In-situ visual monitoring for multi-scale defects detection in laser powder bed fusion

Laser powder bed fusion (PBF-LB) is a promising metal additive manufacturing technology. However, ensuring consistent quality remains a pressing challenge due to defects that arise during the PBF-LB process. These defects significantly impede subsequent molding procedures, emphasizing the urgent need for real-time monitoring technology to steer part production effectively. Present monitoring techniques are limited, typically addressing specific defect types or contour deformations individually and lacking a simple and effective method to achieve multi-scale defects detection for PBF-LB parts. This study aims to pioneer a method that can detect defects and part contour deformations during the PBF-LB process using only an industrial camera as the monitoring device. The proposed method consists of two key components: a semantic segmentation model based on deep transfer learning (SS-DTF model) and a contour deformation detection module. To begin with, an off-axis in-situ monitoring system is deployed with an industrial camera to capture powder bed images for each layer after laser scanning and powder spreading. This enables the production of datasets for post-scanning and post-spreading anomalies. The SS-DTF model is trained on the above two datasets respectively to achieve segmentation for multi-category defects. The contour deformation detection module provides automatic early warning by comparing the true contour of the part with the nominal contour layer-by-layer, using statistical process monitoring method. The proposed method is then used to monitor parts with various geometries under real processing conditions, aiming to clarify the potential correlation between monitoring outcomes and the parts' multi-scale defects in PBF-LB. The results show that the proposed method provides a promising solution for insitu monitoring of the PBF-LB process.