SCA-GANomaly: an unsupervised anomaly detection model of high-speed railway catenary components

With the rapid development of high-speed railways, the safety inspection requirements for supporting devices of high-speed railway catenary components have become increasingly stringent. Currently, defect detection tasks for catenary equipment heavily rely on manual judgment, which is labor-intensive, inefficient, and prone to missing defects. Furthermore, high-speed railway catenary components present a myriad of challenges, including varying sizes, multiple defect types, and a limited number of defects. This complexity makes it difficult to address these issues using a single approach, and Few-Shot defect detection tasks are particularly challenging to achieve using deep learning methods. In this article, we propose a novel SCA-GANomaly model based on the GANomaly network, specifically designed to overcome the challenges of few-shot defect detection by utilizing only normal high-speed railway catenary images for training. The model incorporates selective skip connections and hybrid attention mechanisms, and the loss function is optimized using Earth Mover(EM) Distance, effectively improving image reconstruction and enhancing training stability for high-speed railway catenary images. We conduct comprehensive experiments and evaluations on three types of components: Insulators, Oblique Bracing Wires(OBW), and Puller bolts. The results demonstrate the excellent generalization capabilities of our proposed model, significantly outperforming the GANomaly model in terms of defect detection performance. Specifically, the AUC values for Insulator defect detection increased by 0.15, OBW defect detection by 0.10, and Puller bolts defect detection by 0.12. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.