TY - JOUR
T1 - Defect detection in fused deposition modelling using lightweight convolutional neural networks
AU - Kuriachen, Basil
AU - Jeyaraj, Rathinaraja
AU - Raphael, Deepak
AU - Ashok, P.
AU - Sundari, P. Shanmuga
AU - Paul, Anand
N1 - Publisher Copyright:
© 2024
PY - 2025/2/1
Y1 - 2025/2/1
N2 - Early detection of defects in additive manufacturing (AM) processes has significant benefits such as reducing material waste and improving production time, which results in an increase in overall productivity. Existing methods for defect detection in AM lacks in generalizability, takes more time for training, and finds difficulty in identifying complex defects in the vertical plane in real-time. In addition, there are no datasets freely available for active research. Considering this, in this article, we propose an algorithm that uses simple convolutional neural networks (CNNs), which is an artificial intelligence (AI) technique, to detect major defects in layers during fused deposition modelling (FDM) in real-time. It analyses AM infill patterns to identify irregularities such as staircase, overfill, and void defects. The proposed model is trained using the dataset that has been collected manually and augmented at different scales for building a robust model. The results show that the proposed model is very effective and provides over 97.77% accuracy on real-time images. Furthermore, our proposed model uses fewer convolution layers than popular models, such as visual geometry group (VGG) 19, mobile neural network (MobileNet) V2, residual network (ResNet) 50, and densely connected convolutional network (DenseNet) 121. In addition, we open source the custom-generated datasets that contain staircase, overfill, and void defects images. For future research, we plan to expand dataset diversity and employ real-time adaptive learning.
AB - Early detection of defects in additive manufacturing (AM) processes has significant benefits such as reducing material waste and improving production time, which results in an increase in overall productivity. Existing methods for defect detection in AM lacks in generalizability, takes more time for training, and finds difficulty in identifying complex defects in the vertical plane in real-time. In addition, there are no datasets freely available for active research. Considering this, in this article, we propose an algorithm that uses simple convolutional neural networks (CNNs), which is an artificial intelligence (AI) technique, to detect major defects in layers during fused deposition modelling (FDM) in real-time. It analyses AM infill patterns to identify irregularities such as staircase, overfill, and void defects. The proposed model is trained using the dataset that has been collected manually and augmented at different scales for building a robust model. The results show that the proposed model is very effective and provides over 97.77% accuracy on real-time images. Furthermore, our proposed model uses fewer convolution layers than popular models, such as visual geometry group (VGG) 19, mobile neural network (MobileNet) V2, residual network (ResNet) 50, and densely connected convolutional network (DenseNet) 121. In addition, we open source the custom-generated datasets that contain staircase, overfill, and void defects images. For future research, we plan to expand dataset diversity and employ real-time adaptive learning.
KW - Additive manufacturing
KW - Application of artificial intelligence
KW - Convolutional neural networks
KW - Defect detection
KW - Fused deposition modelling
UR - http://www.scopus.com/inward/record.url?scp=85212397691&partnerID=8YFLogxK
U2 - 10.1016/j.engappai.2024.109802
DO - 10.1016/j.engappai.2024.109802
M3 - Article
AN - SCOPUS:85212397691
SN - 0952-1976
VL - 141
JO - Engineering Applications of Artificial Intelligence
JF - Engineering Applications of Artificial Intelligence
M1 - 109802
ER -