. In this study, we address the problem of defect detection in DED processes through thermal images of melt pools. As an anomaly detection problem, we adopt a semi-supervised approach based on One-Class Support Vector Machine (OCSVM) and Isolation Forest (iForest). We analyze the performance of these models across different feature sets. Additionally, this semi-supervised approach is compared against an unsupervised approach utilizing the same learning algorithms. The results indicate the superiority of the semi-supervised approach for both algorithms. Yet, iForest outperforms OCSVM with an accuracy of 95% and an F1-score of 0.88, demonstrating its robustness in distinguishing defective from non-defective instances. This work provides valuable insights into the applicability of semi-supervised machine learning techniques for real-time defect detection in DED processes. By leveraging thermal imaging data and feature-based anomaly detection models, our findings contribute to the development of efficient, non-destructive quality control mechanisms for additive manufacturing processes.