| dc.description.abstract | Abstract. Monitoring melt pool behavior in laser powder bed fusion additive manufacturing is essential for ensuring process stability and detecting anomalies such as spatter, plume generation, and irregular melt pool shapes, all of which influence part integrity. However, conventional image-based deep learning approaches for this task, while accurate, are computationally intensive and difficult to deploy in real-time production environments. To address this challenge, this article presents a lightweight, feature-driven deep learning framework for multi-label defect classification. We develop a model that leverages a compact set of statistical, morphological, and texture features extracted from melt pool images, enabling concurrent classification of multiple defect types with minimal computational overhead. The dataset used in this study includes melt pool images from the Additive Manufacturing Metrology Testbed (AMMT) at the National Institute of Standards and Technology (NIST), providing both in situ monitoring data and ex situ characterization via high resolution X-ray computed tomography (XCT). Experimental benchmarking against a standard image-based model confirms the efficiency of our approach: it achieves F1 scores exceeding 98% across all categories while reducing model complexity by 16-fold. Furthermore, compared to conventional image-based pipelines, the proposed framework achieves a 2.3× speedup. Crucially, we validate these in situ classifications against ex situ XCT data, demonstrating that specific multi-label defect combinations correspond to measurable grayscale shifts as a proxy for internal porosity. This work thus offers a scalable, physically validated pathway for real-time quality management in additive manufacturing. | |
