Explainable Vision Transformer-Based Real-Time Anomaly Detection for Fused Deposition Modeling

Abstract

Fused Deposition Modeling (FDM) is one of the most widely used additive manufacturing (AM) techniques, especially in consumer-grade 3D printing. However, it suffers from a lack of real-time quality assessment and process control, leading to common print defects such as underprinting, overprinting, and inconsistencies in surface quality. This research proposes the integration of Vision Transformers (ViTs) to enhance real-time anomaly detection in the FDM process using 2D laser scan depth maps. A high-speed KEYENCE LJ-V7000 laser profiler scans the printed surface layer-by-layer, producing depth maps that are analyzed by the ViT model to identify surface defects. The model classifies the surface into four categories: underprinting, overprinting, normal, and empty regions. By utilizing the self-attention mechanism of Vision Transformers, the model captures subtle spatial dependencies and detects complex anomalies that traditional methods might miss. To improve interpretability, we incorporate explainability techniques like attention maps and Grad-CAM to visualize which parts of the depth map contribute to anomaly detection, offering transparency and actionable insights for operators. Our model achieves state-of-the-art (SOTA) performance in anomaly detection, outperforming traditional methods in accuracy and robustness. The proposed solution can improve defect detection accuracy and enable real-time process adjustments based on detected anomalies, enhancing the quality and reliability of FDM printing. This work demonstrates the potential of Vision Transformers for advancing data-driven quality control in additive manufacturing.