<p>Additive manufacturing (AM), fundamentally different from traditional subtractive manufacturing techniques, is a layer-by-layer deposition process to fabricate parts with complex geometries. The formation of defects within AM components is a major concern for critical structural and cyclic loading applications. Understanding the mechanisms of defect formation and identifying the defects play an important role in improving the product lifecycle. Since manually identifying defects in AM parts is time-consuming and error-prone at the same time, the convolutional neural network (CNN) has been demonstrated to be an effective deep learning tool for automated detection of defects for both conventional and AM processes. But a network with optimized parameters including proper data processing and sampling can improve the performance of the architecture. In this study,&nbsp;for the detection of good deposition quality&nbsp;and defects such as lack of fusion, gas porosity, and cracks in a fusion-based AM process, a CNN architecture is presented comparing the&nbsp;classification report and evaluation of different architectural settings and obtaining the optimized result from them. With optical images, a data set of approximately 8500 images was prepared. For supervised learning of the network, these sample images were classified manually into four classes such as lack of fusion, gas porosity, cracks, and good quality. Total 80% and 20% of the images were used with on-the-fly data augmentation for training and validation and testing the network, respectively. Since data set preparation, visualization, and balancing are very important aspects in deep learning to improve the performance and accuracy of neural network architectures, exploratory data analysis was performed for data visualization, and the up-sampling method was implemented to balance the data set for each class. After the data set preparation and visualization, different architecture for the CNN network with the parameters used in the previous study was designed. Dropout and regularization techniques were used to overcome overfitting. By comparing the results for different architectures, the optimal CNN network was chosen for further investigation. To tune the hyperparameters and to achieve an optimized parameter set, a design of experiments was implemented to improve the performance of the network. By localizing the defect location and analyzing the misclassified images, attention maps were generated. The performance of the network with optimized parameters was compared with the results from the previous study. The overall accuracy (&gt;96%) for both training and testing the CNN network presented in this work transcends the current state of the art (92%) for AM defect detection.&nbsp;</p>

A Convolutional Neural Network (CNN) for Defect Detection of Additively Manufactured Parts

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.