Comparative Performance Analysis of CNN Models for Cracking Defect Detection in 3D-Printed Polymer Parts

Abstract Additive manufacturing (AM), or 3D printing, has transformed modern manufacturing by enabling the production of complex shapes with high precision. However, AM faces challenges, such as cracking defects, particularly in parts made from thermoplastics like acrylonitrile butadiene styrene (ABS) and polylactic acid (PLA), which are valued for their strength and durability. This study evaluates five popular pre-trained convolutional neural network (CNN) models, namely DenseNet121, MobileNetV2, ResNet50, VGG16, and XceptionNet, for detecting cracking defects in 3D-printed ABS and PLA parts. The models were assessed using key metrics such as accuracy, loss, F1-score, and receiver operating characteristic (ROC) plots to identify the most effective model for real-time defect detection. A delta 3D printer equipped with a Raspberry Pi camera was used to generate a dataset of 3,705 images by printing various geometries. A design of experiments (DoE) approach was employed to capture images at different stages of the printing process, ensuring a diverse dataset. These images were pre-processed to focus on areas of interest, minimizing the computation required for the CNN models. Transfer learning was used to train the models to classify two categories: Cracking and No_Cracking. The dataset was divided into training (64%), validation (16%), and testing (20%) subsets. XceptionNet achieved the highest accuracy at 99.32%, followed by DenseNet121 at 99.05% and VGG16 at 98.92%. MobileNetV2 recorded 98.65% accuracy, while ResNet50 had the lowest at 95.00%. XceptionNet’s superior performance makes it the best option for real-time cracking defect detection, enhancing production efficiency.