Real-time and high-accuracy defect monitoring for 3D concrete printing using transformer networks

Abstract

Defects and anomalies during the 3D concrete printing (3DCP) process significantly affect final construction quality. This paper proposes a real-time, high-accuracy method for monitoring defects in the printing process using a transformer-based detector. Despite limited data availability, deep learning-based data augmentation and image processing techniques were employed to enable effective training of this complex transformer model. A range of enhancement strategies was applied to the RT-DETR, resulting in remarkable improvements, including a mAP50 of 98.1 %, mAP50–95 of 68.0 %, and a computation speed of 72 FPS. The enhanced RT-DETR outperformed state-of-the-art detectors such as YOLOv8 and YOLOv7 in detecting defects in 3DCP. Furthermore, the improved RT-DETR was used to analyze the relationships between defect count, size, and printer parameters, providing guidance for operators to fine-tune printer settings and promptly address defects. This monitoring method reduces material waste and minimizes the risk of structural collapse during the printing process.