Comparison of machine learning methods for automatic classification of porosities in powder-based additive manufactured metal parts

An outstanding problem of additive manufacturing is the variability in part quality caused by process-induced defects such as porosity. Image-based porosity detection represents a solution that can be easily implemented into existing systems at a low cost. However, current industry porosity detection software utilizes threshold-based methods which require user calibration and ideal lighting conditions, and thus cannot be fully automated. This paper investigates the application of machine learning methods and compares their ability to classify porosities from cross-section images of 3D printed metal parts. Fifty-one features are manually defined and automatically extracted from the images and the most relevant features among them are selected using feature reduction methods. Six machine learning algorithms that are commonly used for classification problems are trained with those features and used for the porosity classification. The decision tree, one of the six machine learning algorithms, yields 85% accuracy with a processing time of 0.5 s to classify porosities from 691 images. However, manual features may not adequately characterize porosity because they are dependent on user's experience and judgment. Alternatively, deep convolutional neural network (DCNN) that does not require user-defined features is used for the classification problem. The comparison results showed that a DCNN yields the highest accuracy of 95% with a processing time of 1.8 s to classify porosities from the same 691 images.