Open Access

ARTICLE

Unsupervised Anomaly Detection System for High-Speed Railway Noise Barrier Using UAV Imagery

Jing Cui¹, Yong Qin^2,*, Yixuan Geng³, Miao Guo^4,*, Xue Yang⁴, Wanyin Shi⁵

1 China Railway Signal & Communication Corp., Beijing, China
2 The State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University, Beijing, China
3 Department of Mechanical Engineering, The University of Hong Kong, Hong Kong, China
4 Faculty of Transportation Engineering, Kunming University of Science and Technology, Kunming, China
5 Yunnan Institute of Economics and Management, Kunming, China

* Corresponding Authors: Yong Qin. Email: ; Miao Guo. Email:

(This article belongs to the Special Issue: Low-altitude Intelligence Transportation Systems: Perception, Decision-Making, Planning, and Optimization)

Structural Durability & Health Monitoring 2026, 20(5), 14 https://doi.org/10.32604/sdhm.2026.081306

Received 27 February 2026; Accepted 09 May 2026; Issue published 24 August 2026

Abstract

Noise barriers (NBs) play a significant role in reducing railway noise and preventing foreign-object intrusion. However, surface damage, corrosion, rust, missing components, and local deformation may gradually reduce their structural reliability and threaten railway operation safety. Because NB anomalies are diverse and defect samples are limited, it remains difficult to build a general detector using conventional supervised learning. To address this problem, this study proposes an unsupervised anomaly detection system for railway NBs using UAV imagery. First, a color-prior-based NB localization algorithm is developed in the HSV color space to extract NB regions without cumbersome pixel-level labeling or localization-network training. Second, a teacher-student-autoencoder framework, denoted UADNet, is designed for NB anomaly detection. In this framework, a GhostConv-based student network learns normal local feature regression, while a skip connection autoencoder (SCAE) models normal global feature consistency. Anomalies are highlighted when both branches fail to reproduce the expected normal responses. Experiments on public anomaly-detection benchmarks and a customized UAV railway NB dataset demonstrate that the proposed system achieves high detection accuracy, accurate localization, and competitive inference speed in complex rail environments.

---

Keywords

---