Today’s data-driven production is increasingly supported by artificial intelligence (AI) with a particular emphasis on machine learning (ML) algorithms to enhance production efficiency. These algorithms excel at detecting patterns within large datasets, making them especially valuable for data-driven tasks like image-based quality control in modern production environments. In additive manufacturing (AM), convolutional neural networks (CNN) are frequently used as part of image-based quality control systems for detecting and classifying printing errors and defects during or after production. However, many of these systems tend to focus on a single defect type only, making it challenging to adapt to different defects, as this often requires significant efforts for manual data annotation as well as retraining of the underlying models. In this paper, a methodology is introduced which can be applied to several different types of defects without requiring extensive manual data annotation and resource-intense training processes, which are typical bottlenecks of today’s image classification systems. By employing a pre-trained CNN (ResNet-50) exclusively for feature extraction and using the k-nearest neighbors (kNN) algorithm for classification, extrusion errors, layershifts, stringing, and error-free prints are classified with satisfactory performance levels with f1 scores ranging from 0.757 to 0.990 and a weighted average of 0.812. This suggests that the proposed methodology could serve as a promising direction for future image-based defect detection systems in fused filament fabrication (FFF).
Bast et al. (Fri,) studied this question.