A major problem in additive manufacturing (AM) processes such as selective laser sintering (SLS) are part defects and irregularities, which have a negative impact on part quality. Currently, destructive and non-destructive testing methods are mostly used after manufacturing to detect such defects. However, in order to enable non-destructive, automated quality control and defect detection at an early stage, computer-aided machine learning (ML) algorithms are increasingly being used. Convolutional neural networks (CNN) based on ML methods are often used for this task and trained with appropriate data for the specific problem.
In the contribution, we present complex CNN architectures trained with specific ML methods such as transfer learning (TL), images from the SLS process, and an image dataset from a completely different feature domain to enable effective, non-destructive, and automatic classification of process images during the SLS process. In this context, a distinction is made between powder bed images where defects can be seen and those without defects. The proposed architectures use established ML models, which were initialized with pre-trained weights from the popular ImageNet dataset. In addition, an adapted classifier was developed for the specific problem of classifying image data with and without visible defects of the SLS process. To evaluate and compare the effectiveness of the presented methods, performance metrics for the ML models were obtained. These metrics ultimately demonstrate the effectiveness of defect detection based on CNN and may provide an alternative method for non-destructive quality control and manufacturing documentation for additively manufactured parts as well as other use cases based on image data.