Zhang, Shaoyu, Han, Quanquan, Zhu, Min, Zhang, Zhenhua, Wu, Defan, Zhao, Peng and Ji, Ze  ORCID: https://orcid.org/0000-0002-8968-9902
2026.
LPBF defect monitoring under limited annotation via adaptive multilayer feature fusion and layer-level defect indication.
Optics & Laser Technology
203
(D)
, 116348.
10.1016/j.optlastec.2026.116348
|
Abstract
Laser powder bed fusion (LPBF) requires reliable layer-wise defect recognition, but pixel-level annotation is costly and chamber images are often degraded by noise, blur and optical interference. This study presents a defect monitoring method trained with only 30 pixel-annotated images and a separate validation set of 30 annotated images. A fully frozen DINOv3-ViT backbone is combined with adaptive multilayer feature fusion, multiscale context enhancement and a lightweight decoder, so that transferable dense features can be used without full encoder fine-tuning. Segmentation outputs are converted into layer-level defect indicators and qualitative indication levels for layer screening. The method achieves an mIoU of 77.96% under this setting, outperforming convolutional and Transformer-related baselines. With strong synthetic noise and only 20% of the training data, the average mIoU remains 58.83%. In a preliminary cross-device and cross-material test on 15 annotated images from one additional machine–material combination, it reaches 72.43% mIoU without retraining. Training takes 37 min, with 5.8 GB average GPU memory and 0.23 s inference per image. Broader industrial applicability remains to be validated on larger multi-build, multi-material and multi-machine datasets.
| Item Type: | Article |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Engineering |
| Publisher: | Elsevier |
| ISSN: | 0030-3992 |
| Date of Acceptance: | 7 September 2026 |
| Last Modified: | 21 Sep 2026 16:00 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/189749 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions