Systematic deep transfer learning method based on a small image dataset for spaghetti-shape defect monitoring of fused deposition modeling

Abstract

As defect detection in the three-dimensional printing process has been essential, especially for fused deposition modeling (FDM), deep transfer learning is a promising technique for monitoring defects owing to the lack of large process datasets and deep learning expertize. This study proposes a systematic method of deep transfer learning based on a small image dataset to detect part failures of FDM printers. First, a small FDM process image dataset, including the spaghetti-shape error as defective, was prepared, and images for training and validation were generated using effective image augmentation techniques to enhance the diversity of the original dataset by considering a variation in part shape and visual monitoring conditions. Then, four popular and representative ImageNet pre-trained CNN architectures were selected: (1) VGGNet, (2) GoogLeNet, (3) ResNet, and (4) EfficientNet, and a deep transfer learning matrix of fine-tuning strategies with a partition of the selected models into multiple training sections by considering their pooling layers was proposed to generate systematic training executions. Finally, the performance and computational requirements, including accuracy, F1-score, parameters, memory size, and computing time for the prediction, were evaluated experimentally. The evaluation results reveal that the proposed method is capable of training models a suitable performance systematically for detecting the spaghetti-shape error. Furthermore, ResNet50 yielded an overall high performance with more than 90 % accuracy in a computing time of 1 s on a small computing device. Moreover, Grad-CAM visualization and a demonstration with a low-cost FDM printer verify the feasibility of the best-trained model for actual field monitoring problems.