A defect detection framework using three-dimensional convolutional neural network (3D-CNN) with in-situ monitoring data in laser powder bed fusion process

Abstract

Laser powder bed fusion (L-PBF) is a promising additive manufacturing (AM) technology for manufacturing complex-shaped metallic parts with high density. Since L-PBF has been rapidly developed and applied in various industries, the quality assurance of the printed part using the process became a topic of primary importance. In this respect, in-situ monitoring techniques have received increased attention in recent years for aiding the quality control of L-PBF process and the certification of the products. This study proposes a novel defect detection framework with a three-dimensional convolutional neural network (3D-CNN) and in-situ monitoring of light intensity for detecting both the lack-of-fusion and keyhole-induced defects. The proposed 3D-CNN model works with a 3D moving window to perform the local inspection using the measured light intensities in three-dimensional space to classify the type of defect. Furthermore, the model predicts the local volume fraction to provide insights into the degree of defect. To perform the classification and regression with a single 3D-CNN, the joint classification and regression approach was adopted to train the model using the results obtained with micro-computed tomography as the ground truth. In order to build the training dataset, the samples with artificial defects were fabricated in different process regimes with energy densities ranging from 19.84J/mm3 to 110.12J/mm3. After the training process, the proposed model was evaluated with the test specimens which contain randomized defects generated due to the excessively low and high energy input. The results showed that the proposed 3D-CNN-based defect detection framework can detect pores greater than 80μm induced by both lack-of-fusion and keyhole mode melting. The sensitivity of the proposed framework was evaluated, showing the true positive rate of 75.69% and 70.47% for lack-of-fusion and keyhole defects with a void volume larger than 0.0003mm3. The prediction of local volume fraction with R2 score of 0.91 was also achieved with the proposed 3D-CNN approach.