Session: 02-02-02: Session #2: Measurement Science, Sensors, Non-destructive Evaluation (NDE) and Process Control for Advanced Manufacturing

Paper Number: 96692

96692 - Detecting Hidden Defects in Additively Manufactured Parts Using X-Ray Computed Tomography and Computer Vision 

Ensuring that a manufactured part is free from defects is among the most important steps in qualifying a part. Additive manufacturing (AM) can produce parts with complex three-dimensional (3D) internal features that are difficult or impossible to inspect using conventional methods like tactile measurements or photographs. These internal features can have defects or deformations that are not easily detected using conventional measurement methods. X-ray Computed tomography (CT) has emerged as a tool that can non-destructively inspect the internal features and identify flaws in AM parts. There is a need for computer vision methods that can automate CT inspection and recognize the many types of defects and deformations that are relevant for AM from CT scan data. Most publications on CT inspection focus on small numbers of parts, and there is also a need to create methods that are scalable and that are capable of automatically analyzing many parts.

In this study we apply convolutional neural networks (CNNs) to automatically detect internal hidden defects in parts from CT scans. We consider a nozzle part that has millimeter-scale internal 3D fluid channels, like those used in modern combustion systems and medical equipment. We produced a total of 155 polymer nozzle parts using a Carbon3D printer, some of which were defect free, and the remaining parts having one or more programmed defects. There was a total of five different types of defects including clogs within the internal channels, holes, and smaller degradations of the internal channel that may arise from problems with the part design, residual polymer, or error in machine instructions. The CT scans were collected on a Rigaku CTLab GX130, approximately four minutes per part, with resolution 90 mm. The 3D scans were sliced in the axial direction of the part, producing about 150 image slides per part. The slices were manually labeled as either defective or defect-free.

For each defect type we train a CNN based on ResNet34 using labeled slices of the CT scan data volume. When training the CNN with 80% of the available data, the classifiers can recognize internal defects on a slice with an accuracy greater than 90%. Defect free parts have no more than 10 slices classified as defective (false positives) and defective parts have no fewer than 10 slices as defective (true positives). By analyzing an entire part as a stack of classified slices, we can accurately classify all test parts as good or defective. We investigate the data efficiency and applicability of these deep learning models and achieve high classification accuracy with as little as 40% of the original training data. This work shows the potential for automatic detection of defects in AM parts using CT scans and computer vision methods.

Presenting Author: Miles V. Bimrose University of Illinois

Presenting Author Biography: Miles Bimrose received his B.S. degree in Mechanical Engineering from The University of Notre Dame in May 2020. His undergraduate research focused aerosol jet printing of functional materials. He joined Professor King’s group in the Fall of 2020, where he is focusing on developing smart automation and metrology for digital manufacturing.

Authors:

Miles V. Bimrose University of Illinois
Tianxiang Hu Zhejiang University
Davis J. Mcgregor University of Illinois
Jiongxin Wang Zhejiang University
Sameh Tawfick University of Illinois
Chenhuo Shao University of Illinois
Zhozhu Liu Zhejiang University
William P. King University of Illinois