Exploring Feature Extraction Strategies for In-Situ Fault Detection of a Metal Additive Manufacturing Process with Time-Series Detection Models

Abstract

Metal additive manufacturing (AM), particularly through Laser Powder Bed Fusion (LPBF), enables the fabrication of complex geometries with applications in aerospace, automotive, and other high-performance industries. These components must often meet stringent requirements for durability, weight, and cost. However, the AM process remains susceptible to internal defects such as overmelting and spatter, which originate in the melt pool and adversely affect part quality. Real-time monitoring of the melt pool using imaging sensors has emerged as a critical approach for detecting such defects, though the high dimensionality of image data necessitates data compression to enable efficient, real-time anomaly detection. This work investigates and evaluates various image compression strategies for their effectiveness in supporting unsupervised, real-time anomaly detection within the LPBF process. Building upon a predictive model trained on a healthy baseline, the detection algorithm identifies faults by assessing deviations in incoming signals. The study reveals that certain compression methods may obscure these deviations, limiting detection sensitivity. Therefore, an appropriate compression technique is essential to maintain detection performance while reducing computational complexity. Both physically intuitive features such as melt pool dimensions and machine learning-based feature extraction methods are examined. The efficacy of each approach is evaluated across several defect types commonly encountered in metal AM builds. These results assess the anomaly distinguishability of each method, offering guidance for the development of robust in-situ fault detection for metal AM.