Sequential pool-based batch-mode active learning based on weighted uncertainty difference sampling for automated warping detection in additive manufacturing

Abstract

Warping deformation is a common defect in parts fabricated through material extrusion for polymers. Despite the need for automated warping detection, developing an object detection model remains challenging due to the limited number of experimental samples, the similarity of time-series warping images, and the labor-intensive nature of image labeling. To facilitate efficient warping detection modeling, this study proposes a sequential pool-based batch-mode active learning framework that iteratively updates the warping detection model after each fabrication step to collect a sequential data pool. The proposed framework enhances modeling effectiveness through a novel data query method called weighted uncertainty difference sampling (WUDS). WUDS selects a batch dataset for new training data instances by considering both the current detection uncertainty and its difference in time-series images. To validate the proposed active learning framework based on WUDS, warping deformation images were collected from recorded video data during the material extrusion of five specimens using a metal–polymer composite filament. A warping detection model based on the You Only Look Once (YOLO) object detection algorithm was then trained using the proposed active learning framework with WUDS, along with three common data query methods for comparison. In addition, the effect of batch size on the performance of the proposed active learning framework with WUDS was analyzed to investigate modeling efficiency. The results demonstrate that the proposed active learning framework with WUDS achieved higher detection performance and lower performance variability than with other data query methods. Furthermore, the proposed active learning framework generated a competent warping detection model despite limited experiments for data collection and a partial set of the data pool. These findings suggest that the proposed active learning approach can contribute to rapid modeling with robust machine learning performance for manufacturing processes.