A machine learning method for defect detection and visualization in selective laser sintering based on convolutional neural networks

Part defects and irregularities that influence the part quality is an especially large problem in additive manufacturing (AM) processes such as selective laser sintering (SLS). Destructive and non-destructive testing procedures are currently mostly used for quality control and defect detection of AM parts after production. In this context, machine learning (ML) algorithms are increasingly being used to enable computer-aided defect detection through automatic classification of manufacturing data. Convolutional neural networks (CNN) based on ML methods are widely used for this task. In this paper, complex transfer learning (TL) methods are presented, which enable the automatic classification of powder bed defects in the SLS process using very small datasets. The proposed methods use the VGG16 and the Xception CNN model with pretrained weights from the ImageNet dataset as initialization and an adapted classifier to classify good and defective image data recorded during part manufacturing. Known performance metrics were determined to evaluate and compare the performance of the models. The VGG16 model architecture achieved the best results for Accuracy (0.958), Precision (0.939), Recall (0.980), F1-Score (0.959) and AUC value (0.982). These results show the effectiveness of defect detection based on CNN and can offer an alternative method for non-destructive quality assurance and manufacturing documentation for additively manufactured parts.