3D Printing Failure Detection: Comparative Analysis of Deep Learning Architectures and Class Imbalance Techniques

Session Number

CMPS(ai) 18

Advisor(s)

Andrew Alini, Asa Harbin,MIT Lincoln Laboratories

Discipline

Computer Science

Start Date

17-4-2025 10:30 AM

End Date

17-4-2025 10:45 AM

Abstract

This research aims to develop systems capable of real-time monitoring and alerting users of potential print failures before they result in material waste and loss of time. We present a novel approach towards early detection and classification of 3D printing failures using various machine learning techniques. Using the existing CAXTON dataset, which focused on optimizing 3D printing parameters, we synthesize a specialized dataset to detect signs of printing failures and trained deep learning models on this new dataset. We compare the effectiveness of various residual network and vision transformer architectures by evaluating both fully trained networks and regression on extracted features from frozen backbones. We also explore class imbalance techniques including oversampling and class weighting. Initial results demonstrate promising results, with a ResNet50 achieving an 89% accuracy. This research contributes to the growing field of applying intelligent systems to additive manufacturing, ensuring more efficient and reliable performance.

Share

COinS
 
Apr 17th, 10:30 AM Apr 17th, 10:45 AM

3D Printing Failure Detection: Comparative Analysis of Deep Learning Architectures and Class Imbalance Techniques

This research aims to develop systems capable of real-time monitoring and alerting users of potential print failures before they result in material waste and loss of time. We present a novel approach towards early detection and classification of 3D printing failures using various machine learning techniques. Using the existing CAXTON dataset, which focused on optimizing 3D printing parameters, we synthesize a specialized dataset to detect signs of printing failures and trained deep learning models on this new dataset. We compare the effectiveness of various residual network and vision transformer architectures by evaluating both fully trained networks and regression on extracted features from frozen backbones. We also explore class imbalance techniques including oversampling and class weighting. Initial results demonstrate promising results, with a ResNet50 achieving an 89% accuracy. This research contributes to the growing field of applying intelligent systems to additive manufacturing, ensuring more efficient and reliable performance.