Digital 3D defect maps: Detecting localised porosity with high-speed melt pool imaging data in LPBF

Abstract

Adoption of metal additive manufacturing for critical applications is hindered by the costs of post-build quality inspection. In-process monitoring offers a promising alternative by enabling parallel construction of digital 3D defect maps for every component manufactured. In this work, we present a system to detect local regions of porosity, containing both keyhole and lack-of-fusion defects, in laser powder bed fusion parts. A coaxial high-speed melt pool imaging setup operating at $20kHz$ acquires feature-rich data, capturing images approximately every $37.5\mu m$ along scan tracks and records over 30 million melt pool images per hour of build time. Using these data, a gradient-boosted decision tree model is trained to classify porosity levels in localised $2mm$ voxel bins. The system achieves a state-of-the-art detection threshold of 0.11% porosity, defined by the standard non-destructive evaluation criterion of 90% probability of detection at 95% confidence. By training on datasets containing realistic, organically generated porosity and demonstrating the most accurate localised porosity detection yet reported, this work represents a significant advance towards practical, industrially relevant in-process defect detection for additive manufacturing.