About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
Comparative Analysis of Data Augmentation Strategies for Defect Classification in Fused Deposition Modelling Additive Manufacturing Using VGG-16 |
| Author(s) |
Tejaswini A Bhosale, Sudarshan Sanap, MAYUR S. SAWANT |
| On-Site Speaker (Planned) |
Sudarshan Sanap |
| Abstract Scope |
Defect categorization in fused deposition modeling (FDM) is essential for quality assurance. However, restricted and imbalanced datasets impede the performance of deep learning models like VGG-16. However, limited and imbalanced datasets hinder the performance of deep learning models such as VGG-16. This study evaluated the impact of three data augmentation techniques-geometric transformations, Generative Adversarial Networks (GANs), and Neural Architecture Search (NAS)-based augmentation-on the classification of common FDM defects: warping, contamination, cracks, porosity, peeling, and good deposition. Experimental results showed that GAN-based augmentation achieved the highest improvement in classification accuracy (+13.2%) but required significant computational resources and time. NAS-based augmentation offered a balanced trade-off (+8.9%) with moderate computational cost and faster data generation. Geometric transformations resulted in modest improvements (+3.8%) with minimal resource usage. These findings demonstrated that advanced data augmentation significantly enhanced the performance of CNN-based defect detection in FDM, supporting the development of more intelligent and scalable manufacturing systems. |