In-situ real-time defect detection, mitigation and self-supervised adaptation based on visual foundation model for material extrusion additive manufacturing

Abstract

Material extrusion has become the most common additive manufacturing (AM) method, but its further industrial applications are limited by low reliability and error susceptibility. Therefore, defect detection and process control are of crucial importance. The lack of theoretical analysis in the closed-loop process control prevents both the rapidity and robustness of defect mitigation. Meanwhile, obtaining sufficient labelled datasets for non-parametric defects is challenging. A real-time visual prediction and fuzzy control system was proposed to achieve rapid and stable defect mitigation. A visual foundation model (VFM) was trained by the dataset with over 560,000 images generated through a visualized automatic annotation system (VAAS). A closed-loop system with VFM was modelled and identified to clarify the control challenges: the time delay and variable response of closed-loop process control, as well as demonstrate the instability of proportional control. Besides, a fuzzy controller was designed to address the control challenges. Additionally, a self-supervised transfer learning (TL) framework is introduced, combining clustering pseudo-label and fine-tuning, for the cross-domain and cross-task adaptation of the VFM. Experiments show that the fuzzy controller significantly reduces the disturbance rejection time to 15.6 % compared with the current method and improves the stability of the system. Through the TL framework, defect detection in robotic-arm fused deposition modelling (FDM) for a specific printed part was achieved with 89.4 % accuracy with the balanced fine-tuning strategy, paving a way for the wider application of defect detection in AM.