Deep Learning-Enabled Infrared Monitoring for In-Process Anomaly Detection in Material Extrusions

Material extrusion is one of the most widely used additive manufacturing (AM) techniques, yet it faces significant challenges related to process stability. One of the primary concerns in this method is maintaining consistent extrusion quality, as underextrusion and overextrusion can critically impact the mechanical integrity, dimensional accuracy, and surface finish of printed components. These defects can lead to structural weaknesses, inter-layer delamination, material wastage, and extended production time, which reduce the overall reliability of the process. Additionally, temperature fluctuations, filament inconsistencies, and print speed variations further contribute to extrusion defects, making process optimization complex. To address these limitations, this study presents real-time defect detection framework that integrates a dual-infrared camera system with deep learning algorithms. This proposed system is designed to continuously monitor thermal characteristics of the extruded material, enabling early identification of extrusion anomalies. By capturing high-resolution thermal images from the front and one side of the in-print object, the system can precisely analyze temperature variations in the extruded filament, which directly correlates with extrusion stability. The collected dataset is then utilized to train and evaluate three deep learning models to classify the defects. These models incorporate Convolutional Neural Networks (CNNs), CNN with Long Short-Term Memory (CNN+LSTM), and CNN with Self-Attention architectures, each optimized for detecting extrusion defects. While CNN+LSTM provides an accuracy of 98.2% (front) and 98.4% (side), CNN + Self-Attention provides 99.3% accuracy when both cameras work combinedly. The integration of LSTM and Self-Attention mechanisms enhances the model’s ability to detect temporal variations, making it particularly effective in addressing gradual extrusion deviations over time.