TY - GEN
T1 - Application of K-Nearest Neighbors Algorithms for Void Classification in Composite Oriented Strand Board
AU - Hu, Wenyue
AU - Eftekhari, Yaser
AU - Callander, Sam
AU - Wang, Xiaoxing
AU - Bowland, Christopher C.
AU - Nguyen, Frank
AU - McCaslin, Jeremy
AU - Schaal, Christoph
AU - Gu, Grace X.
AU - Li, Carina
AU - Jin, Bo
N1 - Publisher Copyright:
Copyright © 2023. Used by CAMX - The Composites and Advanced Materials Expo with permission.
PY - 2023
Y1 - 2023
N2 - Composite Oriented Strand Board (COSB) is an aerospace-grade material that is manufactured using laminated, unidirectional carbon fiber-epoxy prepreg strands. Its manufacturing methodology allows fine-tuning, producing controllable thickness, flatness, and microstructure with quality assurance. To comprehend the microstructure of COSB from X-ray computed tomography (XCT) scan data, it is critical to accurately assess and quantify void content in the post-analysis. The conventional methods of post-analysis used for X-ray micro-computed tomography (micro-CT) have become inadequate due to their time-consuming nature and oftentimes imprecise measurements. In this paper, we present a new approach for enhancing void classification accuracy by utilizing the K-Nearest Neighbors Classifiers (KNN) with the assistance of convolutional kernels. KNN training based on two labels of greyscale thresholding images achieved a 1% error rate, a significant improvement over the three deep learning algorithms (Fully Convolutional Neural Network, U-net, SegNet) in our previous study. When classifying five different void labels, the multi-classifier KNN algorithm with convolutional kernels yields around 2% error rate. Encouraged by these results, we propose using convolutional neural networks (CNNs) to make more complex reasoning and decisions when classifying voids, improving the accuracy of void characterization.
AB - Composite Oriented Strand Board (COSB) is an aerospace-grade material that is manufactured using laminated, unidirectional carbon fiber-epoxy prepreg strands. Its manufacturing methodology allows fine-tuning, producing controllable thickness, flatness, and microstructure with quality assurance. To comprehend the microstructure of COSB from X-ray computed tomography (XCT) scan data, it is critical to accurately assess and quantify void content in the post-analysis. The conventional methods of post-analysis used for X-ray micro-computed tomography (micro-CT) have become inadequate due to their time-consuming nature and oftentimes imprecise measurements. In this paper, we present a new approach for enhancing void classification accuracy by utilizing the K-Nearest Neighbors Classifiers (KNN) with the assistance of convolutional kernels. KNN training based on two labels of greyscale thresholding images achieved a 1% error rate, a significant improvement over the three deep learning algorithms (Fully Convolutional Neural Network, U-net, SegNet) in our previous study. When classifying five different void labels, the multi-classifier KNN algorithm with convolutional kernels yields around 2% error rate. Encouraged by these results, we propose using convolutional neural networks (CNNs) to make more complex reasoning and decisions when classifying voids, improving the accuracy of void characterization.
UR - http://www.scopus.com/inward/record.url?scp=85188539504&partnerID=8YFLogxK
U2 - 10.33599/nasampe/c.23.0117
DO - 10.33599/nasampe/c.23.0117
M3 - Conference contribution
AN - SCOPUS:85188539504
T3 - Composites and Advanced Materials Expo, CAMX 2023
BT - Composites and Advanced Materials Expo, CAMX 2023
PB - The Composites and Advanced Materials Expo (CAMX)
T2 - 9th Annual Composites and Advanced Materials Expo, CAMX 2023
Y2 - 30 October 2023 through 2 November 2023
ER -