Detecting and classifying hidden defects in additively manufactured parts using deep learning and X-ray computed tomography

Automated methods for defect detection are a major goal of intelligent manufacturing systems, and additively manufactured (AM) parts presents unique challenges with complex internal features that are difficult to inspect. X-ray computed tomography (CT) is one of the only methods to inspect the interior of AM parts. This paper shows how deep machine learning (ML) models trained using computer-generated images of defects can automatically identify defects in CT images of real parts that were never previously seen by the model. To create an experimental dataset for testing, we designed a nozzle part having internal three-dimensional (3D) geometries and for some parts introduced intentional defects. Two different resin-based AM processes fabricated 227 parts, some of which were defect free and some of which included intentionally designed defects. CT scans were collected for each part which generated 100,334 cross section image slices that were labeled as defect free (86.4%) or having a defect (13.6%). To train a ML model for defect detection, we developed a novel method to create computer-generated images of defects from defect-free parts. More than 50,000 images of defective parts were generated and used to train a Vision Transformer (ViT) model. The model was tested on 572 defects in experimental parts. The defects that appear in the real parts for testing do not appear in the computer-generated training dataset. The model accurately detects and classifies defective parts with over 90% accuracy. The research demonstrates the potential of synthetic data to train deep learning models capable of detecting previously unseen defects. Such methods could be generalized to many types of part designs and defect types while greatly reducing the time and cost of training ML models for defect detection.