Layer-wise multi-defect detection for laser powder bed fusion using deep learning algorithm with visual explanation

Abstract

In Laser Powder Bed Fusion (LPBF), it is a major challenge to obtain detailed spatial information on different powder bed defects in real-time and simultaneously. Deep Learning (DL) algorithms under the field of Machine Learning (ML) have promoted the intelligent development of the powder bed defect detection method. However, they still need to be further evaluated in terms of detection accuracy and time delay, training data overhead, and model robustness under complex environments. Also, the DL model usually treated as a black-box demands further explanation. Herein, three advanced DL models are constructed using bounding boxes to locate multi-defects quickly and accurately for the powder spreading of LPBF. High detection accuracy is achieved with limited training samples through both data augmentation aiming at expanding image samples and model-based Transfer Learning (TL) used for transferring from the source domain to the target domain by reusing trained models. Further, the data augmentation method is employed to generate low-resolution images with interference to test the robustness of the detection model in harsh environments. Besides, visual feature maps and saliency maps were generated with the Detector Randomized Input Sampling for Explanation (D-RISE) method to help understand the validity of the defect detection process of the DL model. Overall, this work shows that the proposed multi-defect detection algorithms can provide comprehensive information on powder bed defects accurately and quickly.