Unsupervised anomaly localization network for automated fiber placement defect detection under data scarcity

The advancement of defect detection systems for automated fiber placement processes has been greatly accelerated by modern sensing technologies and intelligent inspection algorithms, resulting in enhanced manufacturing efficiency and product quality. Nevertheless, most existing detection algorithms depend on supervised learning frameworks, which are limited by the need for labor-intensive annotation and the scarcity of labeled data. To address these limitations, a novel feature-embedding based unsupervised anomaly localization model, named DenseFlow, has been introduced. This model leverages a pretrained feature extractor and a densely connected multiscale normalizing flow module, enabling robust and efficient anomaly localization by evaluating pixel-wise log-likelihood distributions. To evaluate the proposed method's effectiveness, an anomaly localization dataset, referred to as AFP-AL, was created using a robotic inspection system equipped with a laser profilometer. The dataset construction involved multiphase preprocessing strategies, including the transformation of point clouds into grayscale depth images, sliding window partitioning, and image synthesis operations. Experimental results show that DenseFlow outperforms competing methods adapted to the AFP-AL task, reaching 98.17%, 92.02%, and 69.50% in the threshold-agnostic metrics AU-ROC, AU-PRO, and AU-IoU, respectively. Furthermore, the localization results highlight its resilience to errors introduced by human annotation, which are inherent in supervised learning approaches. Additionally, the proposed methodology is highly adaptable, requiring only minor modifications to process images from various sensors. Its seamless integration into AFP systems enables real-time defect detection, underscoring its potential to advance more efficient and autonomous monitoring systems in AFP manufacturing.