Towards AI-Based Anomaly Detection for 3D Printing with Polymers

The industrial use of Fused Deposition Modeling (FDM) polymers for additive manufacturing has expanded beyond prototyping, yet quality-assurance issues continue to limit its adoption. The mechanical reliability of parts is often compromised by defects such as warping, delamination, stringing, and small layer misalignments, many of which traditional inspection techniques fail to detect. This research combines a review of AI methods for defect detection with the development of a standardized multi-class dataset and benchmarking framework to address the challenge of reliable automated defect detection. The dataset contains 2,638 manually validated RGB images across nine defect categories, including both major and fine-grained defect types. It was created under controlled conditions using a Prusa MK4 printer and made available in a format suitable for contemporary deep learning systems. Using this dataset, ten state-of-the-art convolutional neural networks (CNNs) and transformer-based models were fine-tuned and evaluated under unified protocols. MobileNetV2 and other lightweight CNNs achieved the best results for detecting Spaghetti failures, while the Swin Transformer (Swin-T) reached the highest accuracy (87%) and outperformed other models on subtle defects such as Z-seam and layer shifting. The findings demonstrate that CNNs and transformers achieve their best performance when their architectural design principles are understood and appropriately applied. This thesis provides both a thorough literature synthesis and reproducible benchmarking within a single framework, offering a foundation for future studies on sensor fusion, explainable AI, edge deployment, and real-time adaptive control. The integration of these technological advancements paves the way for scalable and reliable in situ quality assurance in polymer 3D printing through intelligent systems.